CTR Prediction Method Based on Improved Fi-GNN Model
CTR prediction is the probability of users clicking on a given item.It is very important in recommendation system and online advertising.Most CTR prediction models are implicit modeling feature interaction modules,which can not explain mean-ingful high-order feature combinations.This paper proposes a click rate prediction method based on Fi-GNN improved model(Fi-GNN-SKM).For the problem that the expression ability of the adjacency matrix of the baseline model is not flexible enough,a more flexible adjacency matrix is designed to distinguish the importance of different nodes.Secondly,mixed network of experts(MOE)is used to aggregate the information of different spatial neighbor nodes to get better node feature embedding.Finally,for the feature embedding after state update,the deep neural network is used to continue to capture the high-order feature combination and predict the results.Compared with the Fi-GNN model,the experimental results in the Criteo data set show that the AUC value of the Fi-GNN-SKM method increases by 0.56%and the LogLoss decreases by 0.51%.The experimental results in Frappe data set show that the AUC value of Fi-GNN-SKM method increases by 0.37%and LogLoss decreases by 1.84%,which proves the effectiveness of this method.
CTR predictionadjacency matrixmixed network of expertsdeep neural network
万方数据
维普
