Abstract
Recommendation systems play a critical role in assisting customers to discover new products from vast product catalogs. Thanks to the recent advance in recommendation models, e-commerce systems now can assist customers by suggesting products based on purchase records, browsing history, and any user-item interactions resulted from past purchase. Recently, deep learning has significantly improved the recommendation systems, primarily due to the learned hidden patterns that are more effective than representations learned through shallow machine learning models, like matrix factorization. As a result, deep recommendation system hosted on the website’s product page now can generate a better list of items to enhance e-commerce purposes to secure a sale. This study proposes a novel deep learning-based ensemble of user-to-item recommendation systems, i.e., Deep Neural Matrix Factorization (DeepNeuMF), to predict a group of items potentially interesting to customers. In particular, an ensemble of two well-established models, namely, Generalized Matrix Factorization (GMF) and Deep Matrix Factorization (DMF), will enable concatenated learnable features to curate an improved item recommendation list presented to potential users. In addition, this study validates the proposed DeepNeuMF on a private customer-product purchases dataset with public recommendation benchmarks and compares it with several state-of-the-art works. Extensive results demonstrate that DeepNeuMF can leverage both recommendation models for improved performance on both public and private datasets.