Recommender Systems with Two-Stream Pyramid Encoder Network
Xuyang Jin · 2022 2nd International Conference on Computer Science, Electronic Information Engineering and Intelligent Control Technology (CEI) · 2022
Rate prediction has always been one of the most important tasks of Recommender Systems (RSs), and because of the popularity of Transformer in Deep Learning, we design a new deep learning model with Transformer to give a more accurate prediction. Following the idea of two-stream network, we propose a two-stream network architecture that incorporates Users' data and Item separately. Considering the new item may not have similar items in the dataset, we use feature pyramid networks (FPN) to extract a higher semantic segmentation of the item in Item-Stream. As for user data, we use Transformer block to extract the rating habit of a specific user. In this way, we reduce the prediction bias caused by different users' rating habits. In the decoder block, we jointly use the feature representations from the two-stream network to get the final rating prediction for a specific user to the new item. For different scale feature representation, we decode the feature separately and use the average value of predictions. Experimental results demonstrate that our model extracts higher level semantic information of the item. So, we overcome the problem of sparse data matrix.