CPACS: Customer Preference-Aware Customer Service Solutions Recommendation

With the gradual expansion of the work order data of the electric customer service system, the reliance on manual experience has led to low processing timeliness, unable to effectively discover the true demands of customers, and thus unable to provide customers with high-quality solutions. Most of the current methods use deep learning techniques such as Principal Component Analysis and Neural Networks to analyze the semantics of work orders, but it is difficult to fully capture the semantic information hidden in the work order title and work order description, which leads to a decrease in the performance of solution recommendation. Therefore, this paper proposes a Customer Preference-Aware Customer Service Solutions Recommendation (CPACS) model. In order to enhance the representation of work order data, the model uses customer preference information to "query" work order title sequence and work order description sequence separately, generates the temporal dependency representation of these two sequences, and then uses 1-D CNN and Transformer to capture the local and global temporal dependency information of these two sequences. Then, a new information fusion method, Additive Conv-Transformer Skip (ACT-Skip), is proposed to fuse the local and global dependency information in the work order data to improve the solution recommendation performance. The final experiments show that the CPACS model can perform representation learning on work order data more effectively than the baseline model, thus realizing superior performance in the customer solution recommendation task.