Personalized Recommendation Based on Co-Ranking and Query-Based Collaborative Diffusion

Xiao Yang, Zhaoxin Zhang, Qiang WangAAAI 2013 · DOI

Abstract

Traditional collaborative filtering (CF) methods often struggle with data sparsity, the lack of distinction between user expertise levels, and the difficulty of incorporating heterogeneous metadata. This paper introduces a two-layer graph-based recommendation framework designed to address these challenges. The first layer employs a co-ranking mechanism to evaluate user expertise and item quality iteratively. The second layer utilizes a query-based collaborative diffusion process across an n-partite heterogeneous graph, allowing for the seamless integration of metadata and the discovery of high-order relationships. Evaluation on the MovieLens-100k dataset shows that the proposed model significantly outperforms standard baselines, including Matrix Factorization.

Introduction

Recommendation systems are pivotal in modern digital platforms, yet standard Collaborative Filtering (CF) approaches face several persistent issues. First, sparsity hinders the system's ability to find patterns when users have few ratings. Second, equality bias treats all user ratings identically, failing to account for varying levels of user expertise. Finally, rigidity makes it difficult to incorporate heterogeneous data—such as genres, demographics, or tags—into a single mathematical model without increasing complexity or sacrificing scalability. This research proposes treating recommendation as a navigation task on a multi-layered graph to overcome these hurdles.

Methodology

The proposed framework consists of two primary components:

Co-Ranking for Users and Items

To address equality bias, the model performs a co-ranking process on a bipartite graph of users and items. Inspired by PageRank, this iterative process calculates:

  • User Expertise: Users who consistently rate high-quality items receive higher expertise scores.
  • Item Quality: Items rated highly by expert users receive higher quality scores. This weighting ensures that ratings from more credible users have a greater influence on the final recommendations.

Query-Based Collaborative Diffusion

The recommendation task is framed as a query on an n-partite heterogeneous graph. This graph connects users and items to various metadata (e.g., categories, locations, tags). The target user serves as the starting point (query vector), and influence spreads through the graph via "collaborative diffusion." The final recommendations are determined by the accumulation of influence at specific nodes, capturing complex, multi-hop relationships that traditional similarity measures often miss.

Experimental Setup

The model was evaluated using the MovieLens-100k dataset, a standard benchmark containing 100,000 ratings, movie genres, and user demographic information.

Key Results

The proposed model was compared against several industry-standard baselines: Item-based KNN, User-based KNN, Random Walk (RW), and Matrix Factorization (MF).

Performance comparison of the proposed model against baseline methods on the MovieLens-100k dataset.

As shown in the results, the proposed model consistently outperformed all baselines across all metrics, including precision at various ranks (prec@5, prec@10), Mean Average Precision (MAP), and Area Under the Curve (AUC). Notably, it surpassed Matrix Factorization, which was the prevailing state-of-the-art method at the time of publication.

Discussion

Several key insights emerged from this study:

  • Expertise Matters: Accounting for user expertise significantly reduces noise in rating data, making the model more robust against outliers.
  • Flexibility through Diffusion: Framing recommendation as a search problem allows for highly flexible queries. By adjusting the starting nodes in the diffusion process, the system can easily handle complex queries like "Users who like Action movies in a specific location."
  • High-Order Relationships: The graph-based approach captures "multi-hop" relationships (e.g., User A -> Tag X -> Movie B -> User C) that are typically lost in "black box" matrix factorization models.

Conclusion and Future Work

While the two-layer graph approach demonstrates superior performance and flexibility, future research could address its current limitations. Integrating temporal dynamics to account for changing user preferences over time is a key area for improvement. Additionally, exploring more complex, higher-order relationships and optimizing the diffusion process for massive-scale graphs remain important directions for future development.

Summary updated

BibTeX
@inproceedings{yang2013personalized,
  author = {Xiao Yang and Zhaoxin Zhang and Qiang Wang},
  title = {{Personalized Recommendation Based on Co-Ranking and Query-Based Collaborative Diffusion}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2013},
  doi = {10.1609/aaai.v27i1.8534},
}
Figure