Research On Recommender Systems By Transferring Tags

With the exponential growth of information on the Web, recommender systems have become popular applications. Exploiting social tag information has been a pop-ular way to improve recommender systems in recent years.In recent years, transfer learning has emerged as a new learning framework applied to many applications. It can work under the assumption where the data for training can have different feature space or distribution with the data for test.However, recommender systems could not be improved with tags when tags are sparse. We notice that, while the tags are sparse for recommendation on some target domains, related and relatively dense auxiliary tags may already exist in some other more mature application domains. This inspires us to transfer tags to improve recom-mender systems on cross domain. Many previous works have applied transfer learning to recommender systems to solve the cold start and rating sparsity problems. Most of these applications transfer user-item rating patterns. However, these applications require that the rating data on the auxiliary domain should be dense enough. In this paper, we propose a Tag Transfer Learning (TTL) model for effective cross domain collaborative filtering when both the tagging data and the rating data is sparse. TTL transfers the tagging information on another similar domain with denser tagging data in the form of topics to the target domain. Then, TTL uses the tagging data transferred to assist the recommendation on the target domain. TTL has some novel features over traditional collaborative filtering on cross domain. TTL transfers tag topics, a kind of one-way knowledge, instead of user-item rating patterns which is two-way knowledge. TTL is based on the clustering approach but not matrix factorization. TTL also gives a quantitative analysis on "when to transfer"The experiment was conducted on the MovieLens data set. We compared TTL with traditional user-based collaborative filtering and the tag-based recommenders. The experimental results reveal that, when the rating data is sparse(less than11.5), both the MAE value and the RMSE value of our approach are less than the compared two methods which means that our approach outperforms the traditional user-based collab-orative filtering,the tag-based recommenders. Evaluation reveals that our method can relieve the sparsity problem in recommender systems.