[Paper Review] BiRank: Towards Ranking on Bipartite Graphs

Xiangnan He, Ming Gao+2 more|arXiv (Cornell University)|Aug 15, 2017

Complex Network Analysis Techniques32 references3 citations

TL;DR

This paper proposes BiRank, a novel graph regularization-based algorithm for ranking vertices in bipartite graphs by integrating structural link information with prior vertex features (query vectors). It converges to a unique stationary solution and is theoretically grounded via both regularization and Bayesian interpretations, outperforming state-of-the-art methods in popularity prediction and personalized recommendation tasks on real-world datasets.

ABSTRACT

The bipartite graph is a ubiquitous data structure that can model the relationship between two entity types: for instance, users and items, queries and webpages. In this paper, we study the problem of ranking vertices of a bipartite graph, based on the graph's link structure as well as prior information about vertices (which we term a query vector). We present a new solution, BiRank, which iteratively assigns scores to vertices and finally converges to a unique stationary ranking. In contrast to the traditional random walk-based methods, BiRank iterates towards optimizing a regularization function, which smooths the graph under the guidance of the query vector. Importantly, we establish how BiRank relates to the Bayesian methodology, enabling the future extension in a probabilistic way. To show the rationale and extendability of the ranking methodology, we further extend it to rank for the more generic n-partite graphs. BiRank's generic modeling of both the graph structure and vertex features enables it to model various ranking hypotheses flexibly. To illustrate its functionality, we apply the BiRank and TriRank (ranking for tripartite graphs) algorithms to two real-world applications: a general ranking scenario that predicts the future popularity of items, and a personalized ranking scenario that recommends items of interest to users. Extensive experiments on both synthetic and real-world datasets demonstrate BiRank's soundness (fast convergence), efficiency (linear in the number of graph edges) and effectiveness (achieving state-of-the-art in the two real-world tasks).

Motivation & Objective

• To address the lack of a theoretically grounded, generic ranking framework for bipartite graphs that incorporates both structural and prior vertex information.
• To develop a method that converges to a unique stationary ranking while being extensible through principled algebraic and probabilistic interpretations.
• To demonstrate the effectiveness of BiRank in real-world ranking scenarios, including item popularity prediction and personalized recommendation.
• To extend the methodology to n-partite graphs, enabling modeling of more complex relationships such as user-item-aspect interactions.
• To empirically validate BiRank’s convergence, efficiency, and superiority over existing methods on synthetic and real-world datasets.

Proposed method

• BiRank iteratively updates vertex scores by minimizing a regularization function that balances graph smoothness and alignment with the query vector.
• The algorithm is formulated as an iterative power method that converges to a unique solution, ensuring stable and deterministic ranking.
• It is interpreted through graph regularization, where the solution minimizes a quadratic form penalizing deviations from the query vector while preserving graph structure.
• A complementary Bayesian view is established, framing BiRank as a MAP estimation under a Gaussian prior, enabling future probabilistic extensions.
• The method is generalized to n-partite graphs via a tensor-based formulation, enabling ranking in tripartite and higher-order graphs.
• For practical deployment, edge weights and query vectors are parameterized to encode specific ranking hypotheses, such as user preferences or item popularity.

Experimental results

Research questions

• RQ1 How can a ranking algorithm for bipartite graphs be designed to converge to a unique stationary solution while incorporating prior vertex information?
• RQ2 What is the theoretical relationship between BiRank and graph regularization, and how can this be leveraged for principled extensions?
• RQ3 How does BiRank compare to existing random walk-based methods like PageRank and HITS in terms of convergence, efficiency, and effectiveness?
• RQ4 Can BiRank be extended to n-partite graphs to model more complex relationships, such as user-item-aspect interactions?
• RQ5 How effective is BiRank in real-world applications such as item popularity prediction and personalized recommendation?

Key findings

• BiRank achieves state-of-the-art performance in item popularity prediction, outperforming ItemKNN and PureSVD by a relative improvement of 8.3% on average across datasets.
• In personalized recommendation, TriRank (the tripartite extension of BiRank) achieves the best performance, improving by over 10% relative to BiRank and significantly outperforming PageRank and TagRW.
• BiRank converges quickly and scales linearly with the number of edges, demonstrating high computational efficiency.
• The method is robust to data sparsity, outperforming neighbor-based CF methods like ItemKNN on sparse datasets such as Amazon, where items have only 3.9 reviews on average.
• The inclusion of review aspects via TriRank significantly enhances recommendation performance, indicating the value of incorporating fine-grained user feedback.
• Item popularity baselines perform poorly, confirming that modeling personalized preferences is essential for effective recommendation.