Improvement of Accuracy and Time Requirements in Traditional Collaborative Filtering Techniques

Collaborative filtering techniques work by estimating a user’s potential preferences on unconsumed items pertaining to the observed preferences of other users. Existing collaborative filtering approaches suffer from two central issues: data sparsity and difficulty in scalability. Neighborhood-based CF approaches find K nearest neighbors to an active user or K most similar rated items to the target item for recommendation, by means of a similarity measure. To find the similarity between a pair of users (a pair of items), traditional similarity measures use their ratings on their co-rated items. Consequently, it has been observed that traditional similarity measures cannot compute effective neighbors in sparse datasets. As for computational scalability, traditional CF approaches are more appropriate for use in a static environment and incorporating new streaming data(ratings) into the system may be a non-trivial task. In a real-world setting, there are always new users, items and ratings that should be incorporated into the recommender systems in an online manner. So we need to design the incremental collaborative algorithms to give real-time responses to the change of data in an effective and efficient way. In the first part of this work, we proposed a new approach to deal with the problem of collaborative recommendation in sparse datasets by gainfully using the concept of structural similarity of nodes in a complex network after the user-user and item-item networks were generated using traditional similarity measures from the given user-item rating matrix. In the second phase of this work, we propose two hybrid approaches HB-1, HB-2, which utilizes structural similarity of both the constructed user-user and item-item network for finding K nearest neighbors and K most similar items to a target item. Subsequently, in the second part of this project, we focus on designing an incremental approach for collaborative recommendations using NHSM similarity measure which should be able to work with dynamic datasets and update the user-user similarity instantly without retraining the whole model. To do so, we have carefully analyzed the parameter training processing of NHSM similarity-based recommenders to design the incremental update rules for involved parameters reflecting data increments in dynamic environments. Recommendation results on a set of real data show that the proposed approaches outperform existing neighborhood-based CFs w.r.t various evaluation metrics