Interpretable Graph-Language Modeling for Detecting Youth Illicit Drug Use

TL;DR

Interpretability analyses demonstrate that LAMI reveals meaningful behavioral substructures and psychosocial pathways that align with established risk factors for substance use that align with established risk factors for substance use.

Abstract

Illicit drug use among teenagers and young adults (TYAs) remains a pressing public health concern, with rising prevalence and long-term impacts on health and well-being. To detect illicit drug use among TYAs, researchers analyze large-scale surveys such as the Youth Risk Behavior Survey (YRBS) and the National Survey on Drug Use and Health (NSDUH), which preserve rich demographic, psychological, and environmental factors related to substance use. However, existing modeling methods treat survey variables independently, overlooking latent and interconnected structures among them. To address this limitation, we propose LAMI (LAtent relation Mining with bi-modal Interpretability), a novel joint graph-language modeling framework for detecting illicit drug use and interpreting behavioral risk factors among TYAs. LAMI represents individual responses as relational graphs, learns latent connections through a specialized graph structure learning layer, and integrates a large language model to generate natural language explanations grounded in both graph structures and survey semantics. Experiments on the YRBS and NSDUH datasets show that LAMI outperforms competitive baselines in predictive accuracy. Interpretability analyses further demonstrate that LAMI reveals meaningful behavioral substructures and psychosocial pathways, such as family dynamics, peer influence, and school-related distress, that align with established risk factors for substance use.

Figures (12)

• Figure 1: Comparison between traditional methods and our LAMI for illicit drug use detection. Traditional approaches typically treat survey variables in isolation, relying on feature importance metrics or expert-selected variables, which overlook the interdependent relationships among behavioral and psychosocial factors. In contrast, LAMI models survey responses as relational graphs, enabling the discovery of latent structural risk factors and facilitating deeper, more interpretable behavioral insights through graph-level analysis and LLM-based explanations.
• Figure 2: Overview of the LAMI framework. (a) Graph latent relation learning for capturing implicit dependencies among survey variables, (b) a relational graph structure learning (RGSL) layer for inferring latent question–question connections, and (c) an LLM-enhanced predictor for detecting illicit drug use and generating text-based explanations. These components enable interpretable illicit drug use detection among TYAs.
• Figure 3: Model performance using different hyperparameter ($k_{sim}$ and $k_{att}$) settings.
• Figure 4: Training loss and model performance on YRBS with and without relation matrix in the latent relation learning stage.
• Figure 5: Interpretability analysis of LAMI on YRBS and NSDUH datasets, including (1) the top 10 contributing question nodes, (2) the learned structures among these nodes, and (3) the textual analysis based on the LLM's responses during inference.