Volume 17, Issue 4, September 2026

A Multi-Class Deep Learning Approach for Early Detection of Depression Using Social Media Data

Muhammad Sajid, Muhammad Junaid Arshad, Muhammad Owais

Department of Computer Science, Tokyo Metropolitan University, Japan
Department of Computer Science, University of Engineering and Technology (UET), Lahore, Pakistan

Abstract- Depression is a serious mental health issue and early detection of depressive symptoms can help raise awareness and take early action. The social media platforms are full of tremendous quantity of user-generated text that can be used to communicate uncomfortable emotions such as stress, sadness, emotion and more. Many previous studies, however, have focused only on binary classification of depression, and have paid little attention to the severity level of depression. In this paper, a multiclass deep learning approach is proposed for early detection of depression based on the data obtained from social media. The task is presented as a four-class classification task where the classes are non-depressed, mild depression, moderate depression and severe depression. The performance of the models was evaluated for both single-source and multi-source configuration with two datasets measuring depression severity, DEPTWEET and HelaDepDet. Fine-tuning and comparing transformer-based models like BERT, MentalBERT and RoBERTa. To alleviate class imbalance, a balanced leaning strategy was constructed through random oversampling, WordNet-based synonym augmentation and weighted cross entropy loss. The macro AUC-ROC, F1- score, recall, precision, and accuracy were used to assess the performance of the models. From the experimental results, it can be seen that the model based on RoBERTa had good performance in all datasets, especially the balanced learning model, which resulted in the greatest improvement. Balanced RoBERTa that was trained on the balanced DEPTWEET dataset had the best performance with macro AUC of 0.9965, F1-score of 0.97, recall of 0.97, precision of 0.96, and accuracy of 0.97. This study illustrates the effectiveness of using transformer-based deep learning and imbalance-aware training in the multi-class severity classification of depression in social media texts, particularly for underrepresented severity classes.

Keywords- Depression Detection, Social Media Data, Multiclass Classification, Depression Severity, BERT, MentalBERT, RoBERTa, Class Imbalance and Deep Learning

Download full paper PDF format (Page: 1-12)

AI-Driven Multivariable Time-Series Analysis for Optimal Operation of a Monoethylene Glycol and Propane-Based Mechanical Refrigeration Gas Processing System Using CNN, LSTM, DVGAN, Transformer Learning, and Reinforcement Learning

Ndifreke Meyen Uwah, Rowland Obot, Chike Aneke

Computer Engineering Department, University of Uyo, Nigeria
Savannah Energy, Nigeria

Abstract- Natural gas processing plants rely heavily on mechanical refrigeration systems and glycol dehydration units for hydrocarbon dew point control, water dew point reduction, condensate recovery, and flow assurance. This study presents an artificial intelligence-driven multivariable time-series analysis framework for operational optimization and anomaly detection in a gas processing mechanical refrigeration system utilizing propane refrigerant and monoethylene glycol (MEG). Historical industrial process data obtained from a high-pressure gas processing facility were analyzed using an integrated deep learning architecture comprising Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, Deep Variational Generative Adversarial Networks (DVGAN), Transformer Learning, and Reinforcement Learning (RL). The developed framework was applied to analyze process variables associated with HP separator, LP separator, gas/gas exchanger, chiller section, low temperature separator (LTS), and glycol regeneration system. The proposed AI framework successfully identified unstable operating conditions, abnormal process signatures, pressure fluctuations, flow irregularities, and thermal inefficiencies within the refrigeration train. Optimal operating windows for critical process equipment were established using statistically stable operational envelopes derived from normal process behavior. Results showed that maintaining HP pressure within 70–75 barg, LP pressure within 28–35 barg, chiller outlet temperature between 10–15 °C, gas/gas exchanger outlet temperature around 33–36 °C, and LTS temperature between 5–10 °C significantly improved condensate recovery and refrigeration efficiency. In addition, efficient glycol regeneration was achieved at lean glycol temperatures of 22–28 °C, glycol circulation pressures of 45–50 barg, and reboiler temperatures near 200 °C....

Keywords- Mechanical Refrigeration System, Monoethylene Glycol, Propane Refrigeration, CNN, LSTM, Transformer Learning, Reinforcement Learning, DVGAN, Anomaly Detection, Gas Processing Optimization, SCADA and Low Temperature Separator

Download full paper PDF format (Page: 13-20)

Deep Learning-Based Shellcode Detection Using ANN, CNN, and LSTM on UNSW-NB15 Dataset

Ayesha Tahir, Dania Saqib

Institute of Data Science, University of Engineering and Technology, Lahore
Department of Computer Science, UET, Lahore

Abstract- The rapid growth of internet technologies and interconnected digital systems has significantly increased cybersecurity threats and malicious network activities. Among various cyberattacks, Shellcode attacks are considered highly dangerous because they exploit software vulnerabilities and execute malicious instructions within target systems. Traditional Intrusion Detection Systems (IDSs) often struggle to detect modern Shellcode attacks effectively due to highly imbalanced datasets and evolving attack patterns. Therefore, this study proposes a deep learning-based Shellcode detection framework using Artificial Neural Network (ANN), Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM) models on the UNSW-NB15 dataset. The proposed methodology performs binary classification, labeling Shellcode traffic as malicious and all remaining traffic as normal. Multiple data preprocessing techniques, including label encoding, feature scaling using StandardScaler, and the Synthetic Minority Oversampling Technique (SMOTE), are applied to address severe class imbalance. Furthermore, threshold optimization and early stopping techniques are implemented to reduce false positives and improve model effectiveness. Experimental results show that deep learning models combined with oversampling significantly improve Shellcode detection capability. Among all models, ANN achieved the best overall performance with 95.51% accuracy, 89.68% recall, and 0.16 F1-score, while CNN and LSTM also produced competitive results, with strong recall. The results show that deep learning, combined with oversampling methods, can effectively improve intrusion detection performance on highly imbalanced cybersecurity datasets.

Keywords- Cybersecurity, Shellcode Detection, Deep Learning, Artificial Neural Network, Convolutional Neural Network, Long Short-Term Memory, Intrusion Detection System, SMOTE, UNSW-NB15 and Network Security

Download full paper PDF format (Page: 21-26)

Integrating Neural Networks into Resource-Constrained IoT Devices: Challenges and Techniques

Muhammad Saad Azam, M. Junaid Arshad

Department of Computer Science University of Engineering and Technology Lahore, Pakistan

Abstract- With the rapid increase of number of IoT devices, there has been a growing demand for intelligent lightweight devices. However, the constraints of energy, memory and computation of microcontroller-class devices make the deployment of efficient neural networks fundamentally challenging. This survey provides a structured analysis of techniques that can bridge this gap, organized across three dimensions: lightweight neural net-work architectures, model compression techniques and inference frameworks. We examine the tradeoffs of each architecture and if those architectures are IoT friendly. Our analysis shows that architectures like MCUNet and MCUFormer, that target both the design as well as the inference are most well-suited for IoT devices. When it comes to compression techniques, all the tech-niques target different paradigms of compression: Quantization universally benefits IoT devices, Knowledge Distillation (KD) is quite prominent for transformer compression, and NAS has its place for the design of the model itself, where it is more mature and prominent for CNN architectures. The frameworks are analyzed on the basis of memory, portability and execution model and how well-suited are they for MCU-class devices. Lastly, four major challenges are addressed: the memory-accuracy tradeoff, on-device continual learning, transformer deployment gap and hardware limitations for IoT devices.

Keywords- Tinyml, Microcontroller Deployment, Neural Network Compression, Lightweight Architectures, Quantization, Pruning, Knowledge Distillation, Neural Architecture Search, Vision Transformers, Resource-Constrained IoT, Inference Frameworks and Edge AI

Download full paper PDF format (Page: 27-36)

Understanding Common Risks in Software Requirements Engineering: Perspectives from Academia and Industry

Hamed Jaferi, Mohamed Masoor

Faculty of Computer Science, International University of Science and Technology, Sudan

Abstract- Requirements engineering plays a critical role in the success of software development projects, as incomplete, ambiguous, or misunderstood requirements can lead to significant misalignment between customer expectations and the developed software. Such discrepancies may direct development efforts toward inappropriate solutions, resulting in increased costs, delays, rework, or even project failure. Therefore, it is essential to identify not only the risks associated with the requirements engineering phase but also the factors that contribute to their occurrence. This study adopts a secondary research methodology and conducts a comprehensive literature survey to explore common risks in software requirements engineering. The study further examines these risks from the perspectives of both academic research and the commercial software industry, providing a broader understanding of the challenges encountered across different software development contexts. The resulting identification and analysis of risk contributors can serve as a valuable reference for researchers and software practitioners, particularly inexperienced requirements engineers, to improve risk awareness and support more effective requirements elicitation, analysis, and management.

Keywords- Software Requirement Engineering, Academia and Industry, Common Risks and Analysis

Download full paper PDF format (Page: 37-42)

Agentic AI for Real-Time Air Quality Prediction and Autonomous Environmental Response

Muhammad Saad Khalid Khan Suri

Department of Computer Science University of Engineering and Technology Lahore, Pakistan

Abstract- Rapid urbanization and industrial growth have exacerbated air pollution, presenting a critical threat to global public health. While traditional environmental monitoring systems provide passive data visualization, they fail to deliver proactive, autonomous interventions. This paper presents a novel, production-ready platform that integrates real-time Air Quality Index (AQI) prediction with a multi-agent Artificial Intelligence (AI) framework. Utilizing an Extreme Gradient Boosting (XGBoost) model trained on high-resolution meteorological and pollutant data, the system achieves an R2 score of 0.9929 and a Mean Absolute Error (MAE) of 0.85 in AQI forecasting. Furthermore, a Large Language Model (LLM) powered multi-agent architecture evaluates these predictions to autonomously dispatch contextual alerts and mitigation strategies. The system is containerized via Docker and features a high-performance FastAPI backend coupled with a real-time Streamlit dashboard. Experimental results demonstrate the efficacy of combining predictive machine learning with agentic decision-making to transition smart cities from passive monitoring to active environmental management.

Keywords- Air Quality Index, Agentic AI, Machine Learning, XGBoost, Smart Cities, Internet of Things (IoT) and Multi-Agent Systems

Download full paper PDF format (Page: 43-48)

Security and Privacy in the Internet of Things: Threats, Countermeasures, and Future Research Directions

Faisal Arshad, M. Junaid Arshad

University of Engineering and Technology (UET), Lahore

Abstract- The Internet of Things (IoT) has become a critical foundation for smart healthcare, industrial automation, intelligent transportation, smart cities, and connected homes by enabling continuous communication among sensing devices, networks, cloud platforms, and user applications. However, the rapid expansion of IoT has also created serious security and privacy concerns because many devices operate with limited computation, memory, energy, and update capability. These constraints, combined with heterogeneous protocols, weak authentication, insecure firmware, and large-scale connectivity, expose IoT ecosystems to threats such as denial-of-service attacks, malware, spoofing, routing manipulation, unauthorized access, data leakage, and privacy inference. In this paper, the existing research works regarding IoT security and privacy from 2020 to 2026 will be reviewed. The aspects of this topic that will be covered by these studies include but are not limited to the security threats, protection technologies, challenges, and the future trends. This work will discuss security requirements and attacks, cryptographic techniques for protection, authentication and access control, intrusion detection, blockchain solutions, and privacy issues. Moreover, it will provide a comparative analysis of the existing solutions based on their benefits and limitations with respect to scalability, efficiency, inter-operability, and practicality. Future research will focus on lightweight security, AI-based threat detection, zero-trust IoT, edge protection, and privacy preservation.

Keywords- Internet of Things, IoT Security, IoT Privacy, Cybersecurity, Threats, Countermeasures, Access Control, Intrusion Detection, Lightweight Cryptography and Blockchain Security

Download full paper PDF format (Page: 49-56)