TL;DR
This paper introduces CHCF, a novel framework for multi-behavior recommendation that addresses the limitations of existing methods by incorporating user- and item-specific "selection criteria" through adaptive thresholds. This allows for a more nuanced understanding of user preferences beyond simple binary interactions, leading to improved recommendation accuracy.
What Problem Does This Paper Address?
Recommender systems are crucial for navigating the vast amount of information online. In real-world scenarios, users interact with items in multiple ways (e.g., viewing, adding to cart, purchasing). Leveraging these diverse "behaviors" can significantly improve recommendation quality. However, existing methods often struggle with two key issues:
- Implicit feedback ambiguity: A user interacting with an item (e.g., adding to cart) doesn't necessarily mean they strongly prefer it, nor does a lack of interaction mean they dislike it. Current methods often treat all observed interactions as equally positive and unobserved ones as equally negative, missing the subtle differences in preference strength.
- Heterogeneous behavior dependencies: The relationships between different user behaviors (e.g., how likely is a "view" to lead to a "purchase" for a specific user or item) vary greatly. Existing models often fail to capture this dynamic and user/item-specific nature of these dependencies.
Methods
The core of this paper is the Criterion-guided Heterogeneous Collaborative Filtering (CHCF) framework. Here's a breakdown of its key components:
- Non-Sampling Approach: CHCF avoids the complexities and potential biases of negative sampling by treating all unobserved interactions as negative.
- Criterion Learning with Thresholds: Instead of a simple binary (0 or 1) for interactions, CHCF introduces upper and lower thresholds for each user-item pair and behavior type. These thresholds act as "selection criteria," allowing the model to learn more nuanced preference strengths. For instance, an item might need to clear a higher "cart" threshold to be considered a strong candidate for purchase by a particular user.
- User and Item Heterogeneity Matrices: To avoid an explosion of parameters, CHCF models these user- and item-specific thresholds using low-rank user heterogeneity (H) and item heterogeneity (G) matrices. These matrices capture user idiosyncrasies and item properties, respectively, and are learned jointly with the recommendation model.
- Unified Framework: CHCF integrates criterion learning (learning the thresholds) and user preference learning (learning user and item embeddings) into a single, unified framework. This joint optimization allows the model to learn how selection criteria influence preferences and vice-versa.
- Hinge Loss: The framework utilizes a hinge loss function, which is more suitable for learning decision bounds and preference strengths compared to standard regression losses.
- Theoretical Guarantees: The authors theoretically demonstrate that CHCF's loss function is an upper bound of Collaborative Metric Learning (CML) loss, suggesting that CHCF can effectively optimize CML-like objectives in a non-sampling manner.
CHCF is designed to be a general framework that can optimize various underlying collaborative filtering models, such as Matrix Factorization (MF), Generalized Matrix Factorization (GMF), and LightGCN.
Datasets
The authors evaluated their framework on three real-world datasets:
- BeiBei: A dataset for infant products, containing "view," "cart," and "purchase" behaviors.
- Taobao: A large Chinese e-commerce platform dataset with "page view," "cart," and "purchase" behaviors.
- Tmall: Another Chinese e-commerce dataset including "page view," "favorite," "cart," and "purchase" behaviors.
Key Insights and Ideas
- Beyond Binary: The Power of "Selection Criteria": The most significant insight is that simply observing an interaction isn't enough. Users have varying "thresholds" for different behaviors and items, and these criteria are crucial for accurately modeling preferences. CHCF's adaptive thresholds capture this nuance.
- User and Item Specificity Matters: The paper strongly emphasizes that these selection criteria are not uniform. Different users have different decision-making processes, and items have different characteristics that influence how users interact with them. CHCF's ability to model this heterogeneity is key to its success.
- Cascading Relationships are Learned, Not Assumed: CHCF can adaptively learn the relationships between behaviors. For instance, it can discover that "cart" behavior is often a stronger signal for "purchase" than "view" behavior, even if this isn't explicitly programmed.
- Unified Learning for Better Performance: Integrating criterion learning with preference learning in a single framework allows for a more holistic optimization, leading to better overall recommendation performance.
Key Results
- State-of-the-Art Performance: CHCF consistently outperformed existing state-of-the-art multi-behavior recommendation methods across all three datasets and various evaluation metrics (HR@N, NDCG@N). For example, GMF-CHCF showed significant improvements over GHCF, a powerful GCN-based model.
- Efficiency: CHCF demonstrated competitive computational efficiency, often achieving better performance with comparable or even lower training times than other advanced methods, especially those relying on complex GCN architectures.
- Effectiveness on Sparse Data: The framework showed remarkable performance even when users had very few records for the target behavior, highlighting its ability to leverage auxiliary behaviors effectively.
- Robustness to Hyperparameters: CHCF showed stable performance across a range of hyperparameter values, indicating its practical usability.
- Cold-Start Capability: CHCF demonstrated good performance in cold-start scenarios (new users or items), suggesting its suitability for dynamic recommendation platforms.
Limitations and Future Work
- Side Context Knowledge: The authors suggest incorporating side context knowledge, such as user profiles and item textual descriptions, into the criterion modeling process for further improvements.
- More Efficient Models: They aim to extend the CHCF framework to more efficient models for industrial-scale recommender platforms.
- Hyperparameter Tuning: While CHCF is relatively easy to tune, some hyperparameters like the weight for negative entry (
w) and MTL coefficients (λk) still require careful tuning based on the dataset. - Decorated Function Choice: The choice of the decorated function
g(x)can influence performance, and whileg(x) = x^2was found to be effective, further exploration might yield better results.
Summary updated