Fatemah Alghamedy¹, Jun Zhang² and Maryam Al-Ghamdi¹, ¹University of Kentucky, USA and ³University of Jeddah, Saudi Arabia

The cold-start items, especially the New-Items which did not receive any ratings, have negative impacts on NMF (Nonnegative Matrix Factorization)-based approaches, particularly the ones that utilize other information besides the rating matrix. We propose an NMF based approach in collaborative filtering based recommendation systems to handle the New-Items issue. The proposed approach utilizes the item auxiliary information to impute missing ratings before NMF is applied. We study two factors with the imputation: (1) the total number of the imputed ratings for each New-Item, and (2) the value and the average of the imputed ratings. To study the influence of these factors, we divide items into three groups and calculate their recommendation errors. Experiments on three different datasets are conducted to examine the proposed approach. The results show that our approach can handle the New-Item's negative impact and reduce the recommendation errors for the whole dataset.

Collaborative filtering, recommendation system, nonnegative matrix factorization, item auxiliary information, imputation