TLDR: MEMBER is a novel multi-behavior recommender system that addresses the performance gap between recommending ‘visited’ (previously interacted) and ‘unvisited’ (no prior auxiliary interaction) items. It employs a Mixture-of-Experts framework with two specialized experts, each trained with tailored self-supervised learning methods. This approach significantly enhances recommendation quality for both item types, outperforming existing state-of-the-art systems in comprehensive experiments.
In the bustling world of e-commerce, where countless products vie for attention, recommender systems play a crucial role in helping users discover items they might love but otherwise overlook. While many systems traditionally focused on a user’s purchase history, modern multi-behavior recommender systems have evolved to incorporate a wider array of user interactions, such as clicks, views, and items added to a cart, to provide more nuanced and effective suggestions.
However, a recent study by researchers from KAIST, Seoul, Republic of Korea, highlights a significant challenge within these advanced systems: their effectiveness varies considerably between items a user has already interacted with (referred to as ‘visited items’) and those they haven’t (termed ‘unvisited items’). The analysis revealed a substantial gap in recommendation quality between these two item types, and critically, that a single model architecture struggles to perform well on both simultaneously.
To address this dual challenge, Kyungho Kim, Sunwoo Kim, Geon Lee, and Kijung Shin propose a novel multi-behavior recommender system called MEMBER. This innovative framework leverages a ‘Mixture-of-Experts’ (MoE) architecture, where specialized experts are designed to tackle the distinct tasks of recommending visited and unvisited items, respectively.
How MEMBER Works
MEMBER’s core strength lies in its intelligent specialization. Instead of a single, generalized model, it employs two dedicated ‘experts’:
-
Visited-Item Expert: This expert focuses on items a user has already shown some auxiliary interest in (e.g., clicked, added to cart). It’s trained using a self-supervised method called ‘visit-filtering contrastive learning,’ which helps it identify auxiliary behaviors that are strong indicators of a future purchase.
-
Unvisited-Item Expert: This expert is designed to discover and recommend items a user has not yet interacted with through auxiliary behaviors. This is a particularly challenging task, as there’s less direct data. To overcome this, it combines ‘novelty-inferring contrastive learning’ with ‘behavior-generative learning.’ The generative aspect helps the expert learn rich representations even from sparse interaction data by predicting frequent behaviors from rarer ones.
A key innovation in MEMBER is its ‘hard gating’ mechanism. Unlike traditional MoE models that might combine outputs from all experts, MEMBER intelligently selects only one expert for a given user-item pair based on whether the item is ‘visited’ or ‘unvisited’ for that user. This approach not only streamlines the process but also proved more accurate than other combination strategies.
Also Read:
- Deliberate AI Reasoning Enhances Recommender Systems with STARec Framework
- Protecting User Privacy in Recommender Systems: A Deep Dive into Federated Learning Approaches
Impressive Results
The researchers conducted extensive experiments using three real-world e-commerce datasets: Tmall, Taobao, and Jdata. MEMBER was pitted against nine baseline methods, including state-of-the-art multi-behavior recommender systems like MB-HGCN and MULE.
The results were compelling: MEMBER consistently outperformed all baseline methods across various evaluation metrics and settings. In standard recommendation scenarios, it achieved up to a 65.46% performance gain in Hit Ratio@20 over the strongest competitor. More importantly, in item type-specific evaluations, MEMBER showed significant improvements for both visited and unvisited items. The gains for unvisited items were particularly pronounced, demonstrating its effectiveness in addressing this critical and challenging aspect of recommendation.
This research marks a significant step forward in multi-behavior recommendation, offering a robust solution to a long-standing problem. By understanding and explicitly addressing the differences between visited and unvisited items, MEMBER paves the way for more personalized and effective e-commerce experiences. For more technical details, you can refer to the full research paper: A Self-Supervised Mixture-of-Experts Framework for Multi-behavior Recommendation.
