Kernel based Collaborative Topic Regression for Tag Recommendation

Yanwei Guo, Hongrong Cheng, Mingshuang Tang, Jiaqing Luo, Shijie Zhou2016 · DOI

Abstract

Tag recommendation is very helpful for users to organize or categorize online resources like music, photos and articles. In recent years, some models, such as collaborative topic regression (CTR) and its variants, have demonstrated promising performance for tag recommendation. In this paper, we propose a novel Bayesian model, called Kernel based CTR (KCTR) to combine kernel based probabilistic matrix factorization which exploits social networks with topic modeling. In contrast to CTR and its existing variants, KCTR model is capable of keeping the real correlation among items rather than ideally assuming the relations among items are mutually independent, which is hardly satisfied in real world. Experimental results on two real datasets show that our method outperforms the state-of-the-art approaches.

Introduction

Tagging systems are vital for organizing the vast amount of information on the web, yet they suffer from inconsistency due to diverse user vocabularies. Tag recommendation systems aim to stabilize these vocabularies by suggesting relevant tags. The core challenge lies in modeling the ternary relationship between users, items, and tags.

While hybrid models like Collaborative Topic Regression (CTR) have successfully combined content analysis (via Latent Dirichlet Allocation) with collaborative filtering, they often rely on the simplifying assumption that item latent features are mutually independent. In reality, items are frequently interconnected through citations or shared user groups. Ignoring these correlations limits the accuracy of recommendations.

Methodology

The proposed Kernel-based Collaborative Topic Regression (KCTR) addresses the independence flaw by injecting "social awareness" into the recommendation engine.

The Kernel Trick and Social Networks

KCTR utilizes a normalized Laplacian kernel to represent social networks and citation graphs. Unlike traditional models that treat items in isolation, this kernel allows the model to explicitly capture the covariance between items. If two papers cite each other or are frequently grouped together by users, the model learns that they likely share similar latent characteristics and tags.

Generative Process and Optimization

KCTR assumes that tags and items are generated from a Gaussian Process (GP). For items, latent vectors are drawn from a GP centered on the item's topic proportions derived from text content. To solve the model, the authors developed an EM-style algorithm using coordinate ascent for Maximum A Posteriori (MAP) estimation, iteratively refining topic proportions and latent factors.

Experimental Setup

The model was evaluated using two real-world datasets from CiteULike, a scholarly reference management service. The "social network" for these items was constructed by linking papers that shared four or more users or had a direct citation link.

Table 1: Statistics of the citeulike-a and citeulike-t datasets used in the experiments.

Results and Discussion

KCTR consistently outperformed baseline models, including Probabilistic Matrix Factorization (PMF), Kernelized PMF (KPMF), and standard CTR across both datasets.

Figure 1: Recall@L performance comparison on the CiteULike datasets. The left panel compares KCTR against basic PMF variants, while the right panel compares it against LDA-integrated hybrid models.

As shown in the results, KCTR achieves a significant performance boost. On the citeulike-t dataset, KCTR reached a recall of approximately 0.30 at L=100, significantly higher than standard CTR (~0.23) and basic PMF (<0.10). This confirms that incorporating social network information via kernels provides a substantial advantage over models relying solely on text or sparse user-item matrices.

Conclusion

KCTR demonstrates the power of breaking the independence assumption in recommendation systems. By leveraging item correlations through kernel functions, it achieves state-of-the-art performance in tag recommendation. Future work may involve exploring tag-side kernels to capture linguistic relationships like synonyms and further optimizing the computational complexity of kernel-based methods for larger scales.

Summary updated

Abstract

Tag recommendation is very helpful for users to organize or categorize online resources like music, photos and articles. In recent years, some models, such as collaborative topic regression (CTR) and its variants, have demonstrated promising performance for tag recommendation. In this paper, we propose a novel Bayesian model, called Kernel based CTR (KCTR) to combine kernel based probabilistic matrix factorization which exploits social networks with topic modeling. In contrast to CTR and its existing variants, KCTR model is capable of keeping the real correlation among items rather than ideally assuming the relations among items are mutually independent, which is hardly satisfied in real world. Experimental results on two real datasets show that our method outperforms the state-of-the-art approaches.
BibTeX
@misc{guo2016kernel,
  author = {Yanwei Guo and Hongrong Cheng and Mingshuang Tang and Jiaqing Luo and Shijie Zhou},
  title = {{Kernel based Collaborative Topic Regression for Tag Recommendation}},
  year = {2016},
  doi = {10.2991/icesame-16.2016.24},
}
Figure