A Hybrid Recommender System Using Weighted Ensemble Similarity Metrics and Digital Filters

Recommender Systems aim to predict the rating or preference of a user given to an item and provide suggestions of further resources that are likely to be of interest. The critical part of the recommender algorithm is finding the similarity metrics, which yield predictions with different accuracies and varieties. In the case of high dimensional feature space, the recommender systems using traditional similarity metrics suffer from several problems such as cold-start problem, scalability, over-specialization and data imbalance. In this paper, a recommender system using weighted ensemble hybrid similarity metric model is proposed by combining two or more traditional similarity metrics such as Pearson correlation, log-likelihood and Tanimoto coefficient. The digital filter is extended and adapted in order to handle the posterior intractability and spatial smoothing of high dimensional recommender space. The proposed recommender system is implemented using Apache Mahout. The evaluation of the model has been done using three large MovieLens datasets consist of 100 thousand ratings, 1 Million ratings and 10 Million ratings. We provide a quantitative and a qualitative evaluation. Interesting conclusions were extracted from the real-time execution of the proposed system on the above-mentioned data sets. The comparison with the results obtained from the traditional recommender systems shows that the proposed system achieved significant improvement in the accuracy for large data sets.