Research & Publications

Peer-reviewed papers, conference proceedings, journal articles, and preprints focusing on geometric deep learning, graph neural networks, and formal systems verification.

Journal

1 publication

2024·IEEE Access

Classification of Freshwater Fish Diseases in Bangladesh Using a Novel Ensemble Deep Learning Model: Enhancing Accuracy and Interpretability

Effective disease management and mitigation strategies for fish diseases depend on timely and accurate diagnosis. In recent years, artificial intelligence methods—classification algorithms in particular—have become effective instruments for automating fish disease diagnosis. This paper presents two types of ensemble models: i) the baseline averaged ensemble (AE) model and ii) the novel Performance Metric-Infused Weighted Ensemble (PMIWE) model. By leveraging pre-trained models and novel ensemble techniques, we achieve a testing accuracy of 97.53%, corresponding precision, recall, and F1-score of 97%. We also bring about enhanced interpretability and trustworthiness using the Grad-CAM (Gradient-weighted Class Activation Mapping) explainable artificial intelligence (XAI) technique.

Abstract DOI

Conference

4 publications

2026·IEEE

Generative Adversarial Networks (GANs) for Crop Simulation and Enhancement: Exploring Synthetic Data for Agricultural Research

Real-world data collection in agriculture is not only highly expensive but also faces challenges due to seasonal variations and limited generalization capabilities for capturing unusual disease patterns and environmental changes, as is required for deep learning models. We propose a new GAN-based model aimed at crop simulation and disease data augmentation. A class-conditional Deep Convolutional GAN will be designed to generate high-resolution, biologically plausible crop disease images by jointly learning visual structures and class-specific disease characteristics. The real samples in our dataset will be combined with these generated synthetic samples to increase training for crop disease classification. The multi-class plant dataset available in PlantVillage is used in the experimental scenario, with classes for healthy and diseased crops. Extensive ablation experiments confirm that the classifier's performance can be enhanced by GAN-based training compared to state-of-the-art deep learning architectures such as ResNet50, DenseNet121, EfficientNetB3, and VGG16. The accuracy of the proposed model, ResNet50 + GAN, is about 98.65%. This shows not only higher precision levels but also a higher recall rate with an F1-score computed accordingly. This validates the realism of the synthetic images developed. This is an implication of the usefulness of synthetic data in addressing data scarcity and class imbalance in the domain of intelligent agriculture.

Abstract DOI

2026·IEEE

Solar PV Power Generation Forecasting Using an Attention-Based Bidirectional LSTM Network

To provide the reliability and efficiency of the solar energy system. The solar photovoltaic (PV) power generation should be predicted accurately. The generation of solar power is not a linear process and it changes with time depending on the environmental factors. Therefore, it is quite difficult to forecast solar power production properly. We, in this regard, introduce a new hybrid deep learning (DL) model, SP-BiLSTM-ATT, a composite of Bidirectional Long Short-Term Memory model and Attention mechanism to forecast the solar power generation. The model proposed is tested with the help of a solar power generation dataset that has been acquired via the Kaggle data repository. The performance of the proposed model is correlated to the performance of other well-known regression models, such as Decision Tree (dt), K-Nearest Neighbors (KNN), Random Forest (RF) and standard LSTM model. The model proposed has a much greater accuracy when compared to the conventional models with an R2 which is 0.95. The fact that the precision of the suggested model was steadily increased in all the experiments validates the reliability of the model.

Abstract DOI

2026·IEEE

Real-Time Credit Card Fraud Detection Using Graph Attention Networks with Cost-Sensitive Learning

Due to the dynamic expansion and greater efficiency of online payment systems, financial fraud has also become more prevalent and increasingly tech-savvy. For efficient and effective detection, rule-based and table-based cross-learning algorithms often do not account for the relationships and context involved in modern financial fraud. To tackle these problems, a novel real-time fraud detection system is presented that leverages both Graph Neural Networks and the edge perspective. During the system, the data is represented as a graph, leading to better understanding through relation-based learning. To recognize especially influential patterns, a tuned-up variant, the Graph Attention Network, is considered, targeting the discovery of fraudulent patterns through attention-based message transmission between connected entities. To better address class imbalance, a combined performance measure is used that combines the Synthetic Minority Oversampling Technique and Cost Sensitive Learning. During the experiments, a dataset was used comprising 284,807 samples, of which 0.172 % were fraudulent credit card transaction history cases. It is important to note that the proposed system for detection purposes will deliver highly efficient results, with a rapid response time, achieving 99.89 % accuracy.

Abstract DOI

2026·IEEE

Attention-Guided Dual-Energy X-Ray Image Analysis for Prohibited Item Detection Using YOLOv11 and Transformer Fusion

X-ray inspection has become an integral part of public safety screening systems for detecting prohibited items. However, this remains a challenging task due to severe object overlap, low inter-material contrast, and pronounced class imbalance in contraband detection. Dual-energy X-ray imaging provides enhanced material discrimination, but how to effectively exploit this complementary information toward realtime detection remains a challenge. The limitations discussed above are addressed by the proposal of an attention-guided dual-energy X-ray detection framework, known as AG-DEXDet. A transformer-based feature fusion module, combined with a YOLOv11-based one-stage detector, captures global contextual dependencies and improves robustness in highly congested scenarios. This is further combined with the Class-Specific Augmentation Framework to mitigate data imbalance problems across different restricted-item classes. The balance dualenergy X-ray dataset is used for testing the proposed model, comprising 12 contraband categories. The test results revealed that AG-DEXDet achieved 98.09% detection accuracy and a mean Average Precision of 0.9815 while maintaining realtime inference performance. Compared to traditional CNN- or region-based detectors, the proposed approach achieves higher detection accuracy and greater stability in complex screening environments. Effective Attention-Guided Fusion of Dual-Energy Representations toward Real-Time Security Inspection. These results clearly indicate that AG-DEXDet is highly feasible for implementation in practical X-ray screening systems. Future work will focus on extending the framework to larger-scale datasets and to multi-view X-ray imaging scenarios.

Abstract DOI

Preprint

1 publication

2026·SSRN- Preprint

A Predictive Framework for Identifying Fake Reviews on E-Commerce Platforms

Fake reviews in particular are degrading the trust between e-commerce and consumers, because theycan mislead them and misrepresent about the products. To determine this, we propose a hybrid fakereview detection framework with both the linguistic and behavioural signals. A distilled transformerencoder produces first-order rich-bidirectional contextual representations of each review so as tocapture subtle semantic irregularities. These embeddings are then analyzed by a bidirectional LSTMin order to understand long-range sequential dependencies and an attention mechanism is used tohighlight the most deception-worthy words and phrases. In parallel, teaching behaviour features at ascale that is representative of the structure of reviews and statistics of the scripts are fused with deeptextual representation, allowing the model to make use of stylistic anomalies alongside the extractionof contextual meaning, making far more reliable authenticity prediction possible. The experiments onthe Amazon Enhanced Fake Reviews Dataset show the proposed model has an accuracy of 99.85%,99.87% precision, 99.84% recall, and a 99.85% F1-score, which is evenly better than strong baselineproposals. These results validate that the joint utilization of transformer-level semantics, sequentialdynamics, attention-guided saliency, and behavioral evidence is a highly reliable solution to detectfraudulent opinion spam, hence the introduced framework is ready to be used in practice in real onlinemarketplaces for preserving the authenticity of reviews and protecting consumers from misleadingcontents.

Abstract DOI

Dataset

1 publication

2025·Mendeley Data

Student Performance Data Bangladesh

This dataset has been collected to support research on predicting the academic performance of Secondary School Certificate (SSC) and Higher Secondary Certificate (HSC) students in Bangladesh. It comprises responses from many students across various institutions in the country.

Abstract DOI

Articles Under Review

3 publications

· Environmental Science and Pollution Research

AttenFusion-Net: An Attention-Enhanced Deep Learning Framework for Automated Municipal Solid Waste Segregation

The growing problem of municipal solid waste (MSW) generation presents a challenge in the efficient and scalable sorting of urban solid waste. Although deep learning models hold potential for automated waste segregation, their effectiveness is limited by the large Intra-class variance, Inter-class Similarity and Class Imbalance in some real world data. To overcome these limitations, we present a deep learning model with attention mechanisms, ResNet-CBAM, combining a ResNet-50 backbone with a Convolutional Block Attention Module (CBAM) to enhance feature representation through channel and spatial awareness. The approach is tested on the open Waste Segregation Image Dataset, which includes eight classes of biodegradable and non-biodegradable waste. To overcome the class imbalance problem, a combination of data augmentation and undersampling is applied. The findings reveal that our approach can attain an accuracy of 93.09% and an F1 score of 0.9122, surpassing some baseline approaches. The use of attention mechanisms improves the model's discriminative ability, resulting in better classification performance, especially in visually diverse environments. While minor errors still remain among visually similar classes, these results show a stable performance of the proposed model for all classes. In conclusion, the ResNet-CBAM model presents a reliable and transparent solution for waste classification, which can be further integrated into smart waste management and smart cities.

Abstract DOI

·Physical and Engineering Sciences in Medicine

An Explainable Hybrid Deep Learning Approach for Histopathological Cancer Diagnosis

Proper histopathological cancer classification is vital for early diagnosis; however, this is not easy due to the variations in tissue types. In this research, the hypothesis is that a hybrid deep learning model can be developed by combining DenseNet121 and Swin Transformer Tiny using a Cross-Attention Fusion Mechanism, which will combine local texture features with extensive global contextual representations. The model was trained and tested on the Lung and Colon Cancer Histopathological Image Dataset, which consisted of five classes of tissues. Generalization was improved by applying standard preprocessing techniques, including resizing, normalization, and MixUp augmentation. According to experimental findings, the hybrid model achieves a model accuracy of 99.84%, a Cohen's Kappa of 0.993, which is better than that of conventional CNNs (ResNet18, DenseNet121, ConvNeXt-Tiny) and transformer models (Swin-Small). An additional indication that the model emphasizes diagnostically important areas is the use of score-CAM visualizations. These results suggest that the proposed hybrid cross-attention architecture provides a precise and comprehensible framework for cancer diagnosis using AI on histopathological images.

Abstract DOI

·Soft Computing- Springer Nature

Textile Wastewater COD Forecasting Using a Shock-Aware Explainable Gated Hybrid Framework

Textile dyeing industry is a significant contributor of complicated and extremely polluting wastewater. This wastewater has intermittent loads of chemical oxygen demand (COD), stains and other pollutants which puts dangerous effects on the sustainability of the environment and human beings in general. The traditional operation of wastewater treatment plants is reactive and rule-based to a large extent. These methods are ineffective in dealing with the non-linear dynamic character of the effluent of the textile business, resulting in low efficacy and recurring regulatory breach. To overcome these shortcomings, this paper will suggest a new hybrid architecture SAGE-GBTCN (Shock-Aware Gated Ensemble with Gradient Boosting and Temporal Correction Network) to be used in the effective prediction of wastewater pollution. This model combines a gradient boosting ensemble to produce baseline predictions and a parallel temporal network with a residual correction. A shock-sensitive gating system is used to dynamically modify the correction process to consider any sudden, non-stationary changes in the nature of the effluents. This design makes the model very useful in capturing the long-term trends as well as abrupt disruptions within textile wastewater. The suggested SAGE-GBTCN model was tested with the help of data on a full-scale wastewater treatment facility. The findings are shown to be more accurate in prediction and better resistant to abnormal operating condition. The model also demonstrates high possibilities to facilitate active and energy saving management of textile wastewater treatment processes, which will result in an R2 predictive value of 0.942 and a RMSE of 30.30 of COD. Although validated on full-scale industrial WWTP data, the proposed framework targets operational characteristics typical of textile effluent treatment plants, including batch-wise COD loading, abrupt shock events, and chemically driven variability.

Abstract DOI