Sequential Recommender System of Educational Contents with End-to-End Title Feature Extraction for Reducing Utility Gap

Conventional sequential recommender system uses content features extracted by a hand-craft method, and a pre-trained deep learning method. This causes the utility gap between content features and the purpose of sequential recommendations. To reduce the utility gap, we propose a sequential recommender system with end-to-end title feature extraction. Specifically, we develop a convolutional text feature extraction method for title features. Because we can optimize text features for the sequential recommendation in the end-to-end manner, we can reduce the utility gap. Experimental results for real educational contents and teachers’ operating histories on Forestanet show the effectiveness of our proposed method.