As resource spaces become ever larger, the need for tools to help users find pertinent and reliable resources quickly and easily is more and more acute. Recommender systems are an efficient way to tackle the problem of information overload, as they enable to inference from users' past behavior to suggest resources the users have not seen yet. Collaborative filtering, that uses ratings of user on items, has become a very popular technique for recommender systems. One crucial step of item-based collaborative filtering is the computation of pair-wise similarity between items. In order to improve the accuracy of collaborative filtering, we propose a new approach for the computation of similarities, by using mutual information. In this work, only positive and negative evidences are considered for training, whereas classical approaches use ratings. During the prediction step, we do not use the classical $K$ value (KNN neighbors) as the size of the neighborhood of each item, but a size value that depends on the considered item. The proposed method is evaluated on the well-known MovieLens dataset. Our experiments show an improvement in the accuracy of collaborative filtering.