AxDFM:Position Prediction System Based on the Importance of High-Order Features

Abstract

The exploration and combination of high-level features is crucial for many machine learning tasks. At the same time, we cannot ignore the different importance of high-level features. In traditional machine learning predictive models, analyzing and combining the original data and manually making these features will undoubtedly increase the complexity and cost of the system. The emergence of factorization machines can use the vector product to represent the interaction of features, and automatically learn features The combination of to get high-order feature interactions not only reduces the complexity of the system, but also increases the diversity of high-order features. We refer to the depth factorization machine (xDeepFM) to generate high-level feature interactions at the display mode and vector level, and The importance of different features is dynamically learned through the squeeze-incentive (SENET) mechanism, and different weights are used for interaction.Then, use the attention mechanism to extract the importance of the obtained high-order features and assign weights, and finally get the prediction classification through the fully connected layer. We further summarized these methods into a unified model, and named the model the Advanced Attention Depth Factorization Machine (AxDFM).