Stage-Aware Multimodal Popularity Prediction for TikTok Short Videos: Leakage-Safe Features and Modality Attribution

Md Nadim Hosain, Md Soyaeb Bin Kuddus, Md Emadur Rahman Ekra^*, Md Daud Ibrahim

Abstract

Popularity prediction for short videos supports content moderation, creator analytics, and platform research, yet many models mix information from different moments of a video’s life or evaluate on random splits that leak future information into training. In this study, we develop a stage-aware, leakage-safe multimodal framework for TikTok and evaluate it on 11,884 public TikTok posts from 1,606 creators (June–November 2024) with daily engagement snapshots and day-30 view counts. We build descriptors of creator influence, audio and sound, visual and technical attributes, semantic content, interaction structure, and temporal view dynamics from records available at prediction time, train one gradient-boosted tree model per horizon (cold start and days 1, 2, 3, and 5), and test it on the most recent five weeks of posts. Before any engagement is observed, the model identifies future top-20% posts with an area under the receiver operating characteristic curve (ROC-AUC) of 0.942 (follower-count heuristic: 0.882). After one day, it reaches 0.982 for this task and 0.864 for predicting sustained growth, where the Szabo–Huberman log-linear baseline scores 0.677. An exact Shapley decomposition of test ROC-AUC shows that creator influence carries 68% of the cold-start signal and leads day-1 growth prediction, whereas temporal view dynamics lead from day 2 onward; metadata-level content descriptors add little once the creator is known. Feature-level fusion outperforms decision-level fusion during the first days (day-1 growth ROC-AUC 0.864 vs. 0.843). The framework and protocol provide a reproducible reference for leakage-safe TikTok popularity research.

Keywords

TikTok; short-video popularity prediction; multimodal features; gradient-boosted decision trees; temporal data leakage; Shapley value attribution

Cite This Article

Hosain, M. N., Kuddus, M. S. B., Ekra, M. E. R., Ibrahim, M. D. (2026). Stage-Aware Multimodal Popularity Prediction for TikTok Short Videos: Leakage-Safe Features and Modality Attribution. International Journal of Scientific Advances (IJSCIA), Volume 7| Issue 6: Nov – Dec 2026, Pages 983-999. URL: https://www.ijscia.com/wp-content/uploads/2026/11/Volume7-Issue6-Nov-Dec-No.1121-983-999.pdf

Volume 7 | Issue 6: Nov – Dec 2026