动态融合个性化表示和查询序列编码的个性化搜索方法

Personalized search is an important mean to improve information retrieval and enhance user search experience, and has been widely concerned in the industrial and academic fields. Based on the analysis of personalized search task and personalized search dataset, two difficulties of personalized search are summarized: query ambiguity and re-finding behaviors. In this paper, we propose personalized search based on dynamic fusion of personalized representation and query sequence encoding (PRQSE) to address two difficulties of personalized search task. Personalized representation of queries using query units fused with knowledge representation learning to solve the problem of query ambiguity. Encoding user query sequences using recurrent neural networks and a two-layer attention mechanism to capture re-finding behaviors and learn more accurate user interest features. The results of the comparison experiments show that the PRQSE model improves 0.9%-29% in MRR and 2.2%-61.7% in MAP compared with other methods. The effectiveness of personalized representation and query sequence encoding was verified by ablation experiments.