A Bayesian Latent Variable Model of User Preferences with Item Context

Aghiles Salah, Hady W. LauwIJCAI 2018 · OpenReview

TL;DR

This paper introduces Collaborative Context Poisson Factorization (C2PF), a novel Bayesian model that enhances recommendation systems by incorporating item-item contextual relationships alongside traditional user-item interactions. By leveraging how items appear together in contexts like shopping carts or browsing sessions, C2PF significantly improves recommendation accuracy, especially for users with limited interaction history.

What Problem Does This Paper Address?

Recommender systems are crucial for helping users discover relevant items, but they often struggle with sparse user-item interaction data. Traditional methods like Matrix Factorization (MF) decompose user preferences into latent factors but don't effectively capture how items themselves are related. This limitation means that even if two items are similar in some latent aspect, they might not be recommended together if they haven't been interacted with by the same users. The paper addresses this by proposing to use "item context" – groups of items that frequently appear together (e.g., in the same shopping cart or browsing session) – as a rich source of information to infer item relationships and improve recommendation quality.

Methods

The core contribution is the Collaborative Context Poisson Factorization (C2PF) model. C2PF is a Bayesian latent variable model that extends Poisson Factorization (PF), a variant of MF well-suited for discrete data like ratings or purchases.

Here's what's new:

  • Item Context Integration: C2PF models user preferences as being influenced by two factors: (1) the direct interaction between user and item latent factors (standard in PF), and (2) the interaction between the user's latent factors and the contextual latent factors of other items that share contexts with the target item.
  • Generative Process: The model defines a probabilistic generative process where user preferences, item attributes, context item attributes, and "context effects" (how much a context item influences another) are drawn from Gamma distributions.
  • Scalable Variational Inference: To handle the complexity of the Bayesian model, the authors derive a scalable variational inference algorithm. This algorithm uses a mean-field approximation and coordinate ascent updates to efficiently approximate the posterior distribution of the model's latent variables, making it practical for large datasets.

Datasets

The authors evaluated C2PF on six real-world datasets from Amazon.com, specifically the "Also Viewed" lists, which were treated as item contexts. The datasets used were:

  • Office
  • Grocery
  • Automotive
  • Sports
  • Pet Supplies
  • Clothing

These datasets were preprocessed to ensure each user had at least ten ratings and each item had at least two ratings, with identical sets of items and context items.

Key Insights and Ideas

  • Context is King (Especially for Sparsity): The most significant insight is that item context provides a powerful signal for recommendation, particularly when user-item interaction data is sparse. C2PF's ability to leverage these contextual relationships helps overcome the limitations of relying solely on direct user-item interactions.
  • Poisson Factorization's Strength: The paper reinforces the effectiveness of Poisson Factorization over Gaussian-based models for recommendation tasks involving discrete feedback, as it outperforms a strong Matrix Co-Factorization (MCF) baseline even without explicit context modeling.
  • Model Assumptions Matter: Comparing C2PF to other context-aware models like CTPF and COCTPF (which use collective matrix factorization approaches) highlights that C2PF's specific Bayesian formulation and assumptions lead to superior performance.
  • User Representation is Hard with Little Data: The experiments show that C2PF provides the most significant improvements for users with very few ratings. This suggests that when direct user preferences are unreliable due to sparsity, contextual item information becomes a more dominant and effective predictor.

Key Results

  • Performance Gains: C2PF consistently outperformed baseline models (MCF, PF, CTPF, COCTPF) across all datasets and evaluation metrics (nDCG, MRR, Precision@M, Recall@M).
  • Impact of Context: The performance improvements of C2PF over the basic PF model (which degenerates to C2PF without context) directly attribute the gains to the item context modeling.
  • Scalability: The proposed variational inference algorithm is efficient, scaling linearly with the number of non-zero entries in the user-item and item-context matrices, making it suitable for large-scale applications.
  • Robustness to Factor Count: The model's performance was relatively insensitive to the number of latent factors (K), with K=100 offering a good trade-off between accuracy and complexity.

Limitations and Future Work

  • Limitations: While C2PF shows strong performance, the paper acknowledges that the "context set" definition is crucial and can be dataset-dependent. The model's performance might be sensitive to how contexts are defined.
  • Future Work: The authors suggest several promising directions:
  • User-User Social Relationships: Incorporating social connections between users could further alleviate sparsity issues.
  • Combining with Textual Content: Integrating C2PF with models that handle textual item content (like CTPF) could provide a more comprehensive understanding of items.
  • User and Item Evolution: Extending the model to account for how user preferences and item characteristics change over time.

Summary updated

Abstract

Personalized recommendation has proven to be very promising in modeling the preference of users over items. However, most existing work in this context focuses primarily on modeling user-item interactions, which tend to be very sparse. We propose to further leverage the item-item relationships that may reflect various aspects of items that guide users' choices. Intuitively, items that occur within the same "context" (e.g., browsed in the same session, purchased in the same basket) are likely related in some latent aspect. Therefore, accounting for the item's context would complement the sparse user-item interactions by extending a user's preference to other items of similar aspects. To realize this intuition, we develop Collaborative Context Poisson Factorization (C2PF), a new Bayesian latent variable model that seamlessly integrates contextual relationships among items into a personalized recommendation approach. We further derive a scalable variational inference algorithm to fit C2PF to preference data. Empirical results on real-world datasets show evident performance improvements over strong factorization models.
BibTeX
@inproceedings{salah2018bayesian,
  author = {Aghiles Salah and Hady W. Lauw},
  title = {{A Bayesian Latent Variable Model of User Preferences with Item Context}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2018},
  url = {https://openreview.net/forum?id=S1ZHnEzdZB},
}
Figure