Researchers used data from over 11,000 adolescents in a large study to compare different methods for predicting when young people might start using substances like alcohol or marijuana. They found that models that take into account changes over time (longitudinal) performed better than ones that only looked at individual characteristics at one point in time. Adding information about relationships between individuals (graph-based approaches) also improved predictions, especially when combined with longitudinal models. --- Why it matters: This research matters to AI engineers and researchers because it shows how machine learning can be used to tackle complex problems like predicting adolescent substance use. The study's findings highlight the importance of considering temporal patterns and social context in predictive modeling. Source: https://arxiv.org/abs/2608.14578