Multiple-structure Attentional Network for Click-through Prediction in Recommendation System
Chenhui Li, Kaixiang Yi, Minrui Fei, Wenju Zhou, Xian Wei Wu, Yayu Chen · 2021
In recent years, the total amount of data generated by people has increased exponentially, which makes it a great problem to recommend for users, so recommendation system (RS) plays an important role in this process, and it has been widely used in many online service fields such as e-commerce system and science and technology (S & T) service platform. As an important part of the RS, CTR prediction has been studied in depth, from the beginning, machine learning methods include logistic regression (LR), factorization machines (FM) and field-aware factorization machines (FFM) models, to deep learning methods such as multi-layer perceptron (MLP), product-based neural network (PNN), deepFM and xdeepFM models, and later attention mechanism based methods such as AFM and DIN models, they all lack the application of S & T service platform. The multiple-structure attention (Mul-AN) network model is proposed in this paper integrates the low-order as well as explicit and implicit feature interactions without any manual feature engineering for the data, and introduces the attention mechanism into the embedding layer of the model to distinguish the importance of the interaction of different feature, and will be finally applied it to the Hainan Comprehensive S & T Service Platform. Extensive experiments verify that the model can improve the performance and accuracy of S & T service platform CTR prediction.