Applying Machine Learning Techniques to Tag Recommendation in Social Bookmarking Systems

초록

Most social bookmarking systems allow users to assign a set of suitable tags to web resources. The tags annotated in such a collaborative way are used for categorizing, navigating, and searching the gathered information. This kind of classification system (a.k.a. folksonomy) is a defining characteristic of Web 2.0 services. Many social bookmarking systems also provide a tag recommendation service, which could promote and facilitate the tag annotation process as well as keep the folksonomy well maintained. In this paper, we propose a machine learning-based approach to personalized and content-specific tag recommendation in social bookmarking. For each bookmarking activity consisting of a user and a resource, a set of candidate keywords are extracted from three kinds of information sources: the resource description, previously annotated tags to the same resource, and previously annotated tags by the same person. Each candidate keyword is assigned its importance values calculated from past annotation patterns. Then, a machine learning technique is applied to merge the importance information and ranking the candidate keywords. In our experiments, various machine learning techniques such as artificial neural networks, decision trees, and nearest neighbor methods were evaluated by diverse measures, i.e., AUC (area under the ROC curve), recall, precision, F1-measure, and NDCG (normalized discounted cumulative gain) as well as learning and inference time. The evaluation was performed on the ECML PKDD discovery challenge 2009 dataset obtained from a real social bookmarking system. As a result, artificial neural networks outperformed other methods in terms of recommendation accuracy at an acceptable response time. Our approach also performs better than the best result of the challenge especially in top recommendations.

키워드