The Topic-Perspective Model for Social Tagging Systems

Citation

The Topic-Perspective Model for Social Tagging Systems

Caimei lu, Xiaohua Hu, Xin Chen, Jung-ran Park, TingTing He, and Zhoujun Li

Online version

http://www.pages.drexel.edu/~cl389/dataset/kdd10-lu.pdf

Summary

In this paper, authors propose LDA type[1] generative model for social tag annotation. Usually tags associated with a particularly URL belongs either to the content of the URL or the tagger’s perspective about content of the URL. In data mining applications, we would be interested in separating tags associated with the content from tagger’s perspective. In proposed generative model model, we get probability of each tag being associated with content and tagger perspective. In the result section, authors shows that this model improves on previously proposed models for the same task. Tags associated with user perspective can help in improving personalized search.

Motivation for proposed model

• Document is written before a tagger assigns a tag to the document so term generation process for each document should be separated from the tag generation process. They use standard LDA[2] topic model for the term generation process of document.

• When a user generates a tag for a document, it depends either on topic distribution of the document or user’s perspective. They use switch variable to decide whether the user’s perspective or the document topic is used in generation of the tag.

Model

As shown in figure, model is divided in two parts by dashed line. Right part shows the normal LDA[3] generative model. Left part shows how tags are generated. To generate each tag, first an indicator variable x is generated. If x equals 1, then tag is generated using document’s topic distribution. If x is 0 then first a user perspective p is sampled using perspective distribution of user $\theta ^{u}$ then tag t is drawn from the tag distribution of perspective p, $\varphi _{p}$.

Parameter estimation for this model is done using Gibbs sampling.

Experiments and Results

Performance of the model is measured on social bookmarking data set crawled from del.icio.us[4].

Evaluation criterion for experiments is perplexity [5]. As the performance of the model will depend on number of topics and perspectives considered, tuning of these two parameters is done. When number of topics and number of perspective both are set to 80, minimum perplexity was obtained.

Related Papers

• M. Bundschus, S. Yu, V. Tresp, A. Rettinger, M. Dejori, and H.-P. Kriegel, Hierarchical Bayesian Models for Collaborative Tagging Systems, ICDM '09. Ninth IEEE International Conference on Data Mining., IEEE, Miami, Florida, 2009, pp. 728-733.
• D. Newman, C. Chemudugunta, and P. Smyth, Statistical entity-topic models, the 12th ACM SIGKDD international conference on Knowledge discovery and data mining, ACM Philadelphia, PA, 2006, pp. 680 – 686.