Answers are grounded in this paper and cite their sources.
Accurate movie revenue prediction can support release and marketing decisions. Existing box-office models use metadata and static content, while prior trailer-based multimodal studies primarily address movie understanding; the incremental predictive value of trailer audiovisual information and its interactions with other movie data remains insufficiently established. We propose TQGIF-Net, which combines task-adapted queries, relation-aware graph attention, pairwise bilinear interactions, and quality-conditioned mixture-of-experts fusion. Unlike a conventional multimodal Transformer or an MoE fusion alone, the framework jointly models these complementary mechanisms for revenue prediction. A 2000-moviemultimodal dataset is constructed from TMDB data, with 1400 movies for training, 200 for validation, and 400 for testing. The model achieves log-space MAE of 0.462, RMSLE of 0.618, and log-space R2 of 0.681. Its RMSLE is 5.4% lower than that of Transformer Fusion-All (0.653), and adding both trailer modalities reduces RMSLE from 0.665 to 0.618 relative to the metadata–text–poster setting. The findings suggest that trailer audiovisual information can complement static movie features in the evaluated sample.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: