Dual-Perspective Video Popularity Prediction from Audience Response to Platform Context
Year: 2026
Type: conference-paper
Abstract: Social media popularity prediction (SMP) estimates the future attention of social posts and supports recommendation, behavior modeling, and platform optimization. Existing multimodal and retrieval-enhanced methods usually map content, creator, and temporal features directly to a final score, but they underrepresent time-varying publication context such as seasonal trends, campaigns, and real-world events. We propose PEAR, a dual-perspective framework for video popularity prediction. PEAR generates unified LLaVA-NeXT descriptions, retrieves semantically similar historical videos, and combines retrieval statistics with structured metadata in a LightGBM audience-side estimator. It then builds event-conditioned context features from semantic compatibility, lexical/hashtag evidence, temporal activation, and context prevalence, and trains a second LightGBM on out-of-fold residuals to refine the base estimate. PEAR achieves a MAPE of 0.1529 and ranks first in the SMP Challenge 2026 Video Track.
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7225345910