AI for Social Issues

MUHD IZZUDDIN AIMAN BIN ZOLKIPLI

This research focuses on analyzing Human–Wildlife Conflict (HWC) incidents in Malaysia using LLMs, Knowledge Graph (KG), and Graph-based RAG techniques. The study proposes a prototype system that transforms unstructured HWC news articles, particularly elephant-related conflict reports, into a structured and queryable KG to support semantic analysis, information retrieval, and explainable reasoning.

The system extracts entities, relationships, locations, actions, and temporal information from news articles and organizes them into a graph-based knowledge structure. By integrating the KG with GraphRAG, the system enables natural language querying and generates evidence-grounded responses with improved contextual understanding compared to conventional retrieval-based approaches.

The research also investigates how structured knowledge representation can support information exploration, hotspot identification, and explainable analysis for conservation-related problems. Malaysian elephant HWC incidents are used as the primary case study, with future extensions planned toward broader cross-regional HWC analysis, including African datasets under the AJ-CORE / GSRIT collaboration.

LIEHOUAN Franck Arnaud

Severe Acute Malnutrition (SAM) is one of the leading causes of death among children under five, and rapid diagnosis in low-resource settings such as Côte d’Ivoire is critical for timely medical intervention. However, in peripheral health centres, important clinical information is frequently incomplete. A child’s age may be unknown, and anthropometric measurements such as height or weight are often difficult to obtain accurately under emergency or resource-constrained conditions. This severity-biased missingness problem causes conventional AI diagnostic systems to fail precisely for the most vulnerable patients.

This research proposes a missing-data-robust multimodal AI system designed for reliable SAM diagnosis even in incomplete clinical scenarios where important medical information is unavailable. The proposed framework combines an XGBoost-based tabular diagnostic pipeline with a CNN-based oedema detection module operating on photographs of feet and hands. Following clinical diagnostic guidelines, the detection of bilateral oedema directly triggers a SAM alert to support rapid patient referral and treatment, even when anthropometric measurements are missing.

To ensure robustness under real-world field conditions, the system incorporates a gating mechanism that automatically switches to the tabular diagnostic pipeline when image data are unavailable or of insufficient quality.

Esan Damilola Olawale

This research aims to propose a method for assigning unique identifiers to each target across video frames in order to achieve stable segregation and tracking. The intended application is automated bus passenger counting in developing-region public transit, addressing the social challenge of fraudulent ridership reporting, data collection and analysis in Nigerian urban transport. Existing multi-object tracking methods are known to suffer from frequent identity switches in crowded, unstructured in-vehicle environments, making accurate counting difficult. This work seeks to overcome these identity-consistency challenges and to close the domain gap between AI models trained on developed-world data and the conditions found in developing-region transit systems.

HYDARA Ebrima

My research focuses on advancing deepfake detection for explainable and trustworthy AI in video authenticity assessment. Although deepfake detection models have achieved strong predictive performance, it remains difficult for users to understand why an AI system judges a video as manipulated or authentic. Many existing explainable AI methods mainly visualize important regions within individual frames, but video-level deepfake decisions are formed across time. Therefore, explanations should also clarify how detection evidence appears, persists, concentrates, fragments, and influences the detector’s final output throughout the full video. In this research, I develop DeepForensiX (DFX), a detector-centered temporal explanation framework that derives a temporal importance signal from changes in a detector’s output and analyzes the structure of evidence over time. This approach helps identify which moments influenced the detector, how the evidence was organized, whether it was stable, and how strongly it contributed to the final decision. The goal is to support transparent and reliable use of AI in digital forensics, media trust, content moderation, and other socially important applications.

Takumi Kato

In competitive board game applications such as Shogi and Chess, the “Elo rating” system is widely used to classify player strength and facilitate matchmaking between players of similar skill levels. However, these rankings are based exclusively on win-loss outcomes, leaving the rich information contained within game records largely underutilized. This research posits that board game records encompass various indicators of a player’s true strength and explores a novel method for quantifying proficiency by leveraging this data through deep learning.