Collaborative Filtering in Social Tagging Systems Based on Joint Item-Tag Recommendations

Jing Peng, Daniel Zeng,  Huimin Zhao, Fei-yue Wang (CIKM-10)
The paper “Collaborative Filtering in Social Tagging Systems Based on Joint Item-Tag Recommendations” introduces a novel framework for collaborative filtering in social tagging systems. In recent years, social tagging has been gaining wide-spread popularity in a variety of applications. Enabling automated recommendation of various kinds in social tagging systems can further enhance this important social information discovery mechanism. However, all of the previous research focuses on recommendations of either items or tags. If we can leverage tag information and integrate it into our system, it could help to improve the performance of recommendation engine.
Firstly, Peng et al. present a structure that integrates all possible co-occurrence information among the three entities (i.e., User, Item, and Tag) into one framework. Different with other work that usually consider only information from a triple of user, item and tag, they try to capture all information. The framework is able to present not only the relationship between user, item and tag but also the relationship from user and item, item and tag, as well as user and tag. Making use of all available information in the framework, they propose a unified user profiling scheme to present the information of each user. One distinctive feature of this scheme is that it presents an additional information that highly correlates to each item, called Hidden Tag.  Whenever an item is being saved by the user, this Hidden tag will be used automatically. Users will be considered to be similar to a certain extent through the Hidden Tag once they have saved the same item. Moreover, they also assume that if users use the same tags, they should have some common interests. Such information is presented as a Hidden Item in their profiling scheme.
For weighting the entries in the profile matrix, they adapt the weighting formula from other work, which considers the ranking order of each tag in the set of tags co-assigned to a bookmark as well as the size of this tag set. They also present a way to assign weight to Hidden Tag and Hidden Item. Because it’s time consuming to calculate similarities between very large sparse matrices, they use LSA for dimensionality reduction. Based on the low-dimensional, weighted user profiles, they compute the cosine similarity between users, and propose a method to recommend a joint item-tag matrix to each user from the similar users, with the tags representing the topics of the target item that might attract the user. The recommended profile matrix for each user consists of four blocks:
  • The joint recommendation result for real items and tag
  • The Hidden Tag Column holds the pure item recommendation result
  • The Hidden Item Row the pure tag recommendation result
  •  The corner entry is zero 


Finally, Peng et al. demonstrate a solution of generating item recommendation from the joint recommendation matrix. They generate a denser type of joint recommendations based on the two denser pure recommendations and then fuse it with the joint real item-tag recommendation. The final joint item-tag recommendation result is computed as a weighted average of two joint recommendations. The recommended items to each user are obtained by marginalizing the final joint recommendation result.
To evaluate their proposed methods, Peng et al. use three real-world datasets (i.e., Delicious, CiteULike and Bibsonomy) and compare with several existing tag-based recommendation algorithms (e.g., SB, PLSA and FUS-UB). They apply 5-fold cross validation to split the training and testing datasets. They use the Precision, Recall, F1 and Rankscore as evaluation metrics. The results show that:
  • The authors’ proposed recommendation approach significantly outperformed the others, their results are consistently on all evaluation metrics across all the datasets.
  • They also indicate that making use of complete information embedded in tagging data, their approach is more robust than existing methods, especially, on the datasets, of which the tag quality is poor.

By leveraging the complete tagging information for recommender systems, Peng et al. propose a framework for collaborative filtering in social tagging systems which received a better performance compared to different other methods. It shows the power of collaborative filtering technique in social systems, especially recommender systems. Although, there have been previous work taking advantage of social wisdom to improve their systems, such as tagging information. This paper presents a different approach of using social information that brings about better results for the task of item recommendation. Therefore, observing social wisdom, analyzing and applying it with appropriate techniques can help to improve our systems.


Popular posts from this blog

[Talk Summary] Machine Learning and Privacy: Friends or Foes?

SoRec: Social Recommendation Using Probabilistic Matrix Factorization

[Talk Summary 3] Personalized Recommendations using Knowledge Graphs