Improving Memory-based Collaborative Filtering Using A Factor-based Approach

Contents

Collaborative Filtering (CF) systems generate recommendations for a user by aggregating item ratings of other like-minded users. The memory-based approach is a common technique used in CF. This approach first uses statistical methods such as Pearson’s Correlation Coefficient to measure user similarities based on their previous ratings on different items. Users will then be grouped into different neighborhood depending on the calculated similarities. Finally, the system will generate predictions on how a user would rate a specific item by aggregating ratings on the item cast by the identified neighbors of his/her. However, current memory-based CF method only measures user similarities by simply looking at their rating trends while ignoring other aspects of overall rating patterns. To address this limitation, we propose a novel factor-based approach by incorporating user rating average, user rating variance, and number of overlapping ratings into the measurement of user similarity. The proposed method was empirically evaluated against the traditional memory-based CF method and other existing approaches including case amplification, significance weighting, and z-score using the MovieLens dataset. The results showed that the prediction accuracy of the proposed factor-based approach was significantly higher than existing approaches.