A Learning Resource Recommendation Model Based on Fusion of Sequential Information

Abstract

The accuracy of learning resource recommendation is crucial to realize precise teaching and personalized learning. Recently, deep learning based models have been introduced in recommendations. Due to flexibilities of deep networks, several kinds of architectures have been proposed to tackle different problems recently. With the development of computing power, research on sequence data is increasing, which is one of the most important features for learning user’s interests. In order to improve accuracy of learning resource recommendation, we introduce deep learning and propose a deep sequence-fusion network (DSFN) based on fusion of multiple sequential data which is deemed to be more effective in learning resource recommendation than in other field. We take the self-attention mechanism as the core part to design the auxiliary subnet and the prediction subnet for fusing multiple sequential data. The proposed model works with joint training and detached predicting. When training, the two subnets both work by respectively compressing the sequential data into self-attention mechanism. Then the sequences output by self-attention layer flow through multiple feedforward neural networks to produce expected targets. When predicting, there is only prediction subnet working, it performs joint prediction of multiple sequential information by fusing the vector learned on training. The experimental results show that compared with the traditional user-based collaborative filtering algorithm, the proposed model improves the accuracy and recall rate by 20.5% and 13.6% respectively.