Beyond Relevance: Improving User Engagement by Personalization for Short-Video Search

TL;DR

This work introduces $\text{PR}^2$, a novel and comprehensive solution for personalizing short-video search, where $\text{PR}^2$ stands for the Personalized Retrieval and Ranking augmented search system, and achieves the most remarkable user engagement improvements in recent years.

Abstract

Personalized search has been extensively studied in various applications, including web search, e-commerce, social networks, etc. With the soaring popularity of short-video platforms, exemplified by TikTok and Kuaishou, the question arises: can personalization elevate the realm of short-video search, and if so, which techniques hold the key? In this work, we introduce $\text{PR}^2$, a novel and comprehensive solution for personalizing short-video search, where $\text{PR}^2$ stands for the Personalized Retrieval and Ranking augmented search system. Specifically, $\text{PR}^2$ leverages query-relevant collaborative filtering and personalized dense retrieval to extract relevant and individually tailored content from a large-scale video corpus. Furthermore, it utilizes the QIN (Query-Dominate User Interest Network) ranking model, to effectively harness user long-term preferences and real-time behaviors, and efficiently learn from user various implicit feedback through a multi-task learning framework. By deploying the $\text{PR}^2$ in production system, we have achieved the most remarkable user engagement improvements in recent years: a 10.2% increase in CTR@10, a notable 20% surge in video watch time, and a 1.6% uplift of search DAU. We believe the practical insights presented in this work are valuable especially for building and improving personalized search systems for the short video platforms.

Figures (5)

• Figure 1: A brief system overview. the traditional "retrieve-then-rank" pipeline is depicted in blue. The existing non-personalized stages are in green, and personalized modules discussed in this work are highlighted in orange.
• Figure 2: Personalized Retrieval for short video search. We employ collaborative filtering and dense retrieval methods. Each method retrieves topK personalized candidates, then merged and passed to the downstream ranker.
• Figure 3: Behavior modeling of four user action sequences. From left to right are: query GSU, query-target GSU, CP-GSU, and real-time behaviors. The specific definitions are discussed in Section 4.2.
• Figure 4: Analysis of online CTR and Watch Time per Video for different retrieval methods. The plot is based on 24-hour online traffic with over 3B video views.
• Figure 5: A case study from the online search log. The user's implicit need, highlighted in red, absent from the query but discernible through relevant past behaviors.