How Integration helps on Cold-Start Recommendations

The RecSys Challenge 2017 focuses on recommending proper or potentially interested users for job postings. Different from traditional recommendation tasks, most of the items (jobs) to be recommended are cold-items without interaction history. To address this problem, this paper introduces an effective integration method. The whole framework consists integrations in 4 different levels: feature-level, model-level, data-level and approach-level. First, we extract features from users' and jobs' profiles and make further abstractions. Second, we improve the model of Wide and Deep Learning, a method that aggregates both embedding and numeric features. Particularly, to receive better performance, both online and offline user-item interactions are used to train the model. Other methods such as content-based Linear Regression, Item-neighbor, Historical Enhancement and Xgboost are also taken into consideration. Finally, we learn lessons from the field of social choice and experiments show that results of aggregation by voting from different methods give a better performance, receiving 29732 in offline test and 6903 in online test.