Cold Start Purchase Prediction with Budgets Constraints

IJCAI-16 Contest Brick-and-Mortar Store Recommendation with Budget Constraints is about buyer nearby brick-and-mortar stores recommendation. The main task of this competition focuses on predicting nearby store buying action when users enter new areas they rarely visited in the past. The contest has two novelties: first, given huge amount of online user behavior with on-site shopping record of moderate size, we could investigate whether their correlation helps in recommending nearby stores. Second, every merchant’s budget constraints are imposed on prediction. We develop a set of useful features to capture the underlying structure of the data, including merchant-related feature, merchant-location feature and so forth. We also propose to learn topic features with embedding, which represent the user on-site shopping behavior pattern and nearby stores behavior pattern in a shared feature space. We use gradient boosting decision tree as the purchase prediction base model, and use isotonic regression as the purchase probability calibrated model. We use cascade ensemble method, and then develop a merchant recommend framework to decouple the budget constrain. Finally, our team MCMC ranked #2 in the competition.