Weighted sequence loss based recurrent model for repurchase recommendation

2nd International Symposium on Application of Materials Science and Energy Materials

材料科学;能源学

Chen, Pengda^1 ; Li, Jian^1

School of Computer Science, Beijing University of Posts and Telecommunications, Beijing, China^1

关键词: Context information;    Conventional models;    Important features;    Recurrent models;    Repurchase behavior;    Sequential information;    Time step;    Weighted sequences;   

Others  :  https://iopscience.iop.org/article/10.1088/1757-899X/490/6/062062/pdf
DOI  :  10.1088/1757-899X/490/6/062062

学科分类：材料科学（综合）

来源: IOP

【 摘 要 】

Next basket recommendation becomes an increasing concern. Repurchase recommendation, i.e., predicting which products a user will buy again in a user's next order, is a key subproblem. However, most conventional models are not able to extract the whole important features to describe the customer's repurchase process: Context information and sequential information. In our work, we firstly utilize the causal dilated convolutions and recurrent neural network to capture context information and sequential information in different ways. Furthermore, the information extracted by causal dilated convolutions and recurrent neural network is combined at each time step for recommendation. More importantly, to effectively adapt the repurchase recommendation, we introduce a weighted sequence loss, which is able to ignore invalid logloss at special time steps to guide the RNN combined with causal dilated convolutions(RCCNN) training. A deep experimentation shows that RCCNN is able to explain the customer repurchase behaviors, and provide reasonable recommendation.