TL;DR
MBSSL (Multi-behavior Self-supervised Learning) is a new framework that boosts recommendation accuracy by effectively leveraging multiple user actions (like clicking or adding to a cart) to predict a target behavior (like purchasing). It uses a dual-level contrastive learning strategy to handle data sparsity and noise, paired with a novel "gradient surgery" technique that prevents auxiliary tasks from distracting the model from its primary goal.
What Problem Does This Paper Address?
In the world of e-commerce and social media, "target" behaviors—the ones that actually drive revenue, like purchases or downloads—are notoriously sparse. A user might click on a hundred items but only buy one. While these "auxiliary" behaviors (clicks, views, favorites) provide a wealth of data, using them effectively is tricky for two reasons:
- Noise and Sparsity: Auxiliary data is often noisy (a click doesn't always mean intent), and the target data is too thin to provide a strong signal on its own.
- Optimization Imbalance: Most models treat learning from auxiliary behaviors as a secondary task. However, if the model focuses too much on predicting clicks, it might actually get worse at predicting purchases. This "task interference" often leads to sub-optimal performance because the gradients (the mathematical signals used to update the model) from different tasks can conflict in direction or magnitude.
Methods
The authors propose the MBSSL framework, which consists of three main pillars:
- Behavior-aware GNN: They use a Graph Neural Network (GNN) with a self-attention mechanism. This allows the model to learn not just user-item relationships, but also the dependencies between behaviors (e.g., how a "favorite" action typically precedes a "purchase").
- Dual-Level Self-supervised Learning (SSL):
- Inter-behavior SSL: This contrasts the target behavior with auxiliary ones to transfer knowledge. Crucially, they use the Swing algorithm to identify and eliminate "false negatives"—similar users who might be incorrectly treated as opposites in standard contrastive learning.
- Intra-behavior SSL: This focuses solely on the target behavior, using Edge Dropout (randomly removing connections in the graph) to create augmented views. This helps the model stay robust against the noise introduced by auxiliary data.
- Hybrid Manipulation on Gradients (HMG): This is the paper's most novel contribution to optimization. Instead of using fixed weights for different tasks, HMG adaptively "rectifies" the gradients. If an auxiliary task's gradient is too large or points in a direction that conflicts with the target task, HMG projects it onto a normal plane and scales it down. This ensures the auxiliary tasks support rather than supplant the main recommendation goal.
Datasets
The model was evaluated on five real-world datasets:
- Beibei: Infant products e-commerce (Page View, Cart, Purchase).
- Taobao: China’s largest e-commerce site (Page View, Cart, Purchase).
- Tmall: Another major e-commerce platform (Page View, Favorite, Cart, Purchase).
- IJCAI-Contest: A retail system dataset similar to Tmall.
- Videos: An internal dataset from Kuaishou, a short-video platform (Click, Like, Comment, Download).
Key Insights and Ideas
- False Negatives Matter: In contrastive learning, we usually assume any two different users are "negatives." MBSSL recognizes that two users who haven't interacted might still be very similar. By filtering these out using the Swing algorithm, the contrastive signal becomes much cleaner.
- The "Gradient Tug-of-War": The authors visualize how auxiliary tasks can "dominate" the training process. Their HMG strategy treats optimization as a multi-objective problem where the target task always has the right of way.
- Self-Correction: By combining inter-behavior (learning from others) and intra-behavior (learning from self) SSL, the model creates a check-and-balance system that mitigates the risk of "negative transfer"—where learning from clicks actually hurts purchase prediction.
Key Results
MBSSL consistently outperformed ten state-of-the-art baselines (including GNN-based models like LightGCN and other multi-behavior models like GHCF and CML).
- Accuracy Boost: On the Beibei and Taobao datasets, MBSSL achieved improvements ranging from 18.64% to 19.71% in terms of Recall@10 and NDCG@10 over the best-performing baselines.
- Robustness: The model showed significantly less performance degradation when artificial noise was added to the auxiliary data compared to other SOTA methods, proving the effectiveness of the intra-behavior SSL and HMG components.
- Cold-Start Performance: MBSSL performed particularly well for "inactive" users (those with very few interactions), which is a classic pain point for recommendation engines.
Limitations and Future Work
- Structural Information: The authors note that while they capture behavioral dependencies, there is room to more deeply integrate the complex structural information of the user-item graph.
- Computational Complexity: While the authors claim the complexity is comparable to existing SOTA models, the use of the Swing algorithm and gradient manipulation adds overhead that might be challenging for real-time, massive-scale production environments without further optimization.
- Hyperparameter Sensitivity: The performance relies on tuning several key hyperparameters (like the "relax factor" in HMG and the number of eliminated false negatives), which may vary significantly across different domains.
Summary updated