A New Model for Co-Ratings Based Similarity in Collaborative Filtering

Collaborative Filtering recommender systems use the history of user ratings to construct a usually very sparse user-item rating matrix and then use some similarity measurement algorithm to find similar user/items for prediction purpose. Conventional similarity measures such as PCC or Cosine works only on co-rated items and ignore the rating of items that are not rated commonly. Besides, the accuracy of these measures is not very high. Advanced similarity measures such as NHSM and Bhattacharya Co-efficient give comparable accuracy but ignores the weightage of the number of co-rated items. In this paper, we aim to use not only the rating values of co-rated items but also consider the weightage of the number of co-rated items. Moreover, in our improved similarity measure, named Improved_CF we take into account the ratings of non-co-rated items also. Experiments on four publically available datasets (i.e. CiaoDVD, Filmtrust, MovieLens-100K, Epinions) are used to establish the effectiveness of our proposed model. Both prediction and classification accuracy metrics are used as evaluation parameters.