An Item-Targeted User Similarity Method for Data Service Recommendation

Memory-based methods for recommending data services predict the ratings of active users based on the information of other similar users or items, where the similarity algorithm always plays a key role. In many scenarios, we find that the similarity of two users always show different effectiveness when predicting different ratings. Normal similarity algorithms usually do not count the difference, since they originate from statistic and algebra fields and do not directly aim at recommendations. This paper proposes a novel method to amend the user similarity generated by a normal similarity algorithm to more accurately describe the effectiveness of the similarity on a targeted item. We apply our method to improve the Pearson Correlation Coefficient (PCC) algorithm which is one of the most commonly used similarity algorithms. The experiment results on some practical datasets show that our method is slightly better than the original PCC algorithm for predicting ratings in recommendations.