Analyzing item features for cold-start problems in recommendation systems

With the development of web technologies, there is lots of information on the web. For effective web searching, the recommendation systems appear on the web. The recommendation systems provide customized information for the personal users. The conventional processes of the recommendation are generally based on the user preferences for the items. This leads to the cold-start problems for new items in recommending since new items have no user preferences. Although there are some studies to alleviate this problem by utilizing item features such as category information, the studies do not provide the validities of the use of item features. Namely, they just use the item features without analyzing features. If a feature draws meaningful recommendation results, there are some reasons that the feature can draw the results. We try to find these reasons. We calculate the uncertainty of item features by applying entropy in information theory and assume that this uncertainty of item features can show the level of reliability for the recommendation results. We verify our assumption by utilizing some tests in movie domain.