Abstract
Researchers from Tsinghua University have developed CMBPR, a recommendation algorithm specifically designed for the "fragmented time" world of short videos. By combining video category data with a hierarchy of user interactions (viewing, replying, and retweeting), the model significantly improves recommendation accuracy and successfully tackles the "Long Tail" problem—ensuring niche content reaches the appropriate audience.
Introduction
Short video platforms, such as TikTok or Reels, represent a unique challenge for recommendation systems. Unlike traditional movie recommendation engines, these platforms must navigate two primary hurdles: the Long Tail Effect and Complex User Feedback. Short videos are inexpensive to produce and often cover hyper-niche topics, resulting in a distribution where most content receives very few views. Furthermore, social interactions on these platforms are nuanced; a simple "view" is a weak signal of interest, whereas a "retweet" or "share" indicates strong preference.
Proposed Methodology
The CMBPR framework builds upon Bayesian Personalized Ranking (BPR), a standard pairwise ranking method. The authors introduce two major innovations to adapt BPR for the short video ecosystem:
1. Multi-Channel Feedback (2MC-BPR)
The model establishes a priority hierarchy for user interactions: Retweet > Reply > View > No Interaction. By training the model to recognize that a retweeted video should be ranked higher than one that was merely viewed, the system captures the rich nuances of user preference that single-signal models miss.
2. Category-Aided Sampling
Standard BPR typically selects "negative" examples (videos a user did not watch) at random. CMBPR instead utilizes biased sampling, frequently comparing a viewed video to another video within the same category. This "inner-class" comparison forces the model to learn fine-grained preferences—for instance, determining why a user preferred a specific sports video over another in the same category.
An advanced version, CMBPR+, dynamically adjusts this sampling based on a user’s specific history. If a user frequently engages with "Pet" videos, the model focuses more heavily on sampling within that category to sharpen its predictions.
Experimental Setup and Dataset
The study utilized a large-scale, real-world dataset from Sina Weibo, one of China's largest social media platforms. The data pipeline involved deep packet inspection (DPI) to capture clickstreams and crawling text descriptions to categorize content.
Dataset Statistics: * Scope: One week of data from April 2016. * Scale: ~10,000 users and 5,000 short videos. * Categories: 9 distinct categories (e.g., Music, Sports, Movie, Baby&Pet) derived from crawled text. * Interactions: 7.5k retweets, 20k replies, and 76k views.
Key Results
CMBPR and CMBPR+ consistently outperformed state-of-the-art baselines, including Popularity-based ranking, item-based Collaborative Filtering (itemKNN), and standard BPR.
- Accuracy Gains: CMBPR+ achieved a Hit Rate (HR@5) of 0.243 and an NDCG@5 of 0.220, a significant improvement over standard BPR's HR@5 of 0.049.
- Solving the Long Tail: The most dramatic improvements occurred in niche categories. For "Sports" and "Movie" categories—which have relatively few videos—CMBPR nearly doubled the recommendation accuracy compared to models lacking category-aided sampling.
- Behavioral Impact: The researchers validated that incorporating three levels of feedback (View, Reply, Retweet) is significantly more effective than using only two (View and Retweet).
Conclusion and Future Work
CMBPR demonstrates that niche content requires context. By forcing the model to compare similar content, the mathematical representations (latent vectors) of niche videos become much more precise. While the current model relies on text-based classification, future work may explore integrating social graph data (social ties) and temporal dynamics to further refine recommendations as user preferences evolve over time.
Summary updated