Road networks in a locality can be fully planned or fully unplanned. It can even be a mixture of both. New roads to be made need to accommodate existing structures and should also account for future structures or areas that will be built. However, manual planning might sometimes result in suboptimal road structures, as measured by vegetation plantability and road width. The system we propose takes into consideration the metrics that have been mentioned through a rule-based road network generation system that uses Minimum Spanning Trees and the A* algorithm. Unlike existing deep learning-based models, our rule-based approach is explainable and focuses on efficient road network generation instead of mimicking existing networks. The proposed method maintains identical paths while achieving faster runtime by completing generation in 49.45

Extractive Question Answering (EQA) in the biomedical domain presents significant challenges due to complex medical terminology and intricate relationships between entities. Existing systems often perform suboptimally compared to their counterparts in general domains. Current biomedical QA models struggle with accurately identifying named entities and extracting meaningful relationships within medical contexts, thereby limiting their practical utility. This study introduces a novel transformer-based approach for biomedical EQA that incorporates Named Entity Recognition (NER) to identify key medical terms, Relation Extraction (RE) to understand their interconnections, and a BiLSTM layer to enhance contextual comprehension. These components are integrated within an ensemble framework that leverages pretrained biomedical language models. Our methodology systematically evaluated six transformer encoder models and identified Bio+ClinicalBERT as the optimal foundation for developing three progressively advanced biomedical EQA models. The final model — Bio+ClinicalBERT_NER_RE_BiLSTM — incorporating ensemble learning techniques, demonstrated substantial performance improvements: an F1 score of 91. 69%, exact match of 88. 35%, and lenient accuracy of 0.84, marking a 5. 45% performance gain over baseline methods. These results highlight the model’s superior ability to navigate complex biomedical language and extract accurate information from medical texts. The developed system directly benefits biomedical researchers, clinical practitioners, medical students, and healthcare professionals by delivering accurate and rapid access to precise information within extensive medical literature. By bridging the gap between intricate medical texts and effective information retrieval, this work significantly advances automated biomedical knowledge extraction and supports evidence-based medical decision-making.

Bangla Speech Emotion Recognition (SER) is crucial for systems that respond to human emotions, such as in healthcare, education, and safety. However, because to poor datasets, overfitted models, and a dearth of research on adaptive learning and real-world deployment, the discipline has not made satisfactory progress. Most prior research trains on a single small dataset, uses overly complex models, and rarely tests edge feasibility, resulting a gap between research accuracy and practical use. Our work addresses this through four experiments: comparing single versus merged dataset training, large handcrafted versus MFCC-only feature sets, the effect of noise on robustness, and architectures for reducing overfitting while maintaining efficiency. The final system uses MFCC features with a CNN–BiLSTM model, balancing local spectral and temporal patterns, with dropout and careful validation to prevent overfitting. It reaches 82

Bengali Sign Language (BdSL) is a vital communication tool for the Deaf community in Bangladesh, relying on hand, facial, and body movements. This research leverages advanced machine learning and deep learning to develop a real-time BdSL recognition system using a Kaggle-sourced dataset of 1200 RGB images resized to 128 × 128 pixels. Six deep learning models were tested, with MobileNetV2 achieving the highest accuracy of 99.89

Access to legal information presents a significant impediment in developing nations like Bangladesh, where a substantial portion of the populace lacks legal literacy and access to professional counsel. This paper introduces a Multi-Agent Retrieval Augmented Generation (RAG) system specifically designed for Bangladeshi legal information retrieval. The system employs a novel two-agent architecture: a Clarification Agent that refines user queries through Large Language Model (LLM)-driven clarity assessment and interactive questioning, and a primary RAG Agent. The RAG Agent executes a comprehensive process involving document retrieval from a meticulously curated legal corpus, relevance grading, conditional web search, and final answer synthesis. This corpus, compiled from official government websites, legal blogs, and scholarly articles, ensures accurate representation of the Bangladeshi legal landscape. The multi-agent RAG pipeline, through structured collaboration between these autonomous agents, is designed to improve the relevance and contextual accuracy of responses to complex legal queries. A conversational interface built with Streamlit facilitates accessible, real-time user interaction. The Agentic RAG model outperforms baseline and general-purpose legal bots (using GPT-4o-mini and Gemma 3B) on 20 diverse legal queries, showing better contextual relevance, accuracy, and fewer hallucinations. It also supports cost-effective local deployment via the Ollama framework. This research demonstrates a scalable and potentially cost-effective approach to democratizing legal knowledge, which is particularly vital in low-resource settings such as Bangladesh.

The rapid proliferation of smartphones and location-based services has accelerated the development of intelligent systems in various fields, such as tourism, ride-sharing, delivery and shipping services, user engagement, etc. In this context, next Point of Interest (POI) recommendation plays a vital role in predicting where a user will go based on their visit sequences or check-in histories. Recent advances in sequential modeling have led to more accurate recommendations; however, capturing complex interactions across spatial, temporal, and categorical dimensions remains challenging. They are important because they affect users' desire to visit a place at a certain moment. Many current models are inadequate because they either ignore the important influence of POI categories, reduce spatial coordinates to broad region IDs, or conduct temporal analysis exclusively using recurrent neural networks. This paper presents an innovative deep-learning framework Spatio-Temporal Adaptive Fusion Transformer (STAFT), systematically addressing these challenges by seamlessly integrating continuous spatial encoding, category-aware context weighting, and learning sequential pattern. Our model employs Learnable Random Fourier Features (RFF) to transform raw spatial features into richer, high-dimensional embeddings. It also emphasizes relevant POI categories utilizing Category Aware Attention and a Temporal Convolutional Network (TCN) to efficiently capture local temporal dependencies. The Transformer Encoder architecture is used at the core to produce the final prediction. Empirical results from experimentation on real-world datasets demonstrate that our approach performs better than existing models by 4-9% on commonly used metrics.

Malware is malicious software employed to disrupt, disable, or invade operating systems and user hardware without consent. Android, the dominant smartphone operating system, faces enormous malware threats as smartphone usage grows exponentially. Traditionally employed pattern-based malware detection approaches have limitations, which has created interest in artificial intelligence (AI)-based approaches. In this paper, we worked with an Android malware dataset of 355,630 network flow records to build a strong AI-based detection model. The dataset consists of four classes: Android Adware, Scareware, SMS Malware, and Benign, with severe class imbalance. This research was conducted through exploratory data analysis (EDA) and strict feature selection techniques to identify significant features. We applied the synthetic minority over-sampling technique (SMOTE) to address the class imbalance in an optimized ensemble machine-learning pipeline comprising Random Forest, XGBoost, LightGBM, ExtraTrees, and CatBoost classifiers. We also contrasted the models based on fivefold Stratified cross-validation, hyperparameter tuning, and feature conversion. Among the classifiers, the hyperparameter-tuned Random Forest model achieved the best performance of an F2-score of 92

Stock price forecasting remains challenging due to the non-linear, noisy, and non-stationary nature of financial time series. Although LSTMs and Transformer-based models have improved sequential modeling, their ability to scale efficiently to long financial sequences remains limited. Recently, selective state space models such as Mamba have emerged as efficient alternatives to self-attention, offering attention-like performance with linear computational complexity. In this study, we systematically evaluate multiple Mamba-augmented Transformer architectures for stock market price forecasting. We further propose CrossMamba, a novel architecture that models cross-sequence interactions between encoder and decoder representations using a causal Mamba block. Experiments on multiple S&P 500 and Yahoo Finance stocks show that CrossMamba achieves superior short-horizon performance with 5-day input windows (R2 up to 0.963), while Hybrid Bi-Mamba performs best for longer horizons, achieving the lowest MAE of 0.67 for 10-day forecasts. Compared with advanced Mamba-based and Transformer baselines, the proposed models achieve competitive accuracy while maintaining substantially improved computational efficiency. These results highlight the effectiveness of Mamba-augmented Transformers as scalable architectures for financial time series forecasting.

To improve the performance of policy gradient agents for stock recommendations on the Dhaka Stock Exchange, this work explores the use of entropy regularization of the advantage critic function instead of traditional reward shaping. Policy gradient methods, while demonstrably effective for optimizing sequential decision-making tasks, can exhibit instability and converge to suboptimal solutions. Entropy regularization offers a potential remedy by promoting exploration and discouraging premature convergence. This research explores the inherent trade-off between exploration and exploitation within the context of the DSE’s specific market dynamics. The objective is to develop and evaluate adaptive strategies that leverage entropy regularization to improve the robustness and profitability of policy gradient-based recommendations.

Speech Emotion Recognition (SER) in Bangla is a growing yet underexplored area, despite the language being spoken by over 230 million people. While state-of-the-art models such as CNN-Transformer have achieved success in highresource languages, their application to Bangla SER faces critical challenges: scarcity and imbalance of datasets, overfitting of complex architectures on low-resource data, and limited attention to deployment efficiency and adaptability in real-world settings. This paper addresses these gaps through four incremental experiments that guided our design choices. We show that: adding multiple Bangla datasets improves generalization over single-dataset training, extensive handcrafted feature combinations do not outperform MFCC alone, noise-free datasets provide higher robustness, and a hybrid CNN-BiLSTM architecture alleviates overfitting and achieves state-of-the-art accuracy for Bangla SER. Our proposed model achieves 82 % accuracy across seven emotions, outperforming prior Bangla SER baselines while remaining lightweight enough for deployment on Raspberry Pi. Furthermore, we extend SER beyond stand-alone classification by integrating it with Large Language Models (LLMs) for emotionally adaptive dialogue, and by introducing incremental learning for future adaptability. This positions our work as one of the first comprehensive studies to combine accuracy, efficiency, and adaptability in Bangla SER.

Survey data often lacks explicit labeling, making it difficult to apply advanced machine learning techniques for predictive analysis. The study proposes an approach to systematically converting and binarizing survey data, addressing the lack of labeled data in text-based surveys. The survey dataset contains responses from 309 male fertility patients in Malaysia, collected over a one year period from November 2021 to October 2022. The survey data includes information on demographic, occupational stress, job satisfaction, and job performance. A comprehensive comparison is conducted between the traditional normalization method and the proposed method by utilizing eight machine learning and five large language models. Our proposed method demonstrates superior performance with an accuracy improvement of 9.68

In Natural Language Processing (NLP), textual data is foundational, yet it presents substantial challenges, especially for under-resourced languages like Bengali. The complexity and volume of Bengali textual data require sophisticated data cleaning techniques. Traditional methods often neglect critical contextual information essential for effective textual analysis. This study highlights the need for context-aware data cleaning, a methodology that maintains linguistic context while removing noise. The study compares context-aware and traditional data cleaning approaches tailored for Bengali text to improve the performance of contextual transformer-based models. Conventional techniques in this study include symbol and punctuation removal, stop-word elimination, stemming, and removing HTML tags or URLs. In contrast, context-aware techniques involve spelling correction, tagging HTML and URLs, preserving punctuation and emojis, and selectively removing less important words using TF-IDF. The current initiative assesses the impact of these strategies through rigorous dataset curation and extensive training in machine learning, deep learning, and transformer-based models on four prominent Bengali datasets: BEmoC, SentNoB, UBMEC, and EmoNoBa. Results show context-aware data cleaning significantly outperforms traditional methods, particularly in enhancing transformer-based model performance. The developed context-aware data cleaning pipeline integrates various techniques, achieving a baseline accuracy improvement of up to 4

Dengue is a global health crisis and is spread by Aedes aegypti mosquitoes. This paper presents ‘Stinger’, a 3D asymmetric multiplayer serious game to address dengue awareness gaps in rural Bangladesh through culturally relevant gameplay. The game has interactive mechanics that simulate real-world dengue transmission dynamics, teach players to identify and eliminate mosquito breeding sites and adopt preventive behaviors. Setting in a rural Bangladeshi environment, the game has multiplayer functionality to model community-level disease spread and collective prevention strategies. Early results show significant knowledge retention and behavior adoption among players, so game-based learning works for public health education. The game provides a framework for developing culturally adapted serious games to bridge the gap between awareness and action in disease prevention.

Aerial scene classification using satellite and drone imagery is vital for applications like environmental monitoring and urban planning, but the high computational demands of Convolutional Neural Networks (CNNs) limit their real-time use on resource-constrained platforms. Additionally, these models often lack global context awareness, relying mainly on small 3 x 3 kernels that capture fine details but miss broader spatial relationships. While large models can learn complex patterns, lightweight models struggle without adequate data and preprocessing. To address these challenges, we propose a lightweight CNN classifier optimized for fast, real-time aerial scene classification. Inspired by the Inception network, our architecture integrates multi-scale convolutional filters (1 x 1, 3 x 3, 5 x 5, and 7 x 7) to capture both local and global context. To reduce computational overhead, we incorporate Depthwise Separable Convolutions (DSC). To overcome the limitations of a compact model, we apply 5-fold semantic-preserving data augmentation and Contrast Limited Adaptive Histogram Equalization (CLAHE) to enhance feature visibility. While most state-of-the-art models require at least a million parameters, our lightweight CNN achieves an impressive 97.62 % accuracy on the UC (University of California) Merced Land Use Dataset using only 56,293 parameters. Furthermore, we use Gradient-Weighted Class Activation Mapping (Grad-CAM) to assess augmentation, CLAHE, and our architecture's influence on feature attention. This study demonstrates that effective preprocessing with CLAHE and extensive augmentation can narrow the gap between lightweight CNNs and complex models. These results support the use of efficient models for real-time, resource-constrained aerial scene classification, promoting sustainable and accessible Artificial Intelligence (AI) in remote sensing applications.

This study examines the recall value ascendancy of credit card fraud detection performed by Deep Q-Network agents, which is 86 % as opposed to 74 % using ensemble techniques like Random Forests. An enhanced DQN agent is suggested, which uses a two hidden layer neural network to update the target frequency and explicitly penalizes false negatives using an L1-based reward shaping linked to the temporal-difference (TD) error. Additionally, by learning from replay buffers utilizing Synthetic Minority Over Sampling Technique (SMOTE) base oversampling, the DQN agent shows improved adaptation to changing fraud patterns. This study provides insights into the creation of efficient DQN agents for fraud prevention and highlights the potential of Reinforcement Learning (RL) based techniques for increasing the probability of credit card fraud specificity.

Diabetes is a growing public health concern in Bangladesh, and it is one of the leading causes of mortality in the world. This study focused on a data driven machine learning framework for diabetes risk stratification and epidemic control based on the WHO NCD STEPS 2018 survey dataset, consisting of 8,185 respondents. The population is fairly divided by gender (53.5% female, 46.5% male) and residence (51.1% urban, 48.9% rural), and divided into age groups, with the largest $(29.9 \%)$ in the $27-37$ range. Clustering analysis showed that 56.1% of individuals are in the low-risk category, 38.1% in the medium risk category, and 5.8% in the high risk category. Logistic Regression (LR), K-Nearest Neighbor (KNN), Decision Tree (DT), Random Forest (RF), and Support Vector Machine (SVM) were used, and they were evaluated using Accuracy, Precision, Recall, and F1-score. Without resampling, Random Forest achieved the best performance (93.9% accuracy). And after applying SMOTE to addressing class imbalance, Random Forest again outperformed others (98.3% accuracy). By leveraging structured dataset and machine learning algorithms, this study shows it is possible to build an effective and efficient model for predicting the risk of having Diabetes.

In the modern era, where artificial intelligence (AI) intersects with computational literary studies, the analysis of poetry presents both challenges and opportunities. Rabindranath Tagore’s poetry holds immense literary and historical significance, yet understanding its contextual motivations and systematically categorizing it remains complex. This study pioneers a chronological classification of Tagore’s poetry while also exploring subject-based categorization through advanced computational methods. Utilizing an AI-driven approach that integrates machine learning (ML), deep learning (DL), and natural language processing (NLP), we evaluated three ML models with two embedding methods, four DL models with four embedding methods, and five large language models (LLMs). Our methodology achieved 85.99

Earth observation and remote sensing necessitate the use of aerial imaging, and unmanned aerial vehicles (UAVs) are widely utilized for this purpose. Although UAVs can be costly, using inexpensive components could result in issues such as noise, defocusing, and motion blur. Utilizing advanced deep learning algorithms, these images can be effectively restored. However, this necessitates the use of physical effort since each image must be restored manually. Our work introduces a new pipeline that can automatically categorize the degradation types of aerial images and subsequently apply the necessary restoration techniques on degraded images. A lightweight adaptive degradation classifier model was proposed to handle input from images of varying sizes and provide guidance to a image restoration model for achieving the desired task. The degradation classifier is a Fully Convolutional Network capable of categorizing images of any size into 3 distinct classes. It offers guidance to the subsequent image restorer. The restoration model is a pre-trained transformer-based model that efficiently restores different types of image degradations. In addition, we have implemented a pipeline loss to assess the performance of our complete system. To build the system, a dataset is created by extracting cropped patches from high-resolution images obtained through Google Earth. Subsequently, the clean patches were artificially deteriorated by applying different filters to replicate the image degradations. The degradation classifier achieved a mean accuracy of 96.71% across 6 different test scenarios, with a standard deviation of 2.97. The mean pipeline loss was 24.95, with a standard deviation of 1.79. Furthermore, we showcased the efficacy of our pipeline in a practical implementation involving aerial imagery, specifically in the domain of vehicle detection. We conducted training on an object detection model and then evaluated the detection performance on both degraded and automatically restored images. Our experiments demonstrate the efficient performance of our method in automatically restoring degraded aerial images, even though the restoration models were not specifically trained on aerial images. In addition, we examined current limitations and difficulties in creating such systems and offered comprehensive perspectives on future advancements.

In the modern age, with the help of social media, communication has become available for everyone. Offensive text is broadly used in social media to humiliate or threaten someone. Offensive text like a bully, trolling, threats, and sexual harassment are used to demotivate someone. We have gathered a dataset of 44,000 comments from social media. We use five different models: DistilBERT, Multilingual BERT, XLM-RoBERTa-base, XLM-RoBERTa-large, and BanglaBERT. Variations in various parameters, e.g., learning rate, dropout rate, training epoch, early stopping, and batch size, are made to get better results. From our proposed model, the XLM-RoBERTa-base shows the highest accuracy, 83.54%, whereas m-BERT provides the highest AUC value of 0.85.