Collaborative Convolutional Autoencoder for Scientific Article Recommendation
I Nyoman Switrayana, Nur Ulfa Maulidevi · 2022
Data sparsity is one of the main problems in the Recommendation System (RS), especially in the Collaborative Filtering (CF). Matrix Factorization (MF) is one of the most popular CF methods used. However, when the rating data is sparse, the performance will decrease. In this study, a Collaborative Convolutional Autoencoder (CCA) is proposed, the content features of the paper are considered to assist MF in building a user-item matrix. Contextual information from the paper is represented using Sentence-BERT (SBERT) and Convolutional Neural Network (CNN). The proposed model is applied to real-world datasets. The results show that recall and Discounted Cumulative Gain (DCG) are higher overall in all experimental scenarios than baseline. The combination of SBERT and CNN was better at extracting contextual content from papers than baseline. As a result, the generated user-item matrix is more accurate and performance improves. Performance increases significantly when the data is used more sparse. This proves that the proposed model is more robust to data sparsity.