플랫폼 비즈니스 공급자 이탈 예측을 위한 Transformer 기반 모델 연구: 콘텐츠 크리에이터 후원 플랫폼을 중심으로

초록

Purpose: This study compares the predictive performance and interpretability of conventional machine learning models and Transformer-based tabular deep learning models for supplier (content creator) churn prediction in a two-sided market platform, and derives managerial implications for proactive churn retention. Methods: Using operational logs from a private chat-based sponsorship platform, we construct tabular behavioral features capturing creator activity, fan engagement, and sponsorship-related signals. To mitigate temporal leakage, we adopt a time-aware train–validation–test split and apply class-imbalance–aware training strategies. Seven algorithms are evaluated (Logistic Regression, Decision Tree, Random Forest, XGBoost, CatBoost, TabTransformer, and FT-Transformer) and assessed with F1, ROC-AUC, Macro F1, and Matthews Correlation Coefficient (MCC). Interpretability is examined at the feature-group level using tree-based feature importance and Transformer attention weights. Results: Transformer-based models consistently outperform conventional baselines, with TabTransformer showing the strongest overall predictive performance. Interpretability analyses also reveal consistent findings across methods, highlighting creator-side activity and persistence, as well as fan-side engagement, as key contributors to churn prediction. Conclusion: Transformer-based tabular models provide strong predictive performance and actionable interpretability under class imbalance, thereby enabling earlier identification of churn risk and more targeted retention interventions.

키워드