Graph Exploration Matters: Improving both individual-level and system-level diversity in WeChat Feed Recommender (2306.00009v1)

Abstract: There are roughly three stages in real industrial recommendation systems, candidates generation (retrieval), ranking and reranking. Individual-level diversity and system-level diversity are both important for industrial recommender systems. The former focus on each single user's experience, while the latter focus on the difference among users. Graph-based retrieval strategies are inevitably hijacked by heavy users and popular items, leading to the convergence of candidates for users and the lack of system-level diversity. Meanwhile, in the reranking phase, Determinantal Point Process (DPP) is deployed to increase individual-level diverisity. Heavily relying on the semantic information of items, DPP suffers from clickbait and inaccurate attributes. Besides, most studies only focus on one of the two levels of diversity, and ignore the mutual influence among different stages in real recommender systems. We argue that individual-level diversity and system-level diversity should be viewed as an integrated problem, and we provide an efficient and deployable solution for web-scale recommenders. Generally, we propose to employ the retrieval graph information in diversity-based reranking, by which to weaken the hidden similarity of items exposed to users, and consequently gain more graph explorations to improve the system-level diveristy. Besides, we argue that users' propensity for diversity changes over time in content feed recommendation. Therefore, with the explored graph, we also propose to capture the user's real-time personalized propensity to the diversity. We implement and deploy the combined system in WeChat App's Top Stories used by hundreds of millions of users. Offline simulations and online A/B tests show our solution can effectively improve both user engagement and system revenue.

Enhancing Diversity in Recommender Systems: A Graph-Based Approach in the Context of WeChat

This paper addresses key challenges in recommender systems, particularly those concerning the balance of individual-level and system-level diversity in the context of the WeChat feed recommender. Recommender systems generally follow a three-stage pipeline: candidate generation, ranking, and reranking. The authors propose a novel framework that integrates graph-based retrieval strategies and reranking mechanisms aimed at enhancing diversity at both individual and system levels.

Problematic Aspects in Current Systems

The authors identify a common issue in industrial recommender systems where graph-based retrieval is often biased towards heavy users and popular items, thereby reducing diversity. Although Determinantal Point Process (DPP) has been traditionally used to address individual-level diversity in the reranking phase, it suffers from dependencies on item semantics, which may be inaccurate or misleading. The paper argues that optimizing diversity should be a comprehensive process considering both individual-user experiences and overall system diversity.

Proposed Framework and Methodology

The paper proposes an integrated approach that utilizes graph embeddings from the retrieval stage to enhance diversity in the reranking phase. This methodology weakens item similarity biases, leading to enhanced system-level diversity through graph exploration. By incorporating graph embeddings in both stages, this approach improves the selection of items that are less semantically similar but contextually relevant, thereby enriching the diversity presented to end-users.

A key innovation in this approach is the capturing of real-time personalized diversity preferences. User behavior analysis reveals that preferences for diversity are dynamic, prompting the development of a model that adapts these preferences over time. This real-time, personalized adaptation is facilitated by leveraging temporal sliding windows that assess a user’s historical interactions, allowing the recommendation system to adjust diversity in the items presented proactively.

Results and Implications

The proposed method was deployed across WeChat’s Top Stories platform, and its efficacy was validated through both offline simulations and online A/B testing. Empirical results demonstrate significant improvements in user engagement metrics, including increases in video views and total user stay time, as well as boosted advertisement exposure rates. Such improvements suggest the framework's potential to enhance user satisfaction and system revenue simultaneously.

This work has substantial implications for real-world recommendation systems where achieving the right balance of individual-level and system-level diversity can influence long-term user engagement and system sustainability. Additionally, the effective utilization of graph-based methods provides a scalable and efficient solution deployable at a web scale.

Future Directions

Exploring more nuanced user models and extending the approach to various content domains beyond feed recommendations in social platforms could represent potential avenues for future research. Moreover, further refinement of graph-based diversity metrics and their integration with other machine learning models could provide deeper insights and more robust recommendations in industrial settings.

Overall, the paper offers an insightful and practical contribution to the field, setting a precedent for future research on diversity optimization in large-scale recommendation systems.