A collaborative filtering recommender system based on user's time pattern activity

Collaborative filtering has been considerably successful in improving recommender systems both in the literature and commercial applications. Most of the algorithms designed up to now consider users' ratings equally and do not pay attention to the fact that users' interests or requirements might change over the time. In this paper a collaborative filtering based recommender system is designed which tries to find each user's interests to each group of items, thus resulting to a better prediction of ratings a user will give to an item in the near future. This goal is achieved through using the ratings' timestamp, predefined groups of items, and defining a new similarity measure among users. Unlike standard collaborative filtering methods and many new ones in which similarity between users is defined as a single number, in this research we define similarity between users as “group similarity” which is an array of similarity values between items of each group rated by two users. Predefined groups for items e.g. genres for movies, are used as groups for items. Also for calculating similarity, different weights will be dedicated to ratings of each user based on the ratings' timestamp, i.e. a rating with higher timestamp will receive a higher weight. Empirical tests show that our proposed algorithm works better than standard User-based and Item-based collaborative filtering methods in the case of predicting users' interests in the near future with higher precision. Also it is empirically shown that our algorithm works considerably well for cold-start users.