Improvement of image tag recommendation using favorite context in image folksonomy

Abstract

Image tag is used for effective image management in an image folksonomy increasing at an accelerating rate. Tagging in social web site depend on manual annotation of user. Therefore, a lot of images in the image folksonomy are weakly annotated or even not annotated any tag because the manual image annotation is a time consuming and cumbersome works for users. Image tag recommendation is a promising technique which would allow users to select appropriate tags among recommended tags. The image folksonomies are used to improve the result of tag recommendation techniques. When considering the size of the image folksonomy, recommending image tags using the image folksonomy requires high computation costs. The relevance of recommended tags using tagging history of user is depending on the user tagging behavior and suggests limitation topic tags. In this thesis, we proposed an image tag recommendation method using favorite context which explicitly shows users` interests. The favorite context consist of images users` manually marked to represent interests from others photo collections. The effectiveness of the proposed method is verified by comparing the recommendation accuracy when making use of other contexts from the image folksonomy. The performance of image tag recommendation using proposed method shows about 29% relative improved result better than using folksonomy evaluated by P@5. Moreover, we analyzed the affect of social activity of users to the image tag recommendation. We divided users into four groups according to their social activity, and compared the performance of tag recommendation, respectively.