PUBLICATIONS
Journal Articles
- Lexicon and Deep Learning-Based Approaches in Sentiment Analysis on Short Texts.
Social media is an essential component of our personal and professional lives. We use it extensively to share various things, including our opinions on daily topics and feelings about different subjects. This sharing of posts provides insights into someone’s current emotions.
In artificial intelligence (AI) and deep learning (DL), researchers emphasize opinion mining and analysis of sentiment, particularly on social media platforms such as Twitter (currently known as X), which has a global user base. This research work revolves explicitly around a comparison between two popular approaches: Lexicon-based and Deep learning-based Approaches.
To conduct this study, this study has used a Twitter dataset called sentiment140, which contains over 1.5 million data points. The primary focus was the Long Short-Term Memory (LSTM) deep learning sequence model. In the beginning, we used particular techniques to preprocess the data. The dataset is divided into training and test data. We evaluated the performance of our model using the test data.
Simultaneously, we have applied the lexicon-based approach to the same test data and recorded the outputs. Finally, we compared the two approaches by creating confusion matrices based on their respective outputs. This allows us to assess their precision, recall, and F1-Score, enabling us to determine which approach yields better accuracy. This research achieved 98% model accuracy for deep learning algorithms and 95% model accuracy for the lexicon-based approach.
- Enhancing Bangla Fake News Detection Using Bidirectional Gated Recurrent Units and Deep Learning Techniques.
The rise of fake news has made the need for effective detection methods, including in languages other than English, increasingly important. The study aims to address the challenges of Bangla which is considered a less important language.
To this end, a complete dataset containing about 50,000 news items is proposed. Several deep learning models have been tested on this dataset, including the bidirectional gated recurrent unit (GRU), the long short-term memory (LSTM), the 1D convolutional neural network (CNN), and hybrid architectures.
For this research, we assessed the efficacy of the model utilizing a range of useful measures, including recall, precision, F1 score, and accuracy. This was done by employing a big application.
We carry out comprehensive trials to show the effectiveness of these models in identifying bogus news in Bangla, with the Bidirectional GRU model having a stunning accuracy of 99.16%. Our analysis highlights the importance of dataset balance and the need for continual improvement efforts to a substantial degree. This study makes a major contribution to the creation of Bangla fake news detecting systems with limited resources, thereby setting the stage for future improvements in the detection process.
Conference Proceedings
- Aspect-Based Sentiment Analysis of Amazon Product Reviews Using Machine Learning Models and Hybrid Feature Engineering.
While sentiment analysis is a popular and significant research trend, aspect-based sentiment analysis (ABSA) requires more focus from researchers. The customer reviews of headphones and Bluetooth devices on Amazon are the main subject of this study.
Several machine learning (ML) algorithms are used in the study, including Support Vector Machine (SVM), k-nearest Neighbors (KNN), Random Forest (RF), Naive Bayes (NB), Decision Tree (DT), and Logistic Regression (LR). Additionally, a hybrid feature engineering technique combining TF-IDF (Term Frequency-Inverse Document Frequency) and word n-gram is applied, specifically utilizing word n-gram (1,4) in conjunction with TF-IDF.
The results of evaluating these methods showed that, with an accuracy of 91%, SVM with hybrid word n-gram (1,3) produced the best outcomes. The research dataset exhibits imbalance, which is addressed by using the Matthews Correlation Coefficient (MCC) as an additional performance metric. This results in a score of 0.77.
The results show that aspect-based sentiment analysis is effective in gaining insightful information from customer reviews of headphones and Bluetooth devices on Amazon. The SVM algorithm and the designated hybrid feature engineering technique perform better than the others.
- Supervised Machine Learning Approaches to Identify the False and True News from Social Media Data.
The increasing number of online communities and social platforms like Twitter and Facebook has facilitated a level of information sharing never before seen in human history. Consumers are generating and sharing more data than ever before thanks to the proliferation of social media platforms, and some of it is deceptive and has no basis in reality.
Automatically determining whether a text contains misleading data or misinformation is difficult. Before passing judgment on the accuracy of a piece, even a subject matter specialist needs to look into a number of different angles. Here, this research presents machine learning strategies for distinguishing between false and genuine news.
In this study, we have collected data by web scraping and employed several different techniques to train a collection of machine learning algorithms and then compare how well they perform on our datasets. In this work, five machine learning algorithms have been applied to find the best algorithms. After evaluating the model, the research found that the decision tree achieved the best 99.84% model accuracy from this study.
- Computer Vision Based Bangla Sign Language Recognition Using Transfer Learning.
Human societies have relied on communication since ancient times, yet verbal communication poses significant challenges for deaf and hard-of-hearing individuals, necessitating reliance on sign language. Recent advancements, notably through deep learning models, have propelled research in this area.
Acknowledging the need for further progress, particularly in minority languages like Bengali, our study aims to develop a method for image-based Bengali sign language detection. We constructed two independent convolutional neural network (CNN) models, InceptionV3 and Xception, leveraging data from diverse sources.
Remarkably, the Inception V3 model achieved an accuracy of 97%, while the Xception model surpassed expectations with an accuracy of 99.50%. These results signify substantial progress, demonstrating the efficacy of deep learning architectures, especially the Xception model, in accurately interpreting Bengali sign language. Our study demonstrated how transfer learning, when combined with careful optimization, may yield remarkable outcomes in Bengali sign language recognition, which are further enhanced by data augmentation methods.
- Accurate Thyroid Disease Detection with Ensemble Learning Models.
Thyroid disease is a medical condition that prevents a person's thyroid gland from producing enough hormones and significantly affects a person's health. Early detection of thyroid disease and timely treatment can play an essential role in improving the health of thyroid patients.
Our research paper helps to detect this disorder called thyroid disease and suggests an automated system as a complement to conventional treatment for healthcare. We conduct a wide range of studies, including exploratory data analysis, along with processing our data sets using various machine learning techniques. We gain a comprehensive understanding of data distribution and target patients' characteristics through data visualization.
Later, we use a wide range of different machine learning models, such as Decision Tree, Logistic Regression, Support Vector Classifier, Random Forest, Gradient Boosting, Adaptive Boosting, Bagging and K-Nearest Neighbor, etc. to get accurate predictions. Next, we combine the base models and apply various ensemble learning techniques to get even better performance.
Extensive experiments have been conducted in our research, in which the ensemble of the Decision Tree + Logistic Regression model has shown its robustness and proven its effectiveness in thyroid detection by achieving 99.90% accuracy and maintaining a balance between precision and recall. Systematic analysis of different models provides a broad idea of the strengths and weaknesses of the models.
- Deep Learning Based Forecasting Models of Dengue Outbreak in Bangladesh.
In 2023, Bangladesh experienced its most severe dengue outbreak since 2000, necessitating improved prediction methods. This study employs diverse machine-learning techniques to forecast dengue fever occurrences in Bangladesh, aiming for enhanced accuracy and proactive public health measures.
Unlike previous research focusing narrowly on specific variables, this study incorporates a wider range of factors and algorithms. Three models: Gated Recurrent Units (GRU), Recurrent Neural Network (RNN), and Long Short-Term Memory (LSTM) were evaluated using data from January 2000 to December 2022. Official patient data from the Ministry of Health and Family Welfare and meteorological data from the Bangladesh Meteorological Department were utilized.
LSTM exhibited the highest accuracy at 87.98%, surpassing GRU at 79.81%, while RNN trailed at 57.18%. This underscores LSTM’s efficacy in predicting dengue cases in Bangladesh, offering valuable insights for proactive disease management and public health interventions.
- Enhancing Crop Management: Ensemble Machine Learning for Real-Time Crop Recommendation.
The agricultural industry is essential to the world’s food production, and it is critical to use cutting-edge technologies to increase crop productivity. We provide a revolutionary Crop Recommendation System (CRS) that utilizes cutting-edge technology to maximize crop output in response to the pressing need for improvement.
Our study incorporates real-time monitoring of soil conditions, made possible by a custom hardware configuration that includes sensors for temperature, humidity, phosphorus, potassium, nitrogen, and pH measurements. First, we assembled a large dataset with 22 kinds of agricultural production components.
Using many machine learning models, such as ensemble methods and baseline classifiers, we were able to classify crops with an astounding 99% accuracy rate. With the application of these insights, the CRS provides customized recommendations through an easy-to-use user interface for appropriate crops under particular climatic conditions.
Our system’s novel integration of AI-driven decision-making and hardware sensing capabilities promises to transform crop management techniques and provide agricultural stakeholders with useful insights.
- Early-Stage Diabetes Risk Prediction Using Supervised Machine Learning Algorithms.
Diabetes is a common health problem worldwide; it is especially pervasive in Bangladesh. The condition manifests in a person when his blood sugar is consistently high. It also contributes to other health problems like blindness, renal failure, heart attack, and stroke. If you know about the early stage, you can take charge and maybe save someone's life. Sadly, this illness is spreading rapidly.
The purpose of this research was to quantitatively evaluate the effectiveness of many widely used Machine Learning methods. The medical field is only one area that has benefited greatly from recent advancements in Machine Learning technology. Machine learning algorithms come in a wide variety.
Nevertheless, in this research we employ five well-known machine learning algorithms to determine performance metrics: Gaussian Naive Bayes, Random Forest, Support Vector Machine, Logistic Regression, and the Decision Tree classifier. Using real data from diabetic patients in Bangladesh, these algorithms were developed and evaluated. There are 3837 patient records in the dataset, 3057 of which correspond to affected cases and 396 were normal.
Out of 5 different machine learning algorithms, Random Forest achieved the highest 98% accuracy.
- Machine Learning Approach Analysis for Early-Stage Liver Disease Prediction.
Liver disease can cause death in millions of individuals worldwide. Early detection and accurate diagnosis improve patient outcomes and reduce mortality. Machine learning algorithms can detect and diagnose liver diseases more accurately and affordably.
To test machine learning methods for early liver disease diagnosis, this study created a prediction model that distinguished those with liver illnesses from those without. Machine learning algorithms studied included Extra Tree, XGBoost, Random Forest, CatBoost, LogitBoost, Gradient Boosting, AdaBoost, and KNN.
By using oversampling methodologies and assessing metrics like accuracy, recall, precision, and F1-score, the Extra Tree algorithm was found to be the most effective way for early liver problem detection with 99% accuracy. Undersampling, oversampling, and SMOTE reduced class imbalance. This problem is widespread in many machine learning applications.
Machine learning algorithms may improve liver disease detection, early intervention, and healthcare costs. However, data protection, ethics, and healthcare professional training and understanding must be adequately addressed. This research is a major step toward a more accurate and practical liver disease diagnostic tool that could help individuals stay healthy and prevent liver illnesses.
Book Chapters
- Implied Communality Deficit and Heroism.
Implied Communality Deficit: The perceived lack of stereotypically feminine characteristics that can occur when females exhibit stereotypically masculine traits or behavior. Often results in negativity directed towards the female.
There have been many studies on gender biases and stereotyping across fields such as psychology, sociology, and neuroscience. These studies have shown that gender biases and stereotypes exist and can have significant impacts on individuals and society. Stereotypical expectations exist for women to exhibit nurturing and communal traits (e.g., being kind, empathetic, and understanding), rather than powerful or agentic traits (e.g., acting aggressively or dominantly, or being achievement-focused). These stereotypes are widespread and persist despite efforts to counteract them (Heilman 2012; Heilman and Caleo 2018). Children as young as 2 years old understand that there are different expectations for boys and girls (Kane 1996).
Under Review
- Work in progress!