A Method of Purchase Prediction Based on User Behavior Log

In this paper, we propose a method to predict next-one-day-purchase behavior of "Online to Offline"(O2O) items based on huge scale of user behavior log. The overall solution is described in 2 parts: the feature engineering of the user behavior log and the ensemble of different supervised learning models. In the feature engineering section, besides the basic features, we further analyze the behavior of mobile users and propose some special features in O2O scenarios. Those scenarios includes the group-based rank, the transition rate, the centralized proportion, re-buy patterns, geohash-related features and etc. which have improved the model a lot in practice. Besides, group-based rank could be easily extended to other similar business scenarios. Next, model ensemble are tuned in 2 ways: 1) blended models that are trained randomly sampled data to enrich the diversity of the training data to boost the performance; 2) training individual model for different patterns of user-item pair, like the next-day-purchase prediction, re-buy patterns and etc. Finally, a blended of the above models are used to build the prediction result. To evaluate the proposed method, we use the data provided by Ali Mobile Recommendation Competition held in 2015 which consisted of the behavior logs of items from mobile users in one month from Nov. 18, 2014 to Dec.18, 2014, and predict purchase behavior of O2O items in Dec.19, 2014. The result is evaluated under F1 prediction score metric and it achieve a good score as 8.64% that ranks 4th over more than 7,000 teams in the competition.