Deep Attention Factorization Machine Network for Distributed Recommendation System

With the proposal of Microsoft's Deep Crossing, Google's Wide & Deep, FNN, DeepFM and other deep learning models, deep learning recommendation model has become the mainstream of recommendation system. These models can realize interactive learning of different features, but do not consider the impact of attention mechanisms on the prediction results. Therefore, we introduce the multi-head attention mechanism into DeepFM, and propose a deep attention factorization machine network (DAFMN). We add an attention layer between the embedding layer of user behavior features and the FM and DNN layers in this model. So that the accuracy of the recommendation results can be improved according to the user's attention to different intersection features. In addition, we use HDFS and Spark to realize distributed storage and offline computing of features to improve parallel processing capabilities. We conducted comparative tests on two public datasets, and show that the GAUC of the DAFMN model achieves 5.24% and 3.53% improvement in RelaImpr compared with the base model.