Depth Factor Decomposition Machine Based on Multi-kernel Convolutional Neural Network for Click-Through Rate Prediction.

CTR (Click-Through Rate) Prediction is the key to precision marketing of enterprises, which aims to estimate the probability of a user to click on a given item through combination and synthesis of features. In this paper, a novel feature generation method based on multi-kernel convolutional neural network is proposed. The method uses multi-kernel convolutional neural network to generate local features and recombine them to generate new features. By combining it with DNN (Deep Neural Network) further, a unified model Depth Factor Decomposition Machine Based on Multi-kernel Convolutional Neural Network (DFMBMCNN) is defined. DFMBMCNN utilizes multi-kernel convolutional neural networks to identify sparse but important feature interactions, which can learn arbitrary low-order and high-order feature interactions implicitly. Our comprehensive experiments prove that DFMBMCNN outperforms state-of-the-art models and explore the optimal performance of DFMBMCNN.