Ensure Timeliness and Accuracy: A Novel Sliding Window Data Stream Paradigm for Live Streaming Recommendation

TL;DR

This paper proposes a new data stream design paradigm, dubbed Sliver, that addresses the timeliness and accuracy problem of labels by reducing the window size and implementing a sliding window correspondingly and demonstrates that Sliver outperforms two fixed-window data streams with varying window sizes across all targets in four typical multi-task recommendation models.

Abstract

Live streaming recommender system is specifically designed to recommend real-time live streaming of interest to users. Due to the dynamic changes of live content, improving the timeliness of the live streaming recommender system is a critical problem. Intuitively, the timeliness of the data determines the upper bound of the timeliness that models can learn. However, none of the previous works addresses the timeliness problem of the live streaming recommender system from the perspective of data stream design. Employing the conventional fixed window data stream paradigm introduces a trade-off dilemma between labeling accuracy and timeliness. In this paper, we propose a new data stream design paradigm, dubbed Sliver, that addresses the timeliness and accuracy problem of labels by reducing the window size and implementing a sliding window correspondingly. Meanwhile, we propose a time-sensitive re-reco strategy reducing the latency between request and impression to improve the timeliness of the recommendation service and features by periodically requesting the recommendation service. To demonstrate the effectiveness of our approach, we conduct offline experiments on a multi-task live streaming dataset with labeling timestamps collected from the Kuaishou live streaming platform. Experimental results demonstrate that Sliver outperforms two fixed-window data streams with varying window sizes across all targets in four typical multi-task recommendation models. Furthermore, we deployed Sliver on the Kuaishou live streaming platform. Results of the online A/B test show a significant improvement in click-through rate (CTR), and new follow number (NFN), further validating the effectiveness of Sliver.

Figures (8)

• Figure 1: This figure shows that the ground truth CTR undergoes dynamic changes in sync with the live content change in a live room. The game pictures in this figure represent the live content of the corresponding moment in the timeline. We can observe a significant change in live content and ground truth CTR between an hour ago and an hour later, as the anchor switched games.
• Figure 2: Comparison between conventional fixed-window data stream and our approach: conventional fixed-window data streams face a trade-off between the timeliness and accuracy of the live streaming recommender system. Sliver ensures both timeliness and accuracy through the implementation of sliding windows and re-reco strategy. ST and LT denote service timeliness and label timeliness respectively
• Figure 3: An illustration of the timeline for producing a sample in the streaming recommender system. The time interval between request, impression, and sample production is $\tau$ and $\delta$ correspondingly. $Y^{b}$ is a random variable for the moment of the user behavior that happens in the real world after the impression.
• Figure 4: The overview data streaming evolution in Kuaishou live streaming platform from one-hour data stream to five-minute data stream to Sliver data stream with increasing timeliness. One-hour data stream adopts the moment of request as the starting point for the one-hour window. Five-minute data stream adopts the moment of impression as the starting point for the one-hour window. Our proposed Sliver data streaming utilizes 30s sliding window to balance timeliness and accuracy.
• Figure 5: This figure shows the change in labeling accuracy over time in the Kuaishou live streaming platform. Five minutes after impression, the click label accuracy is about 86%, and the like and follow label accuracy is about 80%.