Proceedings of the 2nd International Conference on Big Data, IoT and Machine Learning

For the technological shift, the advancing epoch toward cryptocurrency intensified the impactful method. Metaverse can originate the base operation into a diversified level. The extension of digital marketing contributes to blockchain technology more. Our research demonstrates attested cryptocurrency price evaluation associated with the stock and sentiment. In our research, we have implemented various techniques to predict cryptocurrency prices. Cryptocurrencies like Bitcoin, Ethereum, and Litecoin are the primary focus of this paper. Our research observes the fluctuation in cryptocurrency prices. In our research procedure, deep learning models like LSTM, GRU, LSTM-GRU hybrid, and ARIMA for time series prediction were implemented. The research provides cogent insights into cryptocurrency price prediction fluidity with the stock price and Twitter sentiment on cryptocurrencies. Additionally, the implementation of the LSTM time series model on the combined data depicts the relationship between stock price, Twitter sentiment, and cryptocurrency price pertinence.

The top-k dominating (TKD) query is a method for discovering the most prominent k items from a large dataset. According to the top-k dominance rule, an item is dominant over another if and only if it outweighs the latter in each and every dimension. For example, when a user conducts a search for the best hotels in a city, the most dominant query will provide the top k hotels according to the user’s preferences. The customer can reserve a room at any of the recommended hotels and enjoy the best services available. However, even among the best hotels, opinions might differ widely. User has no way of knowing if this is a better option or not. Because the reviews may be fake or self-made. It would be easy to give many positive fake reviews. The top k dominant query can’t handle this kind of situation. Moreover, criteria of traditional top k dominating queries generate more than k records for a top k dominating query. The problem is magnified by the available dataset when it is incomplete. To solve this problem, we present a new technique for determining the top k dominating queries which is called top k dominance by stability. The stability of each data point is determined, and then the k most stable data points are taken for ranking. The stability of a rating can be found by using the standard deviation. Ratings for a product are more stable and reliable if they clusters around the mean. It is demonstrated through an evaluation that the performance of the proposed technique provides superior results for top-k dominant queries comparing to other methods.

Traditional, sequence alignment-based (AB) methods are time-consuming and show $$NP-hard$$ N P - h a r d complexity for large sequences. That is why alignment-free (AF) approaches become popular among scientists due to low time complexity. Besides, sequence analysis is crucial for gene finding, modification, new variety development, etc. However, existing AF algorithms utilized $$k-mer$$ k - m e r count, histogram, chaos game representation (CGR), etc., but these have low accuracy rates. Therefore, in this research, a novel $$k-mer$$ k - m e r derived transition spatial feature along with standard deviation and the median is used for representing a sequence. Median and standard deviation are extracted from first-order derivatives of $$k-mer$$ k - m e r positions, and the transition is derived from second-order derivatives. The method is tested in six challenging benchmark datasets from different viewpoints and achieved top-rank accuracy for all datasets and saves 23–99% memory consumption. Here, the top accuracy indicates that extracted features are highly efficient to represent the inherent property of a sequence. Moreover, the time consumption is close to state-of-the-art methods and a thousand times faster than the existing MEGA tool. Therefore, industries can use this method with conviction.

Divorce is a legal process that formally ends a marital union between two individuals. In modern times, it is considered one of the major social issues which is rapidly increasing day by day. It not only terminates the relationship between two persons but also affects the harmony of their family members and other associated people. The aim of this work is to propose an Explainable AI (XAI)-based model that extracts significant factors of divorced/happy couples based on the outcomes of the best classifier. In this work, we gathered a divorced predictor dataset from the University of California Irvine (UCI) machine learning repository. This dataset was subsequently preprocessed and employed different classifiers to identify divorced/happy couples more precisely. After evaluating the performance of individual classifiers, SVM showed the best classification performance where its maximum accuracy is 98.23%. Then, we considered one of the most used XAI models called Shapley Additive Explanations (SHAP) to employ SVM for representing which individual feature values are responsible for identifying happy couples/divorce cases. Thus, A40 is found as the most important feature using SHAP analysis.

Rioting is an act of participating in a violent public disturbance, which involves multiple individuals engaging in destructive activities. Such activities can include vandalism, theft from both public and private property, physical assaults on others, and looting. Riots can significantly harm both government and public property, resulting in losses of life, injuries, and property damage. Most of the time, it has been observed that private and public transport turned into the major targets of riots. By detecting potential threats and responding quickly, autonomous vehicles equipped with riot prevention features can help to prevent harm to both individuals and property during a riot. Moreover, riot threat-detecting features can contribute to minimizing the economic impact of riots, which is particularly important for businesses and communities that rely on tourism, trade, and commerce. Despite the development of various safety features in autonomous vehicles, there is currently a lack of effective measures to detect riots and violent public disturbances on roads and highways. In this study, we propose a solution for leveraging the You Only Look Once (YOLO) algorithm to detect six types of road objects and one class of threats for autonomous vehicles. The YOLO version 8 model was trained and assessed on a dataset of road objects including riot threats, and it achieved a maximum accuracy of 97.71%. Additionally, the proposed solution can be coupled with ground robots and unmanned aerial vehicles technology to enable real-time monitoring and treatment of chaotic and risky zones of riot.

The use of social media is widespread in modern culture. People are making purchases on social media sites. Compared to more conventional approaches, digital marketing has shown to be successful. Both academics and researchers can use it to understand better how to use social media to influence consumers’ purchasing decisions. In this study, we proposed a machine learning (ML)-based customer purchase decision prediction system with model explainability. We use eight well-known ML algorithms for prediction and SHAP, SHAPASH, and LIME for model explainability. Among the models, the RF shows its superiority by achieving 93.52% accuracy with 93% precision, 92% recall, and 93% F1 score. We apply the XAI tools on RF and reveal the behind story of customer purchase decisions. We show both the global and local explainability of the model to find the actual cause of online purchases and how the features influence the customer to purchase. This study helps online marketers grow their businesses, and customers can also be benefited from it.

Biomolecular programming encompasses the utilization of diverse chemical reactions to execute computational functions and encode data within proteins and nucleic acids. DNA, also known as deoxyribose nucleic acid, displays remarkably consistent chemical behavior at the molecular level, rendering it a superb substrate for constructing logical operating systems and molecular computers. Field programmable gate arrays (FPGAs) are integrated circuits built upon a matrix of configurable logic blocks (CLBs) interconnected via programmable links. FPGAs are versatile, user-configurable logic devices capable of performing tasks ranging from basic logic gate operations to complex systems. In this research, an FPGA has been meticulously crafted utilizing gates reliant on DNA. Harnessing the distinctive characteristics of DNA-driven computing, DNA-based FPGAs are capable of concurrently performing billions of operations and delivering extensive memory capacity within a confined space. The advantages of DNA-based FPGA logic circuits extend beyond reducing gate counts through additional output state representation; they also enable circuit compression based on input conditions.

Financial institutions, especially the banking sector, have become one of the major pillars of any economy. Given the modern banking platforms, businesses can conduct their activities more smoothly and fast. However, the banking industry is not immune to criticism. Fraudulent individuals are evident in conducting terrorist financing, financial fraud, money laundering, etc. Non-performing loan, default loan, is the major concern of today’s policymakers. Therefore, figuring out a way to control this threat is of utmost importance. Considering this, in this study, we propose a Machine Learning (ML) and Explainable AI (XAI)-based methodology to predict the P2P Bank Load Default (BLD) network and explain the hidden stories of loan default from the customers’ behavior. We employ 10 well-known ML algorithms to predict the BLD from the secondary dataset and apply four XAI tools (SHAP, SHAPASH, ELI5, LIME) on the top performer ML algorithm. The result reveals that Random Forest (RF) outperforms all the algorithms, and it shows 98% accuracy, precision, recall, and F1-score. The XAI tools find the top features for the bank loan default and the contribution of the top features to the model’s performance in terms of predictions. This study can be a guideline for the bank to verify a customer before issuing the loan and can update their policy based on the explainability outputs.

This research aims to address the challenge of constructing rovers for exploring rough and inaccessible terrains, which typically requires significant investments of time and resources. Our approach focuses on developing a small-scale, low-fidelity rover that leverages off-the-shelf parts and can navigate various challenging terrains, such as sand, mud, and brick roads. Our objective is to offer a cost-effective solution that is accessible to a broad range of stakeholders, including hobbyists, researchers, and other individuals interested in exploring harsh environments. By adopting this strategy, we hope to promote greater participation and innovation in the field of planetary exploration.

The identification of violence in real-world scenarios is imperative as it enables the detection of aggressive behavior, thereby preventing harm to individuals and communities. This is crucial for ensuring public safety, facilitating effective crime investigation, promoting child safety, safeguarding mental health, and facilitating social media moderation. Various methods, including handcrafted techniques and deep learning algorithms, can be utilized in surveillance or CCTV cameras, as well as smartphones, to enable timely detection of violent behavior and facilitate appropriate action and intervention. In this study, we introduce VioNET, a novel approach that combines a 3D Convolutional Neural Network and a Vision Transformer with Bidirectional LSTM for the purpose of accurately detecting violence in video data. Since video data is inherently sequential, the extraction of spatiotemporal features is essential to accurate detection. The use of these two deep learning methods facilitates the extraction of maximum features, which are then fused together to classify videos with the highest possible accuracy. We evaluate the effectiveness of our approach by employing three datasets: Hockey, Movies, and Violent Flow, for analysis. The proposed model achieved impressive accuracies of 97.85%, 100.00%, and 96.33% on the Hokey, Movie, and Violent Flow datasets, respectively. Based on the obtained results, it is evident that our method showcases superior performance, outperforming several existing approaches in the field and establishing itself as a robust and competitive solution for violence detection in videos.

With the increasing need for text summarization techniques that are both efficient and accurate, it becomes crucial to explore avenues that enhance the quality and precision of pre-trained models specifically tailored for summarizing Bengali texts. When it comes to text summarization tasks, there are numerous pre-trained transformer models at one’s disposal. Consequently, it becomes quite a challenge to discern the most informative and relevant summary for a given text among the various options generated by these pre-trained summarization models. This paper aims to identify the most accurate and informative summary for a given text by utilizing a simple but effective ranking-based approach that compares the output of four different pre-trained Bengali text summarization models. The process begins by carrying out preprocessing of the input text that involves eliminating unnecessary elements such as special characters and punctuation marks. Next, we utilize four pre-trained summarization models to generate summaries, followed by applying a text ranking algorithm to identify the most suitable summary. Ultimately, the summary with the highest ranking score is chosen as the final one. To evaluate the effectiveness of this approach, the generated summaries are compared against human-annotated summaries using standard NLG metrics such as BLEU, ROUGE, BERTScore, WIL, WER, and METEOR. Experimental results suggest that by leveraging the strengths of each pre-trained transformer model and combining them using a ranking-based approach, our methodology significantly improves the accuracy and effectiveness of the Bengali text summarization.

One of the most significant and difficult tasks in the modern world is rainfall forecast. Rainfall is a complicated and nonlinear phenomenon that requires sophisticated computer modeling and simulation to anticipate with any degree of accuracy. In many regions of the world, daily rainfall totals are dispersed based on various frequency distribution functions. Rainfall has become a significant factor in agricultural countries. We provide a machine learning-based framework for forecasting the total monthly precipitation. In this study, three Bangladeshi weather stations’ daily rainfall was forecasted with a 365-day lead time using backpropagation long short-term memory (LSTM), random forest, and neural network algorithms. The predicted results from the selected algorithms were compared with the observed data to determine prediction precision by mean absolute error (MAE) test results. We found that selected algorithms predicted daily rainfall with reasonable accuracy. Therefore, year-long, month-long, and day-long rainfall can be predicted using these models.

Abuse of illicit substances among students at educational institutions is rapidly reaching crisis proportions in terms of the number of addicts. Substance abuse of all kinds is becoming an increasingly widespread problem. The goal of this research is to determine whether or not it is possible to make a prediction regarding the abuse of intoxicating substances that is frequent among university students in Bangladesh. The visualization of the data from this study indicated parameters or factors connected to the propensity to take drugs. In this particular study, machine learning is used to derive projections and predictions regarding the use of illicit drugs. In this study, Questionnaires were used to gather the data. For the purpose of carrying out this study, the sample consisted of a total of 468 different pupils. The use of Google Forms allowed for the collection of information as a questionnaire. The majority of those who participated were young adults (between the ages of 24 and 29), and the majority of those young adults were male students. At least 30% of the sample for this study admitted to experimenting with drugs for recreational purposes. According to their responses, the most significant factors that lead to student drug use include spending the night at the home of an addicted friend, being subjected to the destructive influence of peers, and smoking cigarettes. The neural network (Multilayer perceptron) algorithm gives an accuracy of 93% which is the highest, furthermore, the random forest algorithm gives the second-best accuracy.

Recognition of Bengali handwritten digits is a fascinating and demanding research problem that has garnered significant interest from researchers in the fields of pattern recognition. In this paper, a task-oriented deep convolutional architecture for recognizing handwritten Bengali digits is proposed. The main goal is to get a high level of accuracy while using a small number of parameters. The proposed architecture is designed to address the challenges posed by the complex and diverse nature of handwritten numerals in Bengali script, while also being computationally efficient with only 1.08 million trainable parameters. The performance of the model was evaluated by conducting experiments on two commonly used benchmark datasets of handwritten numerals in Bengali script, CMATERdb-3.1.1 and BanglaLekha-isolated-numerals. Different augmentation techniques were utilized to enhance the diversity and size of the training set, which led to improved robustness and generalization of the model. On the CMATERdb-3.1.1 dataset, the proposed model achieved an accuracy of 99.28%, and on the BanglaLekha-isolated-numerals dataset, it achieved an accuracy of 99.12%, outperforming several state-of-the-art models with comparable or larger numbers of parameters. The results suggest that this task-oriented model can be an efficient and effective solution for the recognition of Bengali handwritten numerals, with potential applications in document analysis, digitization, and text recognition.

In modern times, the exponential growth of digital images threatens conventional image retrieval frameworks. Shallow machine learning algorithms degrade image retrieval performance owing to the semantic gap between low-level and high-level features. Hence, deep learning (DL) is a possible way to escape from this. Previous studies have yet to determine the superior DL model or combination for similarity-based image retrieval. In this paper, we apply deep features instead of traditional features for image retrieval. We first execute eight pretrained DL models individually and then all pairs combining two, three, and four models. Experimental results show that model combinations outperform single models, with increased accuracy as the number of models increases. We combine a maximum of four models. Our empirical findings from various model combinations indicate that the most effective combination includes ResNet101V2, InceptionV3, InceptionResNetV2, and DenseNet201, yielding a mean average precision (mAP) of 97.55% and mean average recall (mAR) of 19.51%.

Lung cancer is one of the leading causes of cancer-related deaths worldwide. The importance of early detection and treatment in enhancing patient outcomes cannot be overstated. However, due to the disease’s complexity, this can be difficult. Deep learning algorithms have demonstrated encouraging results in effectively identifying and predicting lung cancer in recent years. In this paper, the CNN model we propose was trained to classify lung cancer CT scan images as either malignant, benign, or normal and achieved high accuracy on test set data which is 99.47%. The gradient-weighted class activation map (Grad-CAM) technique was used to create a superimposed image and bounded box image to see the highlighted parts based on which the model has taken a decision to predict the class. This shows which parts of the images were paid more attention to by the model while predicting the class. The model also achieves an F1-score of 99% and precision and recall of 99.48 and 99.47%. The dataset used for this research is available online. Overall, this result exceeds the prior benchmark approach used with this dataset.

Stroke is a disease that affects the arteries leading to and within the brain. Detecting stroke early and conveniently is much more difficult as there is no portable system to detect it. Most of the time the expensive diagnosis method of stroke is out of reach for low- and middle-income countries like ours. Hence, there is a significant necessity for an effective and labor-saving self-diagnosis platform. For the last few years, machine learning and deep learning are used to study medical-related information. Lately, deep learning has very quickly become transformative for health care, offering the ability to analyze data with a speed and much precision. This study gives an automated system to detect the stroke from prepossessed data using CNN and other deep learning models. The proposed methodology is to mainly classify the stroke person’s face from the normal or expressions face. For classification, we passed prepossessed stroke images for training, fed them into various deep architecture, and finally based on the classified expression, we classified normal and stroke patient. The experimental result shows that CNN classification model achieves accuracy 97.145% which is satisfactory. Overall, the aim of this study is to establish a fast and reliable system which will detect stroke on its early stage.

Various organisms, such as bacterial and viral infections, can cause a lung infection known as pneumonia. It is a significant health concern, particularly in developing and underdeveloped countries with high pollution rates, overcrowding, and limited healthcare infrastructure. In order to effectively treat pneumonia and improve survival rates, early detection is essential. The simplest technique for identifying pneumonia is a chest X-ray (CXR) study, but CXR analysis can be subjective and challenging. In this paper, we have developed a method for automatically detecting pneumonia from CXR images by combining transfer learning and an ensemble of three CNN network architectures (InceptionV3, MobileNetV2, and Xception) with the weighted average ensemble method. We evaluated our approach on chest X-ray datasets, achieving a maximum accuracy of 92% and F1-scores of 87% and 92% for normal and pneumonia, respectively. Our proposed method outperforms existing ensemble techniques and other cutting-edge approaches, demonstrating the potential for improving pneumonia diagnosis through deep learning-based approaches.

Road quality assessment is a crucial task for maintaining transportation infrastructure, but it can be challenging and resource-intensive. Recent advances in remote sensing and machine learning have opened up new possibilities for assessing road quality using satellite images. This proposes a comprehensive dataset of approximately 45 k road images, which have been classified into five classes based on road quality. It contains N8, N102, N104, N502, N702, N704, N707, N803, N805, N806, N808, national highways of Bangladesh road images. The dataset has been used to train six deep learning models, including ResNet50, ResNet152, VGG19, DenseNet169, MobileNet V2, and SqueezeNet, on this dataset to identify the roughness levels of the road surfaces. The best accuracy of 82% was obtained from ResNet50. Our analysis shows that ResNet50 performs well on large dataset with noisy and unclear images due to its unique architecture that allows for better gradient propagation during training. This paper also analyzed the relationship between roughness levels and other factors such as traffic volume and road type. The findings of this study demonstrate the potential of satellite-based road quality monitoring for improving transportation infrastructure management and supporting economic development.

This study introduces a new local feature descriptor called evaluation window-based adjacent distance local ternary pattern EAdLTP for image classification. It is created by combining evaluation window EwLBP and adjacent distance local ternary pattern (AdLTP) to achieve robustness by encoding adjacent information. EwLBP produces an evaluation window to reduce noise in the neighbor’s values, and AdLTP captures the relationships between sequential neighbors. The adjacent sub-image window and the adjacent neighbor window are used to calculate the neighbors and extract the binary code that is modified to improve the information of the adjacent neighbors. The final EAdLTP pattern is divided into two parts (EAdLTPU and EAdLTPL), and the feature descriptor vector is obtained by concatenating their histograms. The proposed EAdLTP descriptor is tested on the KTH-TIPS and KTH-TIPS2b datasets and consistently outperforms other fundamental methods by being more robust against noise.

Sentiment analysis is a critical area of study right now. The evolution of social media, websites, blogs, opinions, ratings, and so on. It has expanded significantly along with the development of Internet usage. Through comments, likes, and other interactions with social media posts, people can share their thoughts and feelings. YouTube sentiment analysis has increased as a result of the sharp increase in the amount of user- or viewer-generated data or material on the platform. This study creates a deep learning classifier to analyze YouTube videos and detect the sentiment automatically. We train and assess two long short-term memory-based models. To ascertain which deep learning model on a labeled dataset performs best in terms of accuracy, recall, precision, F1 score, and ROC curve, experiments are conducted. The findings show that a Bi-LSTM-based model, with an accuracy of 71.74%, performs the best overall. The Bi-LSTM not only addresses the issue of long-term reliance, but also takes the text’s context into account. Finally, a comparison is made using experimental findings obtained using various models.

In medical applications, the boundless potential of image processing utilizing Deep Neural Networks has grabbed the interest of researchers. Brain tumor segmentation, which is a crucial piece of task, determines the location and extent of tumor areas. Numerous techniques for segmentation have been suggested by researchers. One significant disadvantage of the existing architectures is the presence of a large number of trainable parameters. It makes the system complex, expensive to train, and unsuitable for integration in low-powered devices. In this paper, we present an efficient, two-stage approach for the effective segmentation of brain tumor from MRI images using RCNN and a modified U-Net. The proposed system was evaluated and verified using a publicly available Figshare dataset (Cheng in, 2017 [1]). The system is low-complex with small number of parameters compared to other existing architectures. It was tested and compared to the original U-Net, and despite having a large decrease in total trainable parameters, it obtained a comparable performance with an accuracy of 99.78%, IoU of 89.76%, and a dice score of 94.53% in our experiments.

Chronic diseases are the primary causes of mortality and impairment in Bangladesh, as well as the leading cause of healthcare expenses. Doctors frequently diagnose chronic disease patients based on symptoms. Following the test reports, patients are then prescribed medication for a certain time. During this time, it is imperative to regularly monitor the patient’s condition. However, these illnesses are expensive to diagnose. In this particular scenario, Fuzzy TOPSIS has the potential to be an effective solution. The goal of using Fuzzy TOPSIS with multi-criteria decision-making (MCDM) in patient follow-up is that instead of depending entirely on diagnosis, we have combined logic-based premises and outcomes based on statistics. This study presents an expert system that can determine a chronic disease patient’s weekly health state. Additionally, the patient’s health is monitored daily to track any changes that may occur. We have selected five distinct chronic conditions and 19 people who are afflicted with five distinct chronic diseases. The laboratory criteria of these diseases are chosen by two feature selection methods. Multiple decision-makers were explored for each condition. In response to the data collection, a new decision matrix has been constructed. The newly created decision matrix is then normalized and weighted. Following this, the patient’s weekly health condition is determined based on the closeness coefficient scores. Finally, variance is utilized to determine the variation from day to day. As a consequence of this, one can take the appropriate measures for further treatment, which reduces both expenses and the amount of time required.

Leukemia is a life-threatening condition affecting people globally, making accurate diagnosis crucial for timely intervention. Consequently, researchers have been exploring automated methods to enable prompt action. The classification of leukemia into multiple subtypes according to WHO standards presents a unique challenge. Unlike binary classification, interclass features are highly similar, leading to misclassification. Ergo, we employ attention mechanisms to tackle this problem. Our proposed deep learning architecture combines transfer learning with attention mechanisms to classify subtypes of leukemia accurately. Using a publicly available dataset of blood cell images that adhered to WHO standards, we illustrate the potency of our approach. Our DenseNet201 with CBAM model achieves a remarkable 99.85% overall accuracy without resorting to data augmentation, surpassing previous methods on this dataset and attaining state-of-the-art results compared to other leukemia literature. To interpret the model’s decision-making process and evaluate the efficacy of the attention mechanism in identifying discriminating features, we showcase GradCAM images and intermediate layer feature maps generated from our custom CNN. The proposed approach enhances leukemia classification accuracy and efficiency, providing clinical decision-making insights.

The classification of different types of date fruit can be challenging due to their varying angles, separation, and exposure to light. To overcome this, image analysis and pattern recognition techniques can be applied. In this study, a proposed workflow involving filter-based feature selection, principal component analysis (PCA), and recursive feature elimination (RFE) was used to select the most salient features for the classification of seven types of date fruit. Machine learning classifiers were then applied to evaluate the performance of the approach, with support vector machine (SVM) outperforming other classifiers. The proposed approach achieved an overall accuracy of 95% using Fisher’s exact test, PCA, RFE, and SVM. Comparison with previous works revealed that the proposed approach was effective in obtaining efficient outcomes. The study demonstrated the potential of using machine learning techniques for date fruit classification, which can ultimately lead to more efficient and accurate grading and sorting processes.

Thyroid disease is a prevalent health problem that requires early detection for effective treatment. However, there is no universal model for detecting thyroid abnormalities efficiently. This study proposes a three-layer thyroid disease detection framework that utilizes different feature engineering techniques to explore various thyroid datasets, improving model performance. We propose a hybrid framework to identify the most relevant features utilizing feature selection techniques contributing significantly to the model’s performance. We evaluate the proposed bagging XGBoost ensemble model’s performance against K-nearest neighbors, extreme learning machines, and random forest classifiers. It surpassed all with 98.44% accuracy with only 53.57% of the dataset’s features. The proposed model outperformed other classifiers with 98.65% accuracy during cross-validation on the second dataset. Stress testing with train–test split ratios of 70–30 and 30–70 produced 94.92% and 94.68% accuracy rates, respectively, that also outperform the benchmark ML models. These findings have significant implications for improving thyroid disease diagnosis using machine learning techniques.

A neurological illness known as Parkinson disease (PD) that affects humans through the nerve cells of the neurological system aging. Unmanageable tremors or jaw, arm, leg, or hand motions are the possible signs. Currently, the only way to diagnose PD is to keep an eye out for its prodromal signs. The speed of therapy can be accelerated by a physician’s expertise being increased by an automated classification approach. This study’s goal is to refine an automated Parkinson’s disease classifier using artificial neural networks (ANNs) and a technique for extracting hybrid features. The prepossessing of the images serves as the initial step in the classification strategy. The analytical properties of the prepossessed pictures are takeout using a mixed feature extraction method that combines stationary wavelet transform (SWT) and gray-level co-occurrence matrix (GLCM) approaches to increase the effectiveness of the classification process. In order to determine, if a patient has Parkinson disease or not, the ANN is used as the final step. The chosen strategy replaces the typical discrete wavelet transform (DWT) method with a blended approach of feature extraction based on SWT, leading to an improved accuracy of classification which is of 84.1%, which is significantly better than modern multi-classification methods.

A brain tumor is a dangerous condition that can be challenging to reliably identify using conventional techniques, such as by looking at MRI scans. To solve this problem, our convolutional neural network (CNN) and transfer learning models were developed to distinguish between the three types of brain cancers that are most frequently found: gliomas, meningiomas, and pituitary tumors. The 7023 MRI scans of the human brain that make up our dataset were separated into four groups based on their tumor status: pituitary, glioma, meningioma, and no tumor. We used an ensemble method to combine pre-trained models and achieved exceptional accuracy in identifying the presence of a tumor, a combination of CNN and VGG16 with an accuracy of 0.97687 in validation data and up to 0.9801 in test data. Our findings showed how effectively and accurately our method classified brain cancers from MRI images. For a brain tumor to be successfully treated and to be life-saving, it must be identified as early as possible.

Brain tumors are a major global health issue, and their detection can be challenging. Typically, doctors visually inspect brain images to locate tumors, but this method is time-consuming and prone to errors. Recently, there has been progress in automated methods for early brain tumor diagnosis. However, challenges remain regarding limited precision and high false positive rate. Thus, an effective approach is needed for accurate tumor classification, utilizing the strong features of the tumor. This study presents a novel approach for brain tumor detection and classification using the fusion of deep features from MRI scans. Preprocessing, including cropping and resizing, eliminates irrelevant information from the brain MRI. After that, transfer learning models (DenseNet-121, InceptionV3, MobileNetV2) extract meaningful deep features, capturing complex patterns in MRI images for accurate tumor detection and classification. The extracted deep features are combined into a single feature vector and utilized as input for both a support vector machine (SVM) and a K-nearest neighbor (KNN) classifier for the final prediction. By combining the deep features into a unified feature vector, the model incorporates more information, resulting in improved classification performance. Two publicly available datasets are used to evaluate the effectiveness of the proposed approach, and the results demonstrate that the combined feature vector outperforms the individual feature vectors. Furthermore, the proposed approach demonstrated superior performance compared to existing methods, achieving the highest classification accuracy. This highlights its potential to support medical professionals in accurately detecting and classifying brain tumors.

Coronavirus disease (COVID-19), an effect of the SARS-CoV-2 virus, is an emerging infectious disease that infects humans due to interspecies transmission. Typically, Spike (S) protein plays a crucial role in entering host cells and beginning infection in contrast to the Membrane (M) and Envelope (E) proteins which are primarily involved in virus assembly. The purpose of this research is to examine the S protein sequences of the human virus to forecast SARS-CoV-2 infection potential. Using computational modeling and bioinformatics approaches, we analyzed the dataset taken from the Virus Pathogen Database except for 47 SARS2-CoV sequences from the SARS2-CoV Database containing aligned sequences of lengths 2396. This study suggests using several machine learning methods to determine infectivity based on the sequence of the coronavirus spike protein. The classical machine learning approach such as k-nearest neighbors (KNNs), Gaussian Naive Bayes (GaussianNB), support vector machine (SVM), decision tree (DT) is used. Bagging Classifier, Gradient Boosting algorithm, and Random Forest are used as the ensemble machine learning algorithm. Artificial Neural Network (ANN) is also applied as a deep learning model. The Gradient Boosting algorithm demonstrates the best accuracy in determining infectivity status. According to our research, the outcome of this study has a chance to accurately forecast coronavirus infectivity from genetic information.

Autism Spectrum Disorder (ASD) is a diverse neurological problem with several contributing factors involving both genetic and environmental variables. The diagnosis of ASD based on neural activity analysis of various signals from the brain, especially electroencephalography (EEG), has drawn attention. A technique for capturing an electrogram of the brain's spontaneous electrical activity is called EEG which is simple to use and non-intrusive. The aim of this study is to classify people with ASD and controls using the most efficient band by examining different sub-bands (such as alpha, beta, and gamma) of the EEG data. Normalized Mutual Information (NMI) is used to create Connectivity Feature Maps (CFMs). Classification is performed through Convolutional Neural Network (CNN). Gamma band is found effective for ASD detection with an accuracy of 96%.

Pneumonia is a serious respiratory infection that can be fatal if not diagnosed promptly. However, even experienced radiologists can find it challenging to arrive at an accurate assessment of pneumonia based on chest X-ray pictures due to their blurriness. In order to solve this issue, we have come up with a novel solution of automated system (ICDP) that uses deep learning techniques to improve diagnostic accuracy. Using deep convolutional neural networks (CNNs), we have built our system. Our model incorporates several layers designed for identifying features in X-ray images and determining whether there is pneumonia or not. In our experiments, we found that our proposed model achieved 95% accuracy in diagnosing pneumonia, and its performance improved with training. Our approach has the potential to improve radiologists’ ability to identify pneumonia, which might ultimately save many lives.

Brain tumors have been considered the world’s, most dangerous disease. Worldwide, approximately 0.25 million people die every year due to CNS tumors and primary cancerous brains. Histopathological examination of biopsy samples is still used in the diagnosis and classification of brain tumors. The existing method is obtrusive, tedious, and sensitive to errors by individuals. To overcome the pitfalls mentioned above for brain tumor multi-class classification, a fully automated deep learning system has been proposed for the early detection of brain tumors in an efficient way. For the purpose of early diagnosis, this research uses Convolutional Neural Networks (CNN) to multi-classify brain tumors. A heatmap image has been produced using the gradient-weighted class activation map (Grad-CAM) technique, and then from the heatmap, a bounded box image has been generated to demonstrate which regions of an image the proposed model devoted significantly more focus to than the other areas. This has been done to show that the proposed model is highly efficient. For the publicly accessible combined dataset, the proposed model attained testing accuracy, precision, recall, F1-score, and AUC scores of 98.75%, 98.00%, 98.00%, 97.76%, and 100%, respectively. The state-of-the-art (SOTA) methods for brain tumor multi-class classification has been outperformed by the proposed way after a number of hyperparameters have been tuned to yield the best outcomes. As a result, this model can help healthcare and radiology professionals check their first screening in order to categorize various kinds of brain tumors.

Ensuring safe drinking water is vital for human health, but it remains a challenging task. School children are vulnerable to various diseases caused by contaminated water. To address this issue, this study proposes a system that utilizes machine learning and IoT to provide a constant supply of pure drinking water for school children. The system collects water parameter data using different sensors, including pH, DS18B20 temperature, and turbidity sensors. The collected data is transmitted to the cloud via an Arduino UNO and ESP 8266 Wi-Fi protocol. The Long Short-Term Memory (LSTM) algorithm is then applied to analyze the data and predict water quality levels. The system's performance was evaluated through several test cases, and the results were promising. The IoT hardware successfully transmitted real-time sensor readings to the machine learning model through the cloud, ensuring drinking water quality. The LSTM model achieved an average 93.33% R2 score and performed real-time forecasts. Overall, the proposed system can help ensure safe drinking water for school children by leveraging the benefits of machine learning and IoT.

Bangla is the sixth most widely spoken language worldwide, making the recognition of handwritten characters in this language an important task. However, recognizing Bangla characters is challenging due to the large number of characters in this language, which is around 400. Furthermore, with the increase in number of characters to recognize, the accuracy decreases. In this research, we propose a transfer learning architecture that branches pretrained convolutional neural network models VGG-16 and DenseNet-121, along with Nadam optimizer. The model is evaluated on Ekush dataset which contains 3,68,776 images over 122 characters or classes. Our proposed model achieved 96.94% accuracy on training data and 96.14% accuracy on test data. A comparison of the various state-of-the-art benchmarks for classifying Bangla Handwritten Characters demonstrates that the proposed model has outperformed previous studies. Our efforts will contribute to the development of an effective tool for automatically recognizing handwritten characters in Bangla.

Voice command-based human-computer interaction (HCI) is becoming useful and practical day by day. Here, we present an open-source voice command-based speech interaction system featuring hands-free interactions of mouse and keyboard without any active internet connection. The usefulness of the application is demonstrated by evaluating the application thoroughly keeping in mind for a motor-disabled person as well as for a normal person. Several participants of different age groups who evaluated the system found that the implemented system worked reliably and helped them complete the task with voice commands only without using mouse and keyboard. In this research, we identify common voice tokens a person would speak to accomplish a human-computer interaction, then we program the tokens to work with major speech recognition platforms such as CMU PocketSphinx, DeepSpeech, and VOSK. Different results were obtained for each platform based on detection rate, accuracy, inference time, CPU usage, system memory usage, and various age group users’ accuracy. In the results section, we present that using our proposed system, the VOSK speech recognition platform outperformed other compared platforms having a 91% successful task completion rate for real-time applications.

As they play a significant role in autonomous driving and traffic safety, traffic sign identification and recognition have recently emerged as one of the most significant fields in image processing and computer vision. Early studies in this field offered several deep learning-based methods for classifying distinct traffic signs using various standard datasets. However, not many researchers focused on creating a dataset of traffic signs in Bangladesh and applying deep learning techniques to recognize them. In this research, we compare and contrast several deep learning models for recognizing traffic signs from the perspective of Bangladesh. We construct a novel dataset with over 2000 images representing thirteen distinct kinds of typical traffic signs in Bangladesh. Using data augmentation, about 8386 images are generated from the original dataset. Subsequently, transfer learning and fine-tuning approaches are applied to nine different deep learning models using this dataset, and the outcomes are compared. Results indicate that ViT had the highest validation accuracy of 99.91% for fine-tuning, while DenseNet201 had the highest validation accuracy of 99.86% for transfer learning. Almost all models attained excellent training and validation accuracy levels, showing that they were able to successfully learn the dataset’s characteristics.

The timely and precise diagnosis of brain tumors is crucial in reducing mortality rates. Although Magnetic Resonance Imaging (MRI) is commonly used as a detection tool for brain tumors, the presence of unwanted regions in MRI and multi-class brain MRI datasets may pose challenges in accurately classifying tumors. This study proposed a two-phase end-to-end framework comprising DenseNet-121-based deep learning to extract features and an ensemble of machine learning methodologies for precise classification. The deep learning-based feature extraction phase effectively extracted essential and discriminant features that were utilized by multiple machine learning models. Preprocessing MRI images to eliminate unwanted regions enhanced the deep learning model’s feature extraction capabilities. The effectiveness of the proposed framework was evaluated by measuring the classification performance of the ensemble mechanism, which achieved an accuracy of 98.86% and an f1-score of 98.76% without any data augmentation. Notably, the random forest attained the utmost accuracy and f1-score among the machine learning approaches used in the ensemble technique.

The progressive nature of Parkinson’s disease (PD) means that it eventually affects all areas of the nervous system and the body that the nervous system controls. The disorder usually shows up as tremors, but it can also make people stiff and slow their general movements. Early detection of PD can help take the necessary steps to prevent it, and machine learning (ML) is one of the best solutions nowadays. In this study, we classify PD from voice signals and find the best data balancing and dimensionality reduction techniques to improve the classification accuracy of the classifiers. We propose a stacking ensemble classifier, namely SVCXRF, that outperforms other classifiers. Logistic regression, random forest, K-nearest neighbour, and multilayer perceptron are used as models, and SMOTE, NearMiss, and SMOTETomek are used as data balancing techniques, and principal component analysis (PCA), independent component analysis (ICA), and truncated singular value decomposition (TSVD) are used as dimensionality reduction techniques. We check all the possible combinations for the ML algorithms. Empirical analysis shows that SMOTE and ICA as preprocessing pipelines and the proposed ensemble SVCXRF show maximum 97% accuracy, precision, recall, and f1-score to classify PD from the voice signal. In the best pipeline design, we classify PD using unsupervised ML method K-means clustering with low prediction accuracy. The proposed ensemble algorithm’s superiority is validated using another PD dataset. The tolerance test shows that the proposed ensemble classifier can predict accurately in any training–testing ratio.

Around the world, aloe vera is grown for both agricultural and therapeutic purposes, making it one of the most popular herbal treatments for topical skin diseases. Although aloe vera is unique in terms of providing nutrients and treating diseases, we fail to protect the leaves from bacteria, which significantly harms farms throughout the world. Identification of the foliar disease is critical because of the serious degree of damage. Aloe leaf infections were categorized using a number of deep learning methods, including VGG19, EfficientNetB5, EfficientNetB6, and EfficientNetB7. Using a smartphone camera, 2770 distinct images of aloe plants with two illnesses were included in the dataset (leaf spot and aloe rust). K-means clustering and histogram equalization were then used to preprocess the data. Finally, EfficientNetB7 outperformed the other experimented methods with a maximum accuracy of 94.11% on a total of 170 images from an unseen test set.

Transliteration is crucial in natural language processing as it enables the conversion between two languages while retaining their phonetic representation. The ability to transliterate the Bengali language is essential for cross-lingual communication. Most Bengali machine transliteration techniques rely on one-hot coding to numerically represent features, which is computationally intensive, requires significant memory, and results in lower accuracy in some cases. Addressing these limitations, we present an efficient feature representation method utilizing binary coding and compare it against the one-hot coding approach using two machine learning models: SVM and random forest. The proposed method is evaluated on the Dakshina dataset and the NEWS 2018 dataset for Bengali to English transliteration. The experimental results indicate that our approach outperforms the traditional methods while requiring significantly less memory and training time. Overall, the proposed method enables a better romanization of frequently used words.

Emotion detection is a task within the field of Affective Computing, which focuses on designing systems and technologies capable of recognizing, interpreting, processing, and simulating human emotions. Building an affective database for Bangla text can have numerous benefits, including improving the accuracy of emotion detection, enhancing natural language processing, and providing a deeper comprehension of how people express emotions in Bangla culture. Identifying emotions from textual data and creating a consistent, reliable, and unbiased emotion dataset pose significant difficulties, especially in low-resourced languages like Bangla. This study addresses the problem of emotion detection in Bangla texts. An emotion dataset was meticulously annotated based on Paul Ekman’s six basic emotions, employing techniques to ensure dataset consistency, reliability, and impartiality. The study achieved an average Cohen’s Kappa score of 87%, indicating nearly perfect agreement between the annotator and the participants. Additionally, various machine learning models and word embeddings were evaluated for emotion detection. The findings demonstrate that Logistic Regression, SVM, and Naïve Bayes yielded the best performance, with Logistic Regression (BOW + Unigram) achieving an accuracy of 68.8% and a macro F1-score of 65%.

Handwritten document images need to be processed carefully to understand meaningful information. For useful collection and future reference, nowadays these documents are now protected in digital library. Historical records, cultural artifacts, and literary assets in Bangla scripts are precious things which should be stored in digitized format so that it can’t be lost. However, character recognition from document images is challenging due to skewness and sometimes it is unmanageable to retrieve information in text format. Moreover, skew rate can deteriorate according to the handwriting style of non-printed document image, especially for Bangla handwritten document image. For this document, characters are not easily separable, but words are discreet with each other. This research addressed this issue with the application of Connected Component Analysis. Although the skew rate can be measured from connected component analysis and bounding box approaches, orthogonal regression is applied to acquire the best fitting line. Experimental results show that the proposed method provides 98.52% accuracy which indicates the effectiveness of the proposed scheme.

Skyline queries have been widely used in a variety of modern database applications. Multicriteria decision-making systems, decision support systems, and recommender systems are among those on the list. As skyline algorithms have so many advantages and may be used in so many various data contexts, they have previously been proposed in a variety of data situations. The scenario of having complete data is not valid in this digital era with a huge amount of data with a missing value. In practical applications, there is a direct dependence between data items and their incompleteness. Besides, dealing with incomplete data may arise problems like losing transitivity and cyclic dominance. With a view to handling these problems, we have proposed a new algorithm that deals with the incompleteness of each data and tries to predict the best data items. We have introduced the concept of the weighting factor. A new dataset has been made comprising rating of famous tourist places in Bangladesh to test the algorithm. Besides, the efficiency has been compared with existing datasets, and our algorithm has shown more practical output than others.

The demand for renewable energy has increased steadily in recent years as people become more aware of their carbon footprint. This has led to a growing need for energy sources that are both sustainable and environmentally friendly. Solar power has emerged as a popular option for generating electricity but has challenges. One of the biggest problems facing solar panels is dust and other garbage buildup, which can reduce their efficiency and output. While keeping solar panels clean around the clock is difficult, automated detection and cleaning systems can help. In this paper, we propose an image processing-based approach that uses a convolutional neural network (CNN) with the popular AlexNet architecture to detect dust on solar panels. Our model achieved an 85% recognition rate for dust detection, which could significantly improve solar panel efficiency. By automating the detection and cleaning process, we can maximize electricity generation and make solar power a viable option for sustainable energy production.

Cloud computing services are available online at any time and from any location. The services for online data access, manipulation, and configuration are provided by cloud computing. The issue of cloud technology’s power usage is addressed in this paper. Methods and algorithms that could lower power usage along with allocating resources are necessary for servers to operate efficiently. Another critical component of cloud technology is load balancing, which permits balanced load distribution among numerous servers to satisfy rising client needs. The contemporary study used several optimization techniques, inclusive of the whale optimization algorithm (WOA), optimization of cat swarms (CSO), BAT-Algorithm, search algorithm for cuckoo (CSA), and particle swarm optimization (PSO), for balancing of loads, power efficiency, and more effective resource distribution to build a productive cloud environment. The consequence indicated that the whale optimization algorithm beats other algorithms in terms of response time, power usage, execution time, and throughput for the settings of seven and eight servers.

Strokes are a leading cause of premature mortality in wealthy nations, and early treatment assistance can significantly prolong a patient’s life. The primary rehabilitative step in the therapy of stroke is determined by how quickly the lesion is identified from MRI images. This will be an essential tool for determining the extent of brain cell damage. However, manual lesion identification takes time and is susceptible to both intra- and inter-observer inconsistencies. In light of this, computerized estimation of the outcome of the ischemic stroke lesion can assist physicians in better evaluating the stroke and providing information on tissue outcomes. This can be achieved by accurately classifying the characteristics employing a convolutional neural network with convolutional layers. Segmentation calls for retaining structural characteristics of pixels in the process of learning the local properties of a picture. Therefore, in the present investigation, a deep learning network is used to segment the Ischemia. The key finding of this study is the extraction of ischemic lesion features through the deployment of the InceptionV3 network and the preservation of information on the z axis through the use of a traditional 3D-U-Net architecture. The trials conducted on the ISLES 2017 dataset yielded an overall segmentation dice coefficient of 0.43. The findings of this study demonstrate that our proposed methodology surpasses previous research.

When users in a digital library read or browse online resources, it generates an immense amount of data. If the underlying system can recommend items, such as books and journals, to the users, it will help them to find the related items. This research analyzes a digital library’s usage data to recommend items to its users, and it uses different clustering algorithms to design the recommender system. We have used content-based clustering, including hierarchical, expectation maximization (EM), K-mean, FarthestFirst and density-based clustering algorithms, and user access pattern-based clustering, which uses a hypergraph-based approach to generate the clusters. This research shows that the recommender system designed using the hypergraph algorithm generates the most accurate recommendation model compared to those designed using the content-based clustering approaches.

Accurate score prediction is essential for teams to develop winning strategies because of the growing popularity of T20 cricket and the significance of setting a challenging target in the first innings. The suggested method entails gathering historical information on T20 matches and applying feature engineering approaches to extract pertinent features. To forecast the first innings score, various regression methods, like XGBoost regression, Lasso regression, and Ridge regression are trained on the dataset. Metrics such as mean absolute error, root mean squared error, and R-squared values are used to assess the performance of the models. The findings demonstrate the potential of machine learning techniques for predicting the first innings score in T20 cricket matches, offering useful information for team strategy. The developed models, implemented codes, and user interface designs are deployed in this link: https://github.com/AST-TheCoder/T20 .

This research paper aims to compare various machine learning algorithms to identify the best career paths for individuals based on Bloom’s taxonomy evaluation. The study uses data from a sample of individuals to train and test the models and evaluates their efficiency based on accuracy rate, precision, recall, and F1-score metrics. The results of the study show that certain machine learning algorithms perform better than others in predicting career paths based on Bloom’s taxonomy evaluation. These findings have implications for career counseling and guidance and may help individuals make informed decisions about their career paths based on their skills, interests, and aptitudes. There are four machine learning models—logistic regression classifier, decision tree classifier, random forest classifier, and K-nearest neighbors’ classifier—which are used to make the comparison of the results. Among the four applied machine learning classifiers (MLCs), random forest beat the other classifiers with an accuracy of 87.77%.

Technology is advancing rapidly nowadays. For this, a large amount of data is being stored in the cloud by tech giants and other organizations. These data include sensitive information like bank transactions, personal healthcare, etc. Whilst these data are used for mining by the data analysts for extracting valuable information, privacy might be breached. That’s why, it should be pre-processed before mining to preserve the privacy of the sensitive data. There are many existing methods to preserve either privacy or utility. However, if we use methods like those, we may lose the privacy of the data or its utility. And if the utility is lost, then it is useless for mining. To preserve both privacy and utility, a hybrid perturbation method called DA3RT is proposed in this study based on derivative, anti-derivative, and geometric rotation. The method is tested with six UCI dataset using three benchmark classifiers. Privacy has been tested with attack resistance, entropy, and other privacy metrics; utility has been tested using accuracy, F1-score, and AUC. The experiment exhibits that DA3RT can preserve privacy as well as utility better than the existing perturbation methods.

Online shopping has become very popular nowadays. It tends to share user experiences of buying products or dealing with the seller by posting reviews. So millions of reviews are being generated daily. It can be difficult for a new customer of that particular product to read all of those reviews and decide whether or not to purchase. In this situation, an overall sentiment(s) of the whole review might help them. Also, people being more creative sometimes post-sarcastic or ironic statements. This may mislead other buyers. The majority of the previous research work regarding product sentiment analysis was confined to two to three sentiments only. Also, a review may express several emotions at once. By doing binary classification, we may miss other emotions present alongside the predicted one. So we have proposed a binary classifier model to separate the sarcastic reviews and then a multi-label classifier model to detect the emotions present in a particular review. We have applied several methods for multi-label classification naming binary relevance, classifier chain, and label powerset on up to four different classifiers. Among them, the OnevsRest classifier along with the support vector classifier as an estimator performed better than the other methods. We have also trained and tested a few binary classifiers for sarcasm detection and got almost the same accuracy of 93.37 and 93.92% for logistic regression and support vector classifier.

Nowadays road accidents are becoming a major global challenge, leading to a high rate of injuries, fatalities, and large economic losses annually. The number of collisions in our nation increasing daily. Developing an accurate model for the prediction of the severity of traffic accidents is a crucial task. Many researchers have already predicted different aspects of road accidents. They have applied different machine learning (ML) algorithms for their work. Whereas a minimal number of researchers have used ensemble-based ML algorithms. In this paper, we have used four ML algorithms such as Random Forest, XGBoost, KNN, and LGBM to form a voting-based ensemble ML model for road accident severity prediction. Multiple road accident datasets have been used to predict the best evaluation for this study. Bangladesh (BD) road accident data from 2017 to 2020 have been collected from Accident Research Institute (ARI) in Bangladesh University of Engineering and Technology (BUET). Another dataset was the USA road accident dataset from 2016 to 2020. Furthermore, SelectKBest and ExtraTreeClassifier algorithms have been used for selecting suitable features for this paper. According to the experimental results, the voting demonstrates the best contribution for both binary and multiclass classification in the BD dataset for achieving the highest accuracy of 96% and 71%, respectively. On the other hand, the voting indicates the highest performance for both binary and multiclass classification in the USA dataset, achieving the accuracy of 92.1% and 87%, respectively. Finally, the result of the comparative analysis shows that voting is providing the best performance than others.

The ocean is a massive expanse of saltwater that spans over 70.8% of the Earth’s surface and holds nearly 97% of the planet’s water. The Southern Ocean is remarkable because it influences worldwide climate patterns and plays a key role in storing a tremendous amount of heat, carbon dioxide, and nutrients. As a result, we are focusing our efforts on solving the mysteries surrounding the Southern Ocean. Therefore, in this study, we attempted to forecast crucial biogeochemical indicators for climate monitoring such as pH (25 $$^{\circ }\text {C}$$ ∘ C ), nitrate content, and relative density of seawater. More research and analysis on these three components might be beneficial because they are critical aspects for ecology, marine biochemistry, and overall climate patterns. This study is conducted to develop new custom kernels (Hyperbolic Sine Kernel, Gaussian Matrix Multiplier Kernel) for mapping the features into desirable space, effectively making assumptions based on existing relevant features using SVR technique. Further, comparing their efficiency against present kernels (on certain constraints) across three dimensions and suggesting a significantly higher-performing kernel.

Research on crimes committed on social media platforms is growing fast as more platforms are made available to the general public (Facebook, Instagram, Twitter, etc.). The focus of this study is on offenses involving verbal abuse directed at a specific individual or group of individuals. As social media platforms continue to develop at an astronomical pace, users’ mental health is being adversely affected as a result. Analysis of public comments is the primary focus of the research, which aims to identify and remove any offensive or improper comments from them. The findings of this study provide a possible solution to this particular issue. We evaluate approximately 13 research papers on verbal harassment comments on social media. In this paper’s review, we attempt to identify and analyze the concealed factors. We also attempt to introduce some tools and algorithms that are used to identify harassment-related comments. We also conducted a survey; the survey questions represent the percentage of social media users who face verbal harassment. In future work, we will attempt to implement LSTM-based comment detectors for identifying instances of online harassment. Also, we provide some justifications for selecting the LSTM algorithm as our future implementation technique.

An increasing amount of harmful effects have been linked to prolonged exposure to abusive language on numerous social media sites. If we want to keep the internet safe and peaceful, we must do something about the epidemic of harsh language. Although studies on the topic of identifying hostile speech have been conducted, the vast majority have only covered the English language. Recent instances in Bangladesh, however, have led to the emergence of inflammatory speech in a variety of languages. Therefore, it is crucial to address this type of harmful material. Unfortunately, Bangla hate speech detection on social media sites such as Facebook and YouTube has been hampered by a lack of available public Bangla datasets. Although some datasets are available online, they are sparse, poorly sequenced, and lack necessary data types. As a means of filling this void, we have compiled a new dataset consisting of 8600 user comments from Facebook and YouTube, which we have divided into the following five categories: sports, religion, politics, entertainment, and others. Following that, we used five distinct models to perform a massive study of abusive language in Bengali. After testing a number of different models, we found that the BERT model had the highest accuracy of 80%. The availability of this dataset greatly aids our contribution to the study of identifying hate speech in Bengali. The same models have also been run on an existing dataset of 30,000 records, where we achieved an accuracy of 97%.

The issue of gender-based abuse is a prevalent concern in today’s society. As technology and social media platforms have become increasingly ubiquitous, these platforms have also become a breeding ground for abusive language and harassment, particularly toward women. In this study, we utilized machine learning techniques—logistic regression, decision tree, random forest, K-nearest neighbors, support vector machine, and Naïve Bayes—to classify abusive text based on gender. The considered dataset in this research comprised comments and posts from various social media platforms, which were preprocessed before being subjected to classification. Experimental analysis revealed that support vector machine demonstrated superior performance in terms of precision, recall, accuracy, sensitivity, and specificity indicating its potential effectiveness in identifying and filtering out gender-based abuse from social media platforms. The findings of this study suggest that machine learning techniques can play a critical role in combating gender-based abuse and harassment online.

The electrification of roads has been identified as an important step toward reducing carbon emissions and promoting sustainable transportation. However, the traditional payment systems for road usage are not well-suited for electrified roads, as they do not take into account the dynamic nature of the charging process. In this paper, we propose a payment system for electrified roads based on blockchain technology and the Delegated Proof of Stake (DPoS) consensus algorithm. Our blockchain-based payment system is designed to provide a secure and transparent platform for the exchange of value between road users and charging stations, while also ensuring fast and efficient payment processing. The proposed payment system addresses several key challenges faced by traditional payment systems, including the need for real-time payment processing, secure and transparent payment tracking, and the ability to handle microtransactions. The DPoS consensus algorithm is used to maintain the integrity of the blockchain and ensure that only valid transactions are included in the ledger.

Assuring the authenticity and seamless transmission of digitally stored student records is one of the main challenges for today’s educational institutions. Higher education institutions are willing to take necessary precautions to prevent the manipulation of these digital records. The process of validating these records can be facilitated by employing blockchain technology that ensures the authenticity, transparency, and immutability of digital data. This will also lessen the significance of third-party document authentication. In this study, we proposed a novel framework for efficiently and securely storing, verifying, and sharing student academic data using blockchain technology in order to prevent potential data fabrication and tampering. Here, we have used Keccak-256 secure hash algorithm for generating a unique and identical 256-bit hash value for a particular student record. The data integrity will be guaranteed by this identical hash value. Overall, this study offers a cost-effective, reliable, and safe data transfer mechanism by utilizing blockchain’s key features, such as digital access rule, immutability, transparency, and self-sovereignty.

Industries dealing with chemicals face a high risk of catastrophic accidents as they handle large amounts of explosive, flammable, and toxic chemicals during storage, transportation, and processing. Common and frequent accident scenarios include fire and explosion, which can trigger a chain of accidents due to their interactive nature. It is, therefore, crucial to have an effective risk evaluation technique. In this study, using a fuzzy inference system, the proposed system can estimate the risk of storing and processing hazardous chemicals. The chemicals’ inventory size, explosiveness, and flammability are used as the input variables, and risk as the output variable. The defuzzified risk value indicates the damage it can cause if accidents happen. Depending on the calculated risk value, we can also determine the precaution level to prevent the accident from happening. By maintaining proper security arrangements, many lives can be saved.

To empower women’s safety, this study presents a framework for real-time monitoring and warning using IoT-based wearable devices. Women’s safety is a major problem throughout the world, and the usage of wearable gadgets to solve this issue is becoming increasingly popular. To detect possible risks and warn users and emergency services in real time, the proposed system employs a mix of sensors, communication technologies, and frontier technologies. A unified dashboard for data analysis and management is also included in the system. The article goes into the technical components of the proposed framework as well as its possible influence on women’s safety. Overall, the suggested framework might help to create a safer atmosphere for women, giving them more confidence and security in their everyday lives.

Accurate and timely analysis of chest CT images is crucial for effectively diagnosing and treating a wide range of respiratory, cardiovascular, and infectious diseases, making it a vital component of modern medicine. Manual interpretation of chest CT can be time-consuming, prone to subjective variability, and potentially error-prone, highlighting the need for automation to improve efficiency and accuracy in analyzing large volumes of images. Researchers are increasingly using convolutional neural networks (CNNs) for chest CT image analysis due to their ability to learn complex features and patterns from large datasets. Despite significant advancements in analytical techniques, the challenge of determining which regions of interest to focus on and how to assign appropriate levels of importance to various features during chest CT image analysis remains an ongoing concern in the field. Our paper introduces a novel approach, utilizing soft and channel attention mechanisms in conjunction with an improved version of EfficientNetB0 to efficiently extract and prioritize critical features essential for accurately detecting and diagnosing lung-related ailments from chest CT scans. We have evaluated the proposed approach on a large dataset of CT scans, with the objective of accurately identifying COVID-19, non-COVID-19, and Community-Acquired Pneumonia (CAP) cases. Experimental results demonstrated exceptional performance, with achieved accuracy of 99.41%, surpassing other state-of-the-art methods in the field.

A report by world quality showed Bangladesh was ranked the second most air polluted country in 2019. The PM2.5 intensity on typical was 83.3 µg per cubic meter. Forests and trees are being cut down for creating buildings. Deforestation impacts badly the air quality of Bangladesh. Another reason for air pollution is overpopulation. Dhaka is one of the most overpopulated cities in Bangladesh. Air quality in Dhaka is 200 micrograms per cubic meter. In this research, we’ve collected some of the most common and dangerous gases, their toxicity, and their effect on the human body. For a healthy atmosphere air quality should be in the range of Nitrogen (N2) 78.084, Oxygen (O2) it is 20.947, Argon (Ar) − 0.934, and Carbon dioxide (CO2) 0.0314 and the others. If the percentage is high then the air becomes polluted. In this paper, we show the main gases in the Dhaka area. We collected the data for CO2, CO, SO2, O3, NH2, and PM2.5 in Air for awarding about the pollution. The objective is to create public awareness through mobile applications or show air pollution’s impact on everyday life. This paper suggested working with govt. or non-govt. environmental organizations to carry out necessary steps.

The incidence of human falls has emerged as a growing public health concern, particularly among the elderly and individuals with disabilities. Fall detection has assumed a vital role in healthcare research, aiming to mitigate the adverse consequences of falls, such as severe medical complications, prolonged treatment, hospitalization, and potential permanent disabilities. Fall detection methods encompass auxiliary equipment-based and computer vision-based approaches, which have gained prominence due to the increasing effectiveness and resilience of the Internet of Things (IoT). Nevertheless, computer vision-based methods give rise to significant privacy concerns, as monitoring human movements may entail capturing sensitive personnel images. To address this concern, this paper proposes a privacy-preserving approach for fall detection, utilizing Federated Learning in conjunction with Transfer Learning to train models. This study implemented various well-known pre-trained models, including VGG16, VGG19, InceptionV3, InceptionResNetV2, and Classic CNN, in a federated environment using a dataset comprising 30 videos. Among these models, InceptionResNetV2 achieved the highest test accuracy of 97.38%.

This paper emphasizes the critical importance of effective plant growth monitoring to optimize environmental conditions and ensure high crop yields in plant factories. The study aims to introduce an IoT-based plant growth monitoring system integrated with the capabilities of the Microsoft Azure cloud platform. To achieve this, IoT devices are strategically deployed in a hydroponic system for real-time data collection on crucial environmental factors and nutrient solutions. Moreover, a separate set of IoT devices has been employed to regulate the nutrient pump and illumination for enhanced control. In this comprehensive research, multiple variables, encompassing temperature, humidity, light intensity, pH, electric conductivity, total dissolved solids, and nutrient temperature are systematically gathered and stored in Microsoft Azure SQL database. Additionally, the inclusion of a camera to capture regular plant images further enriches the dataset. The primary focus of this research centers around the design and implementation of an innovative IoT-based system tailored specifically for data collection in hydroponic farms. By presenting an expansive overview of the design and potential impact of IoT-based systems in this field, the paper proves to be an invaluable resource for researchers and practitioners keen on exploring IoT monitoring and control applications in smart farming systems. As part of future work, the accumulated data will be leveraged for advanced analysis utilizing machine learning techniques, with the ultimate goal of unveiling concealed patterns governing plant growth. This endeavor promises to open up new possibilities for further advancements in controlled environment agriculture and sustainable food production.

The integration of drone technology, IoT, and AI can revolutionize smart town surveillance systems by reducing costs and improving response times to crime. The development of human-friendly drones and IoT technology enables automated city supervision, while the integration of AI allows for immediate reports and actions, overcoming human limitations. Drones can continuously monitor different areas, detecting crime hotspots and providing a substitute for city policies. This smart surveillance system can improve emergency management and ensure a regulated and competent surveillance management approach. Ultimately, this technology can enhance the safety and security of smart towns, making them more effective, efficient, and reliable. The framework of the suggested intelligent surveillance system relies on the utilization of real-time data collection and analysis, integrating drone technology, IoT devices, and AI. The research showcases the system’s efficacy in identifying and addressing criminal activities, making it an invaluable resource for law enforcement agencies. By combining these technologies, surveillance management can become more dependable and effective, ultimately contributing to the advancement of safer and smarter cities. The paper presents a comprehensive outline of the framework and its execution, providing a blueprint for future researchers and stakeholders to replicate and expand upon this endeavor. In this study, we present a superior surveillance system framework that is also implementable.

The lymphatic, bone marrow, and blood systems are all affected by hematological malignancy, also known as blood cancer. Early detection is essential for improved blood medical care and better patient outcomes. In the recent past, deep learning algorithms have developed as useful tools for the analysis and diagnosis of medical images. Deep learning algorithms are used in this study to introduce a cutting-edge technique for identifying blood cancer. Our method involves training a convolutional neural network (CNN) with a large dataset of blood cell data to identify the presence of cancerous cells. We compare our CNN-based approach’s effectiveness with some other CNN-based approaches and demonstrate that it is more effective at detecting blood cancer. We used a variety of preprocessing techniques to provide highly trainable data for our algorithms in order to do this. Five CNN-based algorithms—VGG-19, VGG-16, MobileNet, InceptionV3, and ResNet50—as well as healthy and fragmented data are used in this study. With an accuracy of 95%, ResNet50 achieved the highest accuracy. Our study’s results suggest that deep learning algorithms could be helpful in detecting blood cancer early, leading to a more precise diagnosis and course of treatment for this fatal condition.

Nowadays, breast cancer is one of the most ordinary cancers for women around the world. When breast cells grow and divide in an unrestrained way than forming a mass of tissue that call a tumor and then happen breast cancer. Many researchists have applied many applications to diagnose breast cancer. In this paper, a population-based metaheuristic named chemical reaction optimization (CRO) has been used to optimize the number of features. We have applied metaheuristic algorithms along with machine learning methods to predict breast cancer. From the experimental results, it can be observed that the SVM classifier gives the highest accuracy and f1-score among four classifiers (SVM, XGBoost, random forest, and decision tree). The result of the comparison showed that both the f1-score and accuracy are better than the related methods. To find out the best results on the detection of breast cancer using chemical reaction optimization (CRO) and a minimal number of features is the main target of our paper. We use SVM, decision tree, XGBoost, and random forest as the classifiers. For the experiment, the UCI machine learning dataset has been used. We have tried to find the best results in terms of measurement metrics using the proper technique.

Diseases affecting heart and blood arteries are often referred to as cardiovascular diseases. It is crucial to predict and diagnose accurately cardiovascular disease in order to provide proper treatment. For detection of hidden patterns of cardiovascular diseases from massive amounts of data, deep learning techniques are commonly used. Different predictive models are used in health care to anticipate patient’s future ailments by learning from their prior data. This study seeks to create an intelligent agent for making such predictions automatically. This research mainly contributed to discover the most important attributes that best define the relationship with the target attribute and to find a proper algorithm for the prediction of cardiovascular disease. Using methods like Pearson correlation analysis and permutation feature importance along with effective data preprocessing techniques, the most crucial characteristics of cardiovascular diseases have been determined. A consolidated dataset from IEEEDataPort was used in this study. This study achieved 95% accuracy in predicting cardiac disease using Artificial Neural Network (ANN). This research will help healthcare people to predict a person’s cardiovascular disease at early stage. It will also help researchers discover more ways to anticipate cardiovascular illnesses.