Deep Convolutional Neural Network and Multi-view Stacking Ensemble in Ali Mobile Recommendation Algorithm Competition: The Solution to the Winning of Ali Mobile Recommendation Algorithm

Abstract

We proposed a deep Convolutional Neural Network (CNN) approach and a Multi-View Stacking Ensemble (MVSE) method in Ali Mobile Recommendation Algorithm competition Season 1 and Season 2, respectively. Specifically, we treat the recommendation task as a classical binary classification problem. We thereby designed a large amount of indicative features based on the logic of mobile business, and grouped them into ten clusters according to their properties. In Season 1, a two-dimensional (2D) feature map which covered both time axis and feature cluster axis was created from the original features. This design made it possible for CNN to do predictions based on the information of both short-time actions and long-time behavior habit of mobile users. Combined with some traditional ensemble methods, the CNN achieved good results which ranked No. 2 in Season 1. In Season 2, we proposed a Multi-View Stacking Ensemble (MVSE) method, by using the stacking technique to efficiently combine different views of features. A classifier was trained on each of the ten feature clusters at first. The predictions of the ten classifiers were then used as additional features. Based on the augmented features, an ensemble classifier was trained to generate the final prediction. We continuously updated our model by padding the new stacking features, and finally achieved the performance of F-1 score 8.78% which ranked No. 1 in Season 2, among over 7,000 teams in total.