A novel recommendation method based on general matrix factorization and artificial neural networks

Abstract

Collaborative filtering is a successful approach in relevant item or service recommendation provision to users in rich, online domains. This approach has been widely applied in commercial environments with success, especially in online marketing, similar product suggestion and selection and tailor-made consumer suggestions. However, regardless its market penetration, there are still considerable limitations in terms of the accuracy in proposed recommendations stemming from the high-frequency low-relevance user-item bias, data specificities and individual user patterns and needs that may be hidden in data. We propose a novel recommendation approach that improves accuracy and requires significantly less maintenance compared to traditional collaborative filtering. For the experimental evaluation, we use two real datasets and well-known metrics with the results validating our method. Our proposed method outperforms all the alternative recommendation methods for each of the two datasets and metrics and seems holistically effective against alternatives since it requires less settings to be considered without affecting the output.

TD-DNN: A Time Decay-Based Deep Neural Network for Recommendation System

“…Unlike the basic MF approach, where inner products combine the user and movie ratings, DLCRS performs the element-wise multiplication of user and movie ratings. A novel recommendation strategy based on Artificial Neural Networks and Generalized Matrix Factorization is proposed by Kapetanakis et al [45]. It improves the overall quality of the recommendations and minimizes human intervention and offline evaluations.…”

“…Neural Matrix Factorization (NMF) [45]: In this approach, pairs of users-movies embeddings are combined using Generalized Matrix Factorization (GMF), and then, MLP is applied to generate the prediction.…”