Psychological trait mining of juvenile offending from large language models and online behaviour analytics

Adolescent online traces can reveal early shifts toward harmful trajectories, yet signals are scattered across text, timing, and peer exposure.To address fragile single-modality profiling, this paper proposes an evidence-linked framework that couples behaviour episode graphs with Transformer-based language representations.First, raw logs are segmented into sessions and converted into a heterogeneous interaction graph to encode rhythm and exposure.Then, event-linked texts are embedded to capture stance and intent cues.Finally, an adaptive fusion learner predicts multi-label psychological trait proxies and a risk score with traceable evidence.Experiments on a de-identified dataset of 8,420 users and 3.6 million events show the proposed method achieves AUC 0.879 and F1 0.821, improving over the strongest single-modality baseline by 0.037 AUC and 0.044 F1, with higher precision 0.833 and recall 0.815.The results indicate robust, interpretable profiling for research-oriented prevention.