Hybrid Recommendation from Cross Source Deep Learning Embeddings Based Rating Predictions

Abstract

Collaborative filtering based recommendation systems have data sparsity and cold start problems. Integrating other cues like latent factors in comments, user social profiles etc have been proposed as a solution to the problems in collaborative filtering. But current mechanisms for comments processing don’t effectively process the aspects, sentence conflicts, sarcasm etc. Also they don’t address the rating bias. The current approaches lack implicit finding of aspect orientation of users and fine tuning the recommendation based on aspect orientation. As the result, the accuracy of hybrid recommendation systems is reduced. As a solution to these problems, this work proposes a hybrid recommendation system integrating aspect based bias removal, matrix factorization with residual networks for rating prediction. The bias in rating is removed using a fuzzy Gaussian rating bias removal function. The bias removed rating matrix is converted to feature embedding using matrix factorization and residual network. The feature embeddings are then classified to ratings using softmax classifier. The performance of proposed solution is evaluated against Amazon product review dataset and the solution is found to increase recommendation accuracy by 3% compared to existing works.