User behavior sequence modeling to optimize ranking mechanism for e-commerce search
Chengfu Huo, Yujiao Zhao, Weijun Ren · Proceedings of the 3rd International Conference on Communication and Information Processing · 2017
E-commerce is now playing an increasingly important role in the daily life, and thus search engine is critical for optimizing user experience of online shopping, of which an abundant of researches have been released. The mainstream technology to optimize search engine usually focuses on two basic factors, relativity and personalization, which normally makes use of items' textual information and users' historical behaviors to mine the correlation between search results and users preferences, and to rank the item-list for different users. Specifically, click and non-click are regarded as labels to define whether the current user prefers the given item or not, and then machine learning is introduced to model click-through rate for ranking the item-list returned by search engine. One disadvantage of the above-mentioned approach, however, is that some temporal information and spatial information of user behavior sequence are ignored. Besides, model training using the search log directly leads to positional bias. This paper models the user behavior sequence to optimize ranking mechanism for E-commerce search. In detail, we first propose a sampling framework in accordance with Markov process to acquire training instances. Spatial sampling is implemented in every page view to avoid the positional bias, and temporal sampling is implemented to utilize behavior information of the sequence of page views for feature design. Then LR and DNN are applied for model training. Experimental results are presented at last. To be specific, the testing and training AUC of LR can be 0.614 and 0.784, respectively, and LR has better performance than that of DNN.