[Paper Review] Behavior-Contextualized Item Preference Modeling for Multi-Behavior Recommendation

Mingshi Yan, Fan Liu+4 more|arXiv (Cornell University)|Apr 28, 2024

Behavioral Health and InterventionsPsychology3 citations

TL;DR

This paper proposes Behavior-Contextualized Item Preference Modeling (BCIPM), a novel multi-behavior recommendation framework that learns behavior-specific item-aware preferences via a Behavior-Contextualized Item Preference Network (BIPN) and enhances them with a Graph Convolutional Network (GCN) module. By isolating and transferring only behavior-relevant preferences to the target behavior, BCIPM reduces noise from auxiliary behaviors and achieves state-of-the-art performance across four real-world datasets, significantly outperforming existing methods in HR@10 and NDCG@10 metrics.

ABSTRACT

In recommender systems, multi-behavior methods have demonstrated their effectiveness in mitigating issues like data sparsity, a common challenge in traditional single-behavior recommendation approaches. These methods typically infer user preferences from various auxiliary behaviors and apply them to the target behavior for recommendations. However, this direct transfer can introduce noise to the target behavior in recommendation, due to variations in user attention across different behaviors. To address this issue, this paper introduces a novel approach, Behavior-Contextualized Item Preference Modeling (BCIPM), for multi-behavior recommendation. Our proposed Behavior-Contextualized Item Preference Network discerns and learns users' specific item preferences within each behavior. It then considers only those preferences relevant to the target behavior for final recommendations, significantly reducing noise from auxiliary behaviors. These auxiliary behaviors are utilized solely for training the network parameters, thereby refining the learning process without compromising the accuracy of the target behavior recommendations. To further enhance the effectiveness of BCIPM, we adopt a strategy of pre-training the initial embeddings. This step is crucial for enriching the item-aware preferences, particularly in scenarios where data related to the target behavior is sparse. Comprehensive experiments conducted on four real-world datasets demonstrate BCIPM's superior performance compared to several leading state-of-the-art models, validating the robustness and efficiency of our proposed approach.

Motivation & Objective

• To address the noise introduced by direct transfer of preferences from auxiliary behaviors to the target behavior in multi-behavior recommendation.
• To model user preferences in a behavior-contextualized manner, capturing item-aware preferences specific to each behavior.
• To improve recommendation accuracy in low-data regimes by pre-training initial embeddings to enrich item-aware preference learning.
• To enhance user representation learning by fusing behavior-specific preferences with high-order neighbor information via a GCN module.
• To validate the effectiveness of the proposed method through comprehensive experiments on real-world datasets with diverse interaction patterns.

Proposed method

• BCIPM employs a Behavior-Contextualized Item Preference Network (BIPN) to learn item-aware preferences specific to each behavior, ensuring that only behavior-relevant preferences are used for the target behavior.
• The model integrates interaction data from all behaviors into a unified graph, enabling the GCN enhancement module to capture high-order user-neighbor preferences across the entire interaction network.
• A pre-training module is applied to initialize embeddings, enriching item-aware preferences—especially critical when target behavior data is sparse.
• The final user representation is formed by fusing behavior-specific preferences from the BIPN and collaborative signals from the GCN module, with a learnable hyperparameter λ controlling the balance between the two components.
• The model uses a joint optimization objective that combines the loss from the target behavior and the auxiliary behaviors, with the BIPN ensuring that only relevant preference signals are propagated to the target task.
• The GCN enhancement module operates on a unified interaction graph that combines all behaviors, allowing for richer collaborative signal extraction without behavior-specific separation.

Experimental results

Research questions

• RQ1 How can we reduce noise in target behavior recommendations caused by the direct transfer of preferences from auxiliary behaviors with differing user focus?
• RQ2 To what extent can behavior-contextualized learning of item-aware preferences improve recommendation performance in multi-behavior settings?
• RQ3 Can pre-training of initial embeddings enhance the learning of item-aware preferences, especially in low-data regimes?
• RQ4 How does the integration of a unified GCN on multi-behavior interaction data compare to separate GCN operations per behavior in terms of performance and information retention?
• RQ5 What is the optimal balance between behavior-specific preference learning (BIPN) and collaborative signal extraction (GCN) in multi-behavior recommendation?

Key findings

• BCIPM achieves state-of-the-art performance on four real-world datasets, with significant improvements in HR@10 and NDCG@10 across all benchmarks.
• On the Tmall dataset, BCIPM achieves HR@10 of 0.1414 and NDCG@10 of 0.0741, outperforming the next best method by a substantial margin.
• On the Taobao dataset, BCIPM achieves HR@10 of 0.1292 and NDCG@10 of 0.0716, demonstrating strong generalization across diverse user interaction patterns.
• The ablation study confirms that the proposed aggregation strategy (agg.) significantly outperforms the separate GCN operation method (sep.), indicating the benefit of unified graph learning.
• The parameter analysis shows that a smaller λ (emphasizing BIPN) yields better performance, confirming that behavior-contextualized item preference learning is the primary driver of improvement.
• The model’s performance is particularly robust on sparse datasets like Yelp and ML10M, where pre-training and behavior-specific preference learning mitigate data scarcity effects.