Explainable AI for Social Good: Applications in Mental Health, Public Health Risk, and Environmental Traceability

Abstract

The ubiquitous use of machine learning and AI technology in human-centered domains such as social networks, public health, sustainable trade, and environmental forensics indicates a significant need for an adaptive, interpretable, and generalizable approach in predictive modeling. With the increasing availability of user-generated data, environmental samples, and public health records, AI-driven tools have played a significant role in predictive analysis. However, a persistent challenge remains: in domains with significant societal implications, the availability of data is often inconsistent, unstructured, and lacks fine-grained labels. Furthermore, in these application areas, understanding the prediction becomes as important as the prediction itself, as they guide a more informed intervention strategy. Most of the existing work in this domain struggles to meet this requirement by approaching it from either one size fits all modeling approach or by adapting to a very problem-specific, fine-tuned algorithm that fails to learn the inter-task dependency while not particularly focusing on explainability. Therefore, in real-world scenarios, these tools show an increased risk for practical applicability due to their black-box nature, which leads to a lack of intuitive interpretability for domain experts. These methods can often neglect underlying conditions such as spatial dynamics, socioeconomic disparities, and uncertainty. In fields like population health management and stable isotope forensics, such limitations hinder practical deployment and erode trust. Compounding this issue is the widespread adoption of large language models (LLMs), which, despite their power, are prone to hallucinations and toxicity, undermining their reliability in sensitive domains.This thesis employs active learning, multi-task learning, ante-hoc explainability, post-hoc explanations, and probabilistic Gaussian process modeling to tackle several domains of social computing that range from population mental health, epidemiological outbreak, and forensic environmental tracability analysis. The first work introduces AMMNet, a multi-task active learning model for detecting depression and anxiety from Reddit data. It combines topic-based embeddings and joint task training to improve interpretability and data efficiency over conventional LLM-based classifiers. The main contributions are: 1. It tackles the lack of a fine-grained labeled dataset for Reddit that extends beyond topic-specific subreddits by first curating a labeled dataset and then employing an active learning strategy to help with the training; 2. It proposes a novel multi-task learning model, AMMNet, that outperforms baseline models in the prediction of mental health conditions. 3. A novel model-level explanation behind our prediction due to the introduction of the task-specific feature selector in the task-specific module; and 4. It shows through extensive experiments that for domain-specific classification tasks such as this, a combination of document-level embedding and topic distribution gives the best performance across all the tasks. In the second work, DeMHeM, a multi-task model for identifying bipolar disorder and its comorbidities, is introduced. Through soft parameter sharing and focal loss, the model robustly detects nuanced mental health states and facilitates deeper community-level insight via keyphrase analysis. The main contributions are: 1. development of a novel multitask learning framework for mental health predictions; 2. implementation of a novel and effective multitask optimization algorithm; and 3. exploring post-hoc analysis using the trained model for a more fine-grained understanding of bipolar disorder and its comorbidity. The third work proposes GC-Explainer, an explainable Graph Neural Network for forecasting COVID-19 outbreak severity using only static population features. The model integrates explainability directly into its architecture, enabling transparency without post-hoc methods, and avoids the reliance on real-time or temporal data. The main contributions of the work are: 1. Unlike post-hoc methods for GNN explanation, this work proposes a novel framework, Graph-Covid-Explainer, that simultaneously gives predictions for high-risk areas as well as insights about the most important features during the training of the model. 2. It introduces a novel problem setting that tackles the paucity of historical data to identify high-risk areas during the initial outbreak that can help authorities in better preparing for future crises, and 3. it applies Graph-COVID-Explainer(GCExplainer) on real-world COVID-19 data to show that static features about mobility, socioeconomic status and spatial dependency among regions can be used to make an explainable prediction about the varied degree of severity during the early part of the outbreak, without using historical pandemic data as features.The fourth work proposes to deliver a deployed pipeline that combines Stable Isotope Ratio Analysis (SIRA) and environmental variables using a multi-task Gaussian Process framework. It provides origin tracing for timber samples with predictive uncertainty, significantly improving upon traditional spatial regression approaches. The main contributions are: (1) It presents a comprehensive multi-task Gaussian process modeling framework that supports the incorporation of auxiliary data, such as climate layers, to support origin determination. This enables the incorporation of environmental factors, imputing uncertainty to predictions, and multimodal feature integration; (2) This work is a deployed machine learning pipeline wherein physical samples are collected, subject to tests, and injected into our model to help European enforcement agencies in combating illegal timber trade by demonstrating that a claimed harvest location other than Russia is not viable; and (3) It demonstrates accuracy profiles of our approach in a controlled experiment that illustrates the interplay between SIRA values and atmospheric variables and how they affect our ability to reveal harvest location misrepresentation. This goes beyond traditional ML pipelines that only predict isotope values into an end-to-end approach that supports decision-making by enforcement agencies. The final work combines the concepts of matrix sparsification to extract feature importance with epistemic uncertainty arising from Gaussian processes to make explainable spatial prediction of health outcome in the form of type-2 diabetes. Through a rigorous experimental design, the novel end-to-end Machine Learning framework Deep Graph Gaussian Health Net(DDHG-Net) demonstrates the effectiveness of the model compared to state-of-the-art across different metrics while also providing feature-level insights and uncertainty-aware prediction, making it more suitable for real-world applicability. Our case study on Virginia demonstrates this effectively by identifying a highly prevalent cluster of counties more accurately with high confidence, while uncertain predictions also give insight about which geographical area should conduct a more careful data collection. Overall, the proposed methodological approach laid out in this dissertation promises to be effective in different real-world application domains where explainability is paramount, and the immediate impact of these works lies in greater community welfare.