Abstract
Fan-circle involvement is increasingly associated with organized online collective actions, yet individual-level prediction and interpretable screening remain underdeveloped. Based on a national survey of Chinese fans (N = 435), we reformulate online collective behavior intention into an ordinal three-class classification task (Low/Mid/High) derived from training-set tertiles and propose a band-aware accuracy criterion for tier assignment. A Random Forest (RF) classifier is selected as the primary model and tuned with 5-fold cross-validation. Under the proposed evaluation, RF achieves Accuracy = 0.9318 on a held-out 10% test split, outperforming XGBoost and LightGBM baselines (Accuracy = 0.9091) and an MLP baseline (Accuracy = 0.8636). SHAP explanations identify a compact set of Top-10 item-level predictors for the High class, enabling interpretable identification of modifiable levers. The results complement a theory-driven serial mediation account (fan-circle involvement -> information narrowing -> emotional susceptibility -> collective intention) and provide a deployable, explainable pipeline for research and practice.

This publication has 10 references indexed in Scilit: