A Collaborative Filtering Recommender System Integrated with Interest Drift based on Forgetting Function

Collaborative filtering (CF) is one of the most prevailing and promising approaches in recommender systems. The algorithms precision of collaborative filtering has attracted ever-increasing study of researchers. Traditional user-based approaches for collaborative filtering identify user similarity by analyzing the co-rating items between users and utilize user similarity as predicted weight in order to evaluate the importance of rating from a user on an item. However, other factors are not taken into account, including users’ rating trend and changes of user interest, which will degrade the accuracy of the recommendation result obviously. Therefore, in this paper, user similarity index is introduced to improve user similarity calculation. To assign decreasing weights to dated data, exponential function is implemented to redefine the weight of each item rated at different times. Combining user similarity index with exponential function as the improved algorithm, this paper re-computes the predicted ratings based on traditional user-based CF using Pearson Correlation Coefficient. Experiments on Movielens dataset have shown that the improved algorithm is superior to the traditional one.