摘要
Short-form videos (SFVs) now dominate online entertainment, yet their addictive tendencies correlate with reduced attention, diminished learning motivation, limited psychological support, academic stress, and social anxiety. Current identification of Short-Form Video Addiction (SFVA) relies on manual observation, missing early warning signs. We introduce the first automated, early-detection framework grounded in social network behavior. We collect the first SFVA dataset (3,179 users) capturing users' social interactions, and propose Heterogeneous Edge Augmentation for SFVA Detection (HEAD). HEAD applies graph diffusion to mitigate sparsity and missing values, converts user traces into multiple feature modalities, and builds a heterogeneous social graph modeling interactions among users and their topics of interest. Extensive experiments show that HEAD outperforms baselines in early SFVA detection, demonstrating the value of integrating heterogeneous graph modeling with graph-based signal propagation for behavioral risk identification.