Computing and Computational Intelligence (CCI) 1
Track C8F1 CCI 8.1: Computing & Computational Intelligence (CCI) 8.1
Room: F1. Sipadan I (Level 4)
Chair: Min Keng Tan (Universiti Malaysia Sabah & Modelling, Simulation & Computing Laboratory, Malaysia)
2:30 RainGAN-Kathmandu: a Generative Adversarial Framework for Synthetic Rainfall Augmentation in Urban Road Scene Datasets
Nitesh Kumar Shah (Indian Institute of Information Technology, Allahabad, India); Gadde Jahnavi (IIIT Allahabad, India); Navjot Singh (Indian Institute of Information Technology Allahabad, India); Chandra Prakash Maurya (IIIT Allahabad, India); Satish Singh (IIT Allahabad, India)
Urban roads face major challenges during rainfall, impacting visibility and road texture. Rain degrades the performance of vision-based systems in traffic safety and monitoring. Real-world rainy road datasets are scarce and often lack diversity. This limits the training of models for image translation under rainy conditions. Effective rainy image synthesis is essential for advancing robust autonomous driving systems. This paper introduces a novel methodology to bridge this gap by generating synthetic datasets simulating adverse weather conditions, particularly rainfall, using Generative Adversarial Networks (GANs). We introduce a newly curated dataset of Kathmandu road scenes to provide diverse, real-world clear-weather imagery. Leveraging these datasets, we employ advanced deep learning techniques to synthesize high-quality rainy road scenes. Generated outputs are evaluated using Fréchet Inception Distance (FID) and Learned Perceptual Image Patch Similarity (LPIPS) metrics to ensure realism and perceptual fidelity. The paper outlines the data creation pipeline, model implementation, evaluation strategy, and broader implications, offering a robust framework for weather-affected scene synthesis and road safety research.
2:45 GTCGAN: an Unsupervised Approach for Single Image Deraining
Debesh Kumar Shandilya (IIIT Allahabad, India); Navjot Singh (Indian Institute of Information Technology Allahabad, India)
Single image deraining is a challenging task even today due to the unavailability of real paired datasets. Though many supervised models have been developed, they are trained on paired synthetic datasets, which cannot map real rainy conditions beyond a certain extent, which limits the capability of the supervised models to perform equally well on the real test datasets. The proposed model, Guided TransCycleGAN (GTCGAN), is based on CycleGAN, which uses an unpaired real dataset for training. Further, our model uses both CNN and attention block to exploit local and global relationships in the image. Apart from this, our model uses two extra discriminators to guide generators to generate clean images for the corresponding rainy images. GTCGAN's performance is enhanced by a multi-objective loss function combining edge loss for visual quality and Structural Similarity Index Measure loss for structural integrity. Our model outperformed the state-of-the-art in terms of average PSNR and average SSIM on three publicly available datasets.
3:00 Functional Connectivity-Driven Diagnosis of ADHD and ASD via Transfer Learning on Rs-fMRI Data
Khushi Jain, Akanksha Upadhyay and Vaishali Chawla (University of Delhi, India); Navjot Singh (Indian Institute of Information Technology Allahabad, India); Bharti Rana (University of Delhi, India)
Neurodevelopmental conditions such as Autism Spectrum Disorder (ASD) and Attention Deficit Hyperactivity Disorder (ADHD) alter functional connectivity (FC) among the brain regions. ADHD and ASD affect 5-6% and 1-2% of the population, respectively and hence need immediate attention to develop automated differential diagnosis. This work proposes an automated diagnostic system for ADHD and ASD by investigating FC among brain regions. We used publicly available resting-state functional Magnetic Resonance Imaging (rs-fMRI) data for ADHD, ASD and age-matched healthy (HC) subjects. The rs-fMRI data helps to analyse the temporal activity in the brain, and computing FC using time-series data helps to capture the temporal correlation among regions. We used five diverse FC measures and performed three classification tasks: ASD vs. ADHD, ASD vs. HC and ADHD vs. HC. Further, we fine-tuned the pre-trained VGG19 model to develop a decision model. The best accuracy of 99.09% for ASD vs. ADHD, 78.18% for ADHD vs. HC and 65.46% for ASD vs. HC demonstrates the efficacy of the proposed method. The disrupted regions are also identified using SHapley Additive exPlanations (SHAP), an explainable artificial intelligence method, that are in accordance with the literature. The work demonstrates the capability of deep learning in analysing neuroimaging data for clinical and research applications.
3:15 Multi-Scale Temporal Attention Convolutional Neural Network for Sleep Stage Classification Using Single-Channel EEG
Aikendrajit Ningthoujam and Shaik Rafi Ahamed (Indian Institute of Technology Guwahati, India)
Accurate and efficient classification of sleep stages is essential for advancing widespread sleep monitoring solutions. This study presents a novel multi-scale temporal attention Convolutional Neural Network (CNN) tailored for single-channel EEG (Fpz-Cz) sleep stage analysis. Our architecture innovatively combines temporal attention mechanisms across multiple scales within a convolutional neural network framework. By emphasizing temporal relationships in the EEG signal across varied receptive fields, the model effectively captures discriminative features without the computational overhead of channel-wise operations common in multi-channel systems. This multi-scale strategy enables the network to autonomously identify and prioritize significant patterns, ranging from short-term transients to longer rhythmic structures. The proposed design strikes an optimal balance between performance and computational efficiency compared to existing deep learning approaches. Evaluated on the Sleep-EDF-20 dataset, our model achieves a competitive accuracy of 76.46% and a Cohen's kappa κ of 0.69, with approximately 141.5k parameters. This research highlights the potential of multi-scale temporal attention in CNNs for accurate and computationally efficient single-channel EEG sleep stage classification, paving the way for future advancements and practical applications.
3:30 Life Data Analysis (LDA) Using Weibull Distribution Method for Determination of Reliability for HPSV and LED Street Light
Nor Diana Ruszaini (TNB Research Sdn Bhd, Malaysia); Khairuddin Abdullah and Asnawi Mohd Busrah (TNB Research, Malaysia)
This study applies Life Data Analysis (LDA), particularly Weibull distribution modeling, to evaluate and compare the reliability, failure patterns, and operational lifespans of High-Pressure Sodium Vapor (HPSV) and Light Emitting Diode (LED) streetlights using real field data. Although LED streetlight technology gained international traction around 2012, its long-term performance was not well documented at the time. In Malaysia, Tenaga Nasional Berhad (TNB) began widespread deployment of LED streetlights in 2016. This study utilizes actual operational and failure data from thousands of units to conduct a statistically rigorous analysis. LDA is a methodical reliability engineering approach that uses time-to-failure data to predict product life and identify failure trends. The Weibull distribution, a versatile tool in LDA, was employed to model failure behaviors of both HPSV and LED lights. In addition to failure modeling, the Mean Time to Failure (MTTF) was calculated for both lighting technologies, serving as a fundamental indicator of reliability. The results demonstrated that LED streetlights possess a significantly higher MTTF compared to HPSV units, confirming their longer lifespan and superior durability in outdoor conditions. These findings reinforce the reliability advantages of LED systems and align with their design intent of long-term, low-maintenance operation.
3:45 S-LIME: an Explainable SVM-Based Framework for Human Activity Recognition Using LIME
Shaharier Kabir (American International University-Bangladesh, Bangladesh); Muhammad Masud Karim (Bangladesh University of Engineering and Technology, Bangladesh)
Human Activity Recognition (HAR) has emerged as a pivotal application in healthcare, assistive technology, and smart environments, leveraging wearable sensor data to classify human actions. While Support Vector Machines (SVMs) are widely used for HAR due to their robustness in handling high-dimensional data, their lack of interpretability restricts their deployment in real-world applications requiring transparency and trust. To bridge this gap, we propose S-LIME (SVM-LIME), an explainable AI (XAI) framework that integrates Local Interpretable Model-Agnostic Explanations (LIME) with SVMs to enhance model interpretability in HAR systems. S-LIME enables domain experts, medical practitioners, and end-users to understand the feature contributions that drive SVM-based activity classifications, making AI-driven HAR more transparent and accountable. We evaluate S-LIME on benchmark HAR datasets, demonstrating its ability to generate human-interpretable insights while maintaining high classification accuracy. The framework provides instance-level feature explanations, highlighting the key sensor signals influencing activity classification. Additionally, we introduce LIME-based feature ranking, which identifies and optimises critical sensor signals, further improving model performance. Our findings show that S-LIME achieves state-of-the-art accuracy while providing an interpretable decision-making process, paving the way for reliable, transparent, and ethically sound HAR models.
4:00 Predicting Fall Risk in Parkinson's Patients Through Sequential Gait Analysis and Machine Learning
Gaurav Sharma, G (Bennett University, India); Nehvanshika Nehvanshika, Niharika Anand and Gaurisha Singh (IIIT Lucknow, India)
Falls pose a critical threat for patients with Parkinson's disease, as difficulties in walking and balance greatly increase the likelihood of falling. This study aims to improve fall risk prediction accuracy and computational efficiency using wearable sensor data and advanced learning models. We analyze real-world gait data collected via inertial measurement units and implement both classical machine learning (ML) techniques and deep learning frameworks aimed at recognizing those susceptible to falls. Feature extraction strategies focusing on gait speed, stride variability, and symmetry measures were applied to enhance model performance and reduce dimensionality. Among classical approaches, a Random Forest classifier optimized with Recursive Feature Elimination (RFE) and XGBoost achieved a balanced accuracy of 77%, with sensitivity and specificity of 70% and 84%, respectively. A long-short-term memory (LSTM) model that uses temporal stride sequences significantly outperformed traditional models, achieving balanced precision of 95.97%, sensitivity of 93.29% and specificity of 98.65%. Comparative analysis with baseline literature demonstrated substantial performance gains, highlighting the ability of time-sensitive deep learning methods to predict the risk of falls in PD. The findings underscore the feasibility of deploying robust, real-time predictive systems in clinical and home-monitoring environments, with future work directed towards multimodal integration and scalable deployment.
Computing & Computational Intelligence (CCI) 2
Track C8F6 CCI 8.2: Computing & Computational Intelligence (CCI) 8.2
Room: F6. 505 Sepilok (Level 5)
Chair: Pei Yee Chin (Universiti Malaysia Sabah, Malaysia)
2:30 Performance Analysis of Multimodal Fusion Techniques for Predicting Student Engagement States
Deepika Suranjini Silva and Nadeeshani Wickramage (Sri Lanka Institute of Information Technology, Sri Lanka); Jayakody Arachchilage Don Chaminda Anuradha Jayakody (Curtin University Technology, Sri Lanka & Sri Lanka Institute of Information Technology, Sri Lanka); Pradeep Abeygunawardhana (Sri Lankan Institute of Information Technology (SLIIT), Sri Lanka)
Effective monitoring of student engagement is vital for enhancing learning outcomes, especially in digital and hybrid classroom settings where traditional physical cues such as eye contact and posture are often limited. This study investigates the performance of machine learning models combined with multimodal data fusion techniques to predict student engagement levels using the Emotional Monitoring Dataset. The dataset integrates behavioral (e.g. facial expressions), physiological (e.g. heart rate, skin conductance), and environmental (e.g. ambient noise, lighting) signals, all labeled with discrete engagement states: Highly Engaged, Moderately Engaged, and Disengaged. Three widely used machine learning models, Random Forest (RF), XGBoost, and Multilayer Perceptron (MLP) were evaluated using both early and late fusion strategies. Results demonstrate that the late fusion approach consistently outperformed early fusion across all models, with XGBoost achieving the highest accuracy (R² = 0.99999) and lowest error (MSE = 3.88e−07), indicating near-perfect prediction. Random Forest also yielded strong results, while MLP showed competitive performance under the late fusion configuration. These findings highlight the effectiveness of decision-level integration in handling multimodal data and support the use of ensemble and deep learning approaches for developing intelligent, real-time attention monitoring systems in education. The proposed methodology offers promising applications in personalized learning and emotion-aware educational technologies.
2:45 Real-Time Road Damage Classification and Severity Detection System
Victor Sebastian D. Bondoc, Efren Jr. D. Pastores, Julius Nikolai D. Bernardo, Gary Clyde T. Rabe, Renato T. Panis Jr and Jevon A. Silvano (Technological University of the Philippines, Philippines); Immanuel Jose C Valencia and Ira Valenzuela Estropia (De La Salle University, Philippines); Ryan Reyes (Technological University of the Philippines, Philippines); Lean Karlo S. Tolentino (Technological University of the Philippines, Philippines & Mapua University, Philippines); Jessica S. Velasco (Technological University of the Philippines, Philippines); Mark P Melegrito (Technological University of the Philippines Manila, Philippines)
In the Philippines, road pavement damage poses significant threats to road safety, often leading to accidents, vehicular damage, and increased maintenance costs. Typically, actual road surveying is conducted when inspecting road pavement damage, a traditional method that is time-consuming when generating reports about the surveyed road pavement. Requiring labor, expenses, and subjective evaluations, resulting in inaccuracies and inefficiencies in assessing road conditions, which delays maintenance and repair. The main objective of this study is to develop a real-time road damage classification and severity detection using YOLOv8. Using a moving vehicle, a high-resolution camera is mounted, capturing images of road pavement damages in real-time, processed by the NVIDIA Jetson Nano, YOLOv8 as the deep-learning model classifying road pavement damage classes: potholes, cracks, alligator cracks, pumping, and depression, determining it identified severity levels based on the standards and guidelines set by the Department of Public Works and Highways (DPWH). A VK-162 GPS module is incorporated to geotag each detected road damage by recording its latitude and longitude coordinates in real-time for accurate location mapping and detailed reporting of road conditions. The system was evaluated based on its accuracy of detection and the tagged location. The findings of the study suggest that the system is a viable, low-cost, and scalable assessment tool to conduct preliminary assessments of road conditions, providing significant opportunities to reduce manual labor while improving the speed and efficiency of data collection and asset and infrastructure maintenance planning.
3:00 Predicting Hourly Electricity Demand Using Fuzzy Logic: Integrating Environmental Factors for Accurate Forecasting
Ahmed Intekhab Rohan (Islamic University of Technology, Gazipur, Bangladesh); Md Ismail Hossain (American International University Bangladesh, Bangladesh); Hasanur Zaman Anonto and Abu Shufian (American International University-Bangladesh, Bangladesh); Mohammad Shah Paran (Lamar University, USA); Toriqul Islam (American International University-Bangladesh, Bangladesh); Md Shakhawat Hossain (Lamar University, USA)
This study investigates the distribution of electricity demand is distributed over temperature, humidity, wind speed and seasonal variations. Using data covering the timeframe of five years (2021-2025), the objective behind this research is to understand the interdependent nature of the electric demand and different external conditions of the environment. Preliminary analysis indicates that the demand for electricity varies based on the time of the day, the season and the day of the week. Peak demand is at noon in winter at 27,000 units and declines to 22,000 units during post-midnight. Electric demand positively correlated with temperature and humidity, where a notable increase in demand was detected when temperature increased from 5°C to 8°C and when humidity increased from 60% to 66% The wind speed variation of the demand shows that the demand is less than 2m/s and higher than 5m/s. The impact of the seasonal change is also studied in this paper and the results show that the demand is higher in the winter and summer due to the heating and cooling need respectively while the demand is lower in the spring and autumn. The findings deliver specific intelligence on optimizing how energy management strategies can best be employed, specifically for ensuring that demand drops are anticipated ahead of time around peak seasons and extreme weather.
3:15 MapReduce and K-Means Clustering Method for Long Text Summarization on Large Language Model
Moh. Rosy haqqy Aminy (Sepuluh Nopember Institute of Technology, Indonesia); Diana Purwitasari, Dwi Sunaryono, Ilham Gurat Adillion, Dini Adni Navastara and Bilqis Amaliah (Institut Teknologi Sepuluh Nopember, Indonesia); Hilmil Pradana (Sepuluh November Institute of Technology, Indonesia); Yoga Yustiawan (Pusan National University, Korea (South))
The rapid growth of digital information has resulted in an explosion in the volume of online text documents, ranging from news articles, and financial reports, to scientific papers. This condition triggers the need for an automated summarization system that can present important information in a concise and easy-to-understand manner. The application of Large Language Model (LLM) based summarization models on long documents still faces limited input length challenges. This paper proposes a long text summarization method based on LLM by combining MapReduce and K-means Clustering. Qwen2.5-7B model Instruct is used with Low-Rank Adaptation-based instruction fine-tuning technique. Long documents are divided into chunks, then converted into embedding and clustered using the K-Means algorithm. The clustering results are used to select the most relevant information representation, which is then summarized at the Map stage. The partial summaries are then merged and re-summarized at the Reduce stage to produce one final summary. This approach is capable of processing documents that exceed the token limit of LLM. To evaluate the summarization results, the ROUGE (Recall-Oriented Understudy for Gisting Evaluation) metric is used. An experimental evaluation was conducted on banking sector documents in the BRI dataset and showed that the MapReduce and Clustering method significantly improved performance over the direct truncation approach. Our approach achieves ROUGE-1, ROUGE-2, and ROUGE-L scores of 0.416, 0.118, and 0.219 respectively compared to direct truncation scores of 0.320, 0.090, and 0.168.
3:30 Fuzzy Rule-Based and Multinomial Logistic Regression-Based Models on Selecting Alternative Routes for Maysan Road, Valenzuela
John Lemar M Tirao (Mapua University, Philippines); Jocelyn Buluran (Mapúa University, Philippines); Jordan Velasco (Pamantasan ng Lungsod ng Valenzuela, Philippines)
Maysan Road, a critical national highway in Valenzuela, National Capital Region, Philippines, has long suffered from traffic congestion due to its corridor role in linking MacArthur Highway and General Luis Street, necessitating an effective route management solution. With the rapid increase in vehicular volume brought by urban and economic expansions, this research aimed to propose a traffic management scheme using a fuzzy rule-based model to recommend proper alternative route selection by integrating physical and traffic characteristics of the road network with road user perceptions. A three-phase methodology was employed: (1) data gathering of route conditions and road user perceptions survey, (2) development of fuzzy-rule-based (FRB) system and multinomial logistic regression (MNL) models for route selection, and (3) validation through statistical testing to determine the most efficient route. Parameters such as weather, socio-economic activities, time of the day, day of the week, and level of service were used as inputs in the models. The FRB model tended to emphasize Route 3 under congested traffic and good weather conditions. The MNL was fitted and confirmed the statistical adequacy of all input variables, with varying significance across routes (McFadden's pseudo R-squared value of 0.2623); model validation also showed Route 3 has the highest average predicted probability (21.09%) followed closely by Routes 4 and 1. This research concludes that the hybrid modeling approach that combines statistical and rule-based inference provides an interpretable framework for localized traffic management.
3:45 An Advanced Hybrid Strategy for Detecting Fraud in Mobile Money Services
Kaung Wai Thar and Thinn Thinn Wai (University of Information Technology, Myanmar)
Mobile payment systems are growing in popularity as more people use smartphones, which also attracts fraudsters. As a result, mobile money fraud is increasing, especially in developing countries. However, security concerns in mobile money services have gained attention because weak security has kept many customers away. Fraud is not a new problem, but it still costs the global economy billions of dollars each year. Financial transaction data, including mobile money transactions, are mostly labeled as legitimate, making fraud detection difficult. Researchers have developed various fraud detection methods using various machine learning techniques like LGBM, random forests, SVM, deep learning, neural networks, XGBoost, and logistic regression have been tested to identify fraudulent transactions through data preparation, feature engineering, and model building. However, machine learning models trained on imbalanced data tend to be biased toward legitimate transactions, making them unreliable. This study aims to use machine learning classifiers to detect fraud in mobile money transfers. The data comes from real-time transactions that mimic common fraud schemes. This research explores the development and evaluation of ML models for fraud detection and proposes a solution using a hybrid ML approach with effective hyperparameter tuning, and the Synthetic Minority Over-sampling Technique (SMOTE). The results show that the proposed approach improves accuracy by addressing data imbalance. The findings contribute to the development of better fraud detection systems for mobile money services.
4:00 Medical Datasets for Machine Learning in Brain Tumor Diagnosis and Segmentation: A Review
Priyanka Singh (VIT-AP University, India & Victorian Institute of Technology (VIT), Australia); Jayendra Kumar (Vellore Institute of Technology, India); Samineni Peddakrishna (National Institute of Technology Silchar, India); Banee Bandana Das (SRM University Andhra Pradesh, India); Saswat Kumar Ram (SRM University, Amaravati, Andhra Pradesh, India)
Brain tumor detection through machine learning has gained significant traction due to its potential for early diagnosis, accurate classification, and automated segmentation in clinical settings. The success of such models is closely tied to the quality and availability of annotated datasets. This survey presents a comprehensive review of major publicly available datasets-including BraTS, Figshare Brain MRI, TCGA-GBM/LGG, REMBRANDT, CQ500, and IBSR-highlighting their imaging modalities, annotation protocols, dataset sizes, and clinical relevance. Special emphasis is placed on BraTS for segmentation and Figshare for multi-class classification. Genomics-integrated datasets like TCGA and REMBRANDT support multi-modal learning, while IXI and CQ500 are valuable for pretraining and emergency diagnostic models. The survey identifies key limitations, such as inter-observer variability, class imbalance, and inconsistent annotation formats. It also underscores the need for more diverse, standardized, and richly labeled datasets. By evaluating the strengths and weaknesses of existing resources, this work provides guidance for selecting suitable benchmarks and suggests future directions such as federated learning and synthetic data augmentation to improve clinical robustness.