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) itemsbased on huge scale of user behavior log. The overall solution isdescribed in 2 parts: the feature engineering of the user behaviorlog and the ensemble of different supervised learning models. Inthe feature engineering section, besides the basic features, wefurther analyze the behavior of mobile users and propose somespecial features in O2O scenarios. Those scenarios includes thegroup-based rank, the transition rate, the centralized proportion, re-buy patterns, geohash-related features and etc. which haveimproved the model a lot in practice. Besides, group-based rankcould be easily extended to other similar business scenarios. Next, model ensemble are tuned in 2 ways: 1) blended models that aretrained randomly sampled data to enrich the diversity of thetraining data to boost the performance, 2) training individualmodel for different patterns of user-item pair, like the next-daypurchaseprediction, re-buy patterns and etc. Finally, a blendedof the above models are used to build the prediction result. Toevaluate the proposed method, we use the data provided byAli Mobile Recommendation Competition held in 2015 whichconsisted of the behavior logs of items from mobile users in onemonth from Nov. 18, 2014 to Dec.18, 2014, and predict purchasebehavior of O2O items in Dec.19, 2014. The result is evaluatedunder F1 prediction score metric and it achieve a good scoreas 8.64% that ranks 4th over more than 7,000 teams in the competition.