Reproduce “ContentCTR: Frame-level Live Streaming Click-Through Rate Prediction with Multimodal Transformer” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiaxin Deng, Dong Shen, Shiyao Wang, Xiangyu Wu, Fan Yang, Guorui Zhou, Gaofeng Meng
ContentCTR, a multimodal transformer architecture, predicts Click-Through Rate (CTR) at the frame level for live streaming content by integrating visual, audio, and text features with streamer identity. This approach achieves superior performance in identifying engaging moments, yielding a +2.9% CTR increase and +5.9% live play duration in production on KuaiShou's platform.
Key insights
- Frame-level prediction is essential for capturing dynamic user interest fluctuations within live streams, unlike traditional item-level approaches that average engagement over entire broadcasts.
- Temporal misalignment between different modalities (visual, audio, text) in live streaming data needs explicit handling, which Dynamic Time Warping (DTW) can address for robust multimodal feature fusion.
- A boundary-aware pairwise loss function effectively prevents over-optimization and stabilizes training in ranking tasks, leading to more consistent performance improvements compared to standard pairwise losses.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.