A new weighting algorithm for collaborative filtering

Abstract

Similarity measure is an important part of user based collaborative filtering, where the prediction of the rank of the item depends on the similarity between users. In this paper, a novel weighting technique for the similarity measure is proposed in order to increase the accuracy of collaborative filtering. In this technique, the user weight of each category is calculated using the ranks of items that belong to these categories, this process converts the user-item model to user-category model. The sum of the differences between two users' models forms the first part of the weight and it can be used as a similarity measure. The second part of the weight is generated depending on the categories of the item that its rank will be predicted, the active user's weights and the similar users' weights of these categories. Using this weight as a similarity measure then multiplied it with the cosine similarity has improved the collaborative filtering accuracy almost between 2% and 0.6% by changing the number of similar users for MovieLens 100K dataset.