Addressing Item Cold Start Problem in Collaborative Filtering-Based Recommender Systems Using Auxiliary Information

Abstract

AbstractRecommender systems (RS) are becoming increasingly popular and are emerging as a viable answer to the problem of information overload. The collaborative filtering-based (CF) recommender system has trouble recommending new items or to new users because there are no ratings for new items or by new users. These issues are known as new item cold start (ICS) and new user cold start (UCS) problems. In many domains, such as ecommerce, we have auxiliary information such as an image and a textual description for each item. These auxiliary data can be used to produce feature vectors for new items, which can subsequently be used to find similarities between new items and existing items in the system, allowing for their recommendation, and alleviating item cold start problem. In this paper, we propose to extract feature vectors from images and textual descriptions of items using deep neural networks such as pretrained VGGNet and BERT, and then use these feature vectors for similarity calculation, allowing CF-based RS to predict ratings of new items accurately or with low error. We present 3 models that use only image, only text, and both images and text as auxiliary information, and tests on a large dataset from Amazon show that the suggested strategy is effective, with an MAE ranging from 0.77 to 0.86 on a 1 to 5 rating scale.KeywordsItem cold start problemItem-based collaborative filteringRecommender systemsDeep learningAuxiliary information