Informatics Institute of Technology

This paper proposes a hybrid Transformer-LSTM autoencoder for student final grade prediction, addressing the limitations of existing models that capture either short-term or long-term learning patterns but not both. The model leverages the Transformer’s ability to extract long-range dependencies and LSTM’s strength in modeling sequential short-term behavior, enabling a comprehensive representation of student learning trajectories. Using the Open University Learning Analytics dataset, the approach integrates anomaly detection via reconstruction error to proactively flag at-risk students. Experimental results demonstrate that the hybrid model outperforms LSTM-only and Transformer-only baselines, achieving an F1-score of 0.85 and an AUC-ROC of 0.92, thereby offering a more reliable framework for early intervention. This contribution highlights both the methodological novelty and the practical utility of the hybrid architecture in educational data mining, making the approach a promising tool for supporting timely and personalized student interventions.

Murraya paniculata L., a widely grown ornamental species, exhibits notable phenotypic variation across diverse agro-ecological regions. This study investigated morphological diversity in leaves, flowers, and fruits across Sri Lanka and established a field gene bank for long-term conservation. Surveys were conducted in 13 districts representing varied climatic conditions, and accessions were collected using multi-stage sampling. Twenty-four qualitative and quantitative traits were assessed over two years through field and laboratory characterization. Geographical distribution was mapped using GIS, and ten-year climatic data were incorporated into correlation analysis and Principal Component Analysis (PCA). Morphological traits showed significant associations with precipitation, temperature, and latitude. PCA identified key traits such as petal length, petal width, leaf length and width, petiole length, and internode length that explained 84.75% of the total variation. Accessions from cooler, high-rainfall regions tended to have larger vegetative and floral structures, whereas more compact morphologies dominated in warmer, drier areas. All accessions were successfully established in a field gene bank. Overall, the study demonstrates that climatic, and geographic factors strongly shape morphological variation in M. paniculata, underscoring its high phenotypic plasticity and providing valuable insights for breeding, horticulture, and conservation.

Adverse Drug-Drug Interactions (DDIs) pose a critical challenge under polypharmacy. Traditional methods (rules, classical ML) struggle with scale and novelty in freetext literature. The proposed framework PharmaMap-LLM a DDI extraction framework that instruction-tunes a biomedical LLM via QLoRA for memory-efficient do main ad aptation on a single RTX 3090. The system uses an agentic flow: a Predictor (MedLLaMA-3-8B) with two Validators and a clinician-facing Risk agent so that predictions, confidence, a nd <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{r}$</tex> ationales are cross-checked for explainability. Training leverages the DDIExtraction 2013 corpus and an instruction dataset derived from DrugBank pairs (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\boldsymbol{\sim} \mathbf{2 0, 0 0 0}$</tex>), with entity normalization (SciBERT tokenizer). The work reports hardware and training arguments (learning rate, LoRA ranks, batch/sequence sizes) to support reproducibility. On DDIExtraction 2013, PharmaMap-LLM attains F1 (0.87), Precision (0.89), and Recall (0.85), surpassing prior text-mining baselines. Human evaluation by a clinical pharmacologist (200 cases) rated 83 % of responses as accurate or better (mean 4.2/5); an LLM-as-judge audit (GPT-4, 2,000 outputs) found 92.3 % correctness. Against GPT-4o on 500 shared prompts, validity was similar; the difference <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$w$</tex> as not significant b y M cNemar's <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\chi{ }^{2}$</tex> (continuity-corrected), <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{p} \approx \mathbf{0. 4 0}$</tex>. Brief error analysis highlights under-detection of pharmacodynamic patterns and missing dose/time qualifiers, which inform our mitigation strategy (PD-focused exemplars, retrieval of threshold cues, calibrated confidence). Overall, the proposed system achieves an explainable, reproducible, and deployment-friendly DDI tool for clinical decision support.

This paper introduces a hybrid spatial fusion model for near real-time player recognition in T20i cricket. The system combines ResNet50 for deep feature extraction fused with Support Vector Machine (SVM), K-Nearest Neighbors (KNN) base classifiers, and a Logistic Regression meta-classifier in a stacking ensemble. A novel, domain-specific dataset was developed to capture challenges such as occlusions, variable lighting, and distant camera views followed by an expert evaluation and validation of the dataset. The proposed model achieved 98.14% accuracy, 98% precision, and 98% recall, significantly outperforming standalone deep learning models. These results highlight the advantages of combining deep and traditional machine learning methods. The approach is both robust and resource-efficient, offering a practical solution for sports analytics and laying the foundation for future work on temporal data and transfer learning.

Jazz reharmonization is a complex musical technique that improves harmonic frameworks by inventive chord progressions, augmenting the depth and emotional resonance of pieces. This paper examines the use of artificial intelligence, namely transformer networks, in the automation of jazz reharmonization, ensuring melodic coherence and artistic fidelity. This review compiles current research to highlight essential approaches, including early rule-based systems, neural networks, and advanced transformer models while emphasizing their benefits and limits. Particular emphasis is placed on the difficulties associated with dataset preparation, assessment measures, and the adaptation of AI systems to the intricacies of jazz harmony. The evaluation, via a critical examination of previous studies, identifies areas for enhancement, including real-time applications, multi-modal data integration, and cross-genre adaptation. The insights provided seek to direct future research and position transformers as a revolutionary instrument in computational musicology. This document is an extensive reference for scholars and professionals investigating the convergence of artificial intelligence and jazz music.

In the context of increasing digital accessibility and the need for inclusive communication technologies, the demand for automated sign language recognition has grown significantly. However, Sinhala Sign Language (SSL), used by the Sri Lankan deaf community, remains underexplored due to the lack of labelled datasets and the challenges of developing recognition models for low-resource languages. This study addresses this gap by introducing an SSL dataset, specifically curated to support real-time recognition of 384 commonly used words in daily and work-related contexts. The methodology for dataset creation involved expert consultations, a structured data collection process, and multi-step labelling. These steps ensured the accurate labeling of video data, which can be used for machine learning tasks like gesture recognition, sentence formation, and speech translation. The proposed dataset lays the foundation for SSL recognition models that are fast, lightweight, and optimized for low-resource environments. Additionally, we explore the performance of initial recognition models using this dataset, demonstrating its applicability in real-world scenarios. The dataset and research outcomes provide valuable resources for the development of inclusive AI-driven communication tools for the Sinhala-speaking deaf community. This contribution addresses the pressing need for better technological support for underrepresented languages in the field of sign language recognition and processing.

In response to a pervasive lack of legal awareness among Sri Lanka, this research introduces a law constitution chatbot utilizing Natural Language Processing (NLP). The chatbot is designed to offer precise legal guidance by extracting pertinent laws from the Sri Lankan constitution based on user input. Building on existing studies at the intersection of NLP and the legal domain, this chatbot addresses the lack of relevant legal context for Sri Lanka. Reinforcement learning is integrated to refine the chatbot’s behavior through user feedback, ensuring optimal selection of the most relevant laws. An exploration of applications in countries like Canada and India, where NLP and machine learning have successfully addressed legal information gaps, focuses on various legal domains in Sri Lanka with the aim of empowering the public. The research methodology involves leveraging the Sri Lankan constitution as the dataset, extracting keywords from legal documents, and employing various NLP models, and limitations of the application. A comprehensive literature review reveals the potential energy of NLP techniques in legal chatbots. The global analysis of legal chatbots significantly contributes to the theoretical foundation of this project. In summary, this project seeks to advance legal chatbot technology in Sri Lanka, utilizing NLP and reinforcement learning techniques to empower citizens with legal knowledge and enhance overall legal awareness, thereby addressing the critical issue of legal unawareness and fostering a more informed society.

Osteoarthritis (OA) is a degenerative joint disease marked by the gradual breakdown of cartilage and changes in the underlying bone. The condition is most prevalent in older adults and is often exacerbated by joint injuries and mechanical stress. Knee osteoarthritis (KOA) is one of the common disability adjusted life years. There is no cure for the disease, but a combination of lifestyle changes, medications, physical therapies, and sometimes knee transplants can help manage the condition. Despite being time-consuming and highly prone to user variation, manual segmentation, diagnosis, annotation of knee joints, and personalized treatment recommendations remain common procedures for diagnosing osteoarthritis in clinical practice. This research addresses these limitations and reduces diagnostic errors and time and enhance the efficiency and accuracy of knee osteoarthritis diagnosis. The proposed KOA detection model utilizes a deep learning-based approach with Convolutional Neural Networks (CNNs) to identified knee joints in x-rays, Binary classification to determine the presence of disease, a pre trained network(VGG16) and custom CNN used to classify the Kellgren-Lawrence (KL) grades, and random forest model recommending and validating treatment options based on severity. The proposed model has shown promising results by obtaining 98% overall accuracy.

In the contemporary digital era, mental health awareness has improved, yet challenges remain in therapy adherence and social stigma associated with seeking help for mental disorders such as depression, anxiety, and PTSD. Premature discontinuation of psychological therapies is prevalent, particularly among younger demographics like millennials and Generation Z, who often face significant societal judgment. This research paper presents a computer science project that explores the potential of gamification and micro-AI to enhance retention and completion rates in cognitive behavioral therapy (CBT). The project develops an interactive application that employs gamification strategies to increase engagement and motivation in therapy sessions. Additionally, it integrates micro-AI for real-time mood analysis, offering personalized feedback and support, thereby deepening users' connection to the therapeutic process. With functional testing pass rate of 94%, the primary goal of this innovation is to improve mental health outcomes by mitigating the issues of adherence and premature therapy dropout, providing a scalable and technologically integrated approach to mental health interventions. This study evaluates the effectiveness of these interventions through quantitative metrics and user feedback, aiming to contribute valuable insights into the intersection of technology and mental health.