Cold Start Reduction Using Residual Neural Networks in Recommender Systems

Abstract

Recommender Systems (RS) are extremely important in some industries because they can generate significant revenue when used efficiently or serve as a significant way to differentiate from competitors. There are three major categories of models for accomplishing the RS: collaborative filtering methods (CFM), and content-based methods (CBM). The cold start problem (CSP) is one of the most problems still facing many researchers. where the CSP describes the difficulty in making recommendations when the users or items are new, continues to be a significant challenge for CF. Traditionally, this problem has been addressed by conducting an additional interview process to determine the user (item) profile before making any recommendations. This paper presents a hybrid model for CSP reduction based on the Residual Neural Networks (ResNets). We used the ULMFiT dataset based on the Arabic Wikipedia corpus with (30k) vocabulary. Different scenarios are used for evaluating the experiments using a supervised machine learning algorithm. The experiment results showed the ability of the proposed method to reduce the CSP, with an F1 score reaching (96.96 %).