Teen substance-use risk is predicted better by history than by a snapshot

A one-off set of features predicts the onset of substance use worse than a teenager's history of change does. That is the main practical conclusion of a study on data from roughly 11,860 participants in the ABCD Study: for early risk detection what matters is not only the child's profile today but their trajectory.

The authors compared models for predicting first alcohol sip, alcohol use, marijuana use, and combined alcohol and marijuana use. They used ordinary tree-based models, recurrent neural networks and Temporal Graph Convolutional Networks — models that take account of links between people and features.

Among the single approaches, temporal XGBoost performed best: it uses temporal dynamics and generally beat the graph models. But the graphs still proved useful as an extra source of information. When the predictions of temporal XGBoost and T-GCN were combined through score-level stacking, the result was the best for every outcome studied: AUC-ROC above 0.79.

In other words, social context here is no magical substitute for observational data. Links to family and school, along with profile similarity, can add the missing signal, but the foundation is the sequence of changes over time. Among the important features the authors highlight deviant peer behaviour, age, externalising behavioural symptoms, parental monitoring, cultural norms and neighbourhood context.

There is an important limit to interpretation too. An AUC above 0.79 describes model quality on a research task, not a guaranteed forecast for a particular teenager, still less a diagnosis. The abstract does not spell out how the model would work inside a real prevention programme, so turning this result into an automatic risk label would be premature.

The practical lesson is broader than substance use: if the task involves early prediction of behaviour, an archive of observations may prove more valuable than the most sophisticated algorithm. And a graph neural network makes sense not because it is more fashionable but only if its signals genuinely complement the temporal model. If you had to pick one priority for such a system, would you first gather a longer history of observations or build a graph of social ties?

Source: cs.AI updates on arXiv.org