Abstract: Based on users’ session logs, a TF-Ranking recommendation method that integrated XGBoost and gated recurrent unit was proposed. This method used gated recurrent unit to learn user session data. Firstly, XGBoost was used to extract features, and experimental results showed that XGBoost could overcome the defects of traditional data model and greatly reduced the complexity of data models while maintaining their original attributes. Secondly, an improved Dropout network was used to process data, leading to a recall rate improved by 1.32%. Finally, by training data based on Learning to Rank method in combination with pairwise method, a positive sort recommendation list with strong relevance to query contents was provided for users. Experiments were conducted on the data set of Trivago Recsys Challenge 2019. The results show that the proposed algorithm could improve the recall rate and the average reciprocal ranking, and could be applied to large-scale data recommendation.

Key words: recommender system, TF-Ranking recommendation, gated recurrent unit, boosting tree model, Dropout network, dwell time