A New Users Rating-Trend Based Collaborative Denoising Auto-Encoder for Top-N Recommender Systems

Abstract

To promote online businesses and sales, e-commerceindustry focuses to fulfill users' demands by giving them top set ofrecommendations which are ranked through different ranking measures.Deep learning based auto-encoder models have further improved theperformance of recommender systems. Astate-of-The-Art collaborative denoisingauto-encoder (CDAE) models user-item interactions as a corruptedversion of users rating inputs. However, this architecture stilllacks users' ratings-Trend information which is an important parameterto recommend top-N items to users. In this paper, buildingupon CDAE characteristics, we propose a novel users rating-Trendbased collaborative denoising auto-encoder (UT-CDAE) whichdetermines user-item correlations by evaluating rating-Trend(High or Low) of a user towards a set of items. This inclusion of auser's rating-Trend provides additional regularization flexibilitywhich helps to predict improved top-N recommendations. Thecorrectness of the suggested method is verified through different rankingevaluation metrics i.e., (mean reciprocal rank, meanaverage precision and normalized discounted gain), for various inputcorruption values, learning rates and regularization parameters.Experiments on standard ML-100K and ML-1M datasets showthat suggested model has improved performance overstate-of-The-Art denoising auto-encodermodels.