Advertising CTR Prediction Based On Multiple Sets Of Features And Model Fusion

With the continuous development of online advertising,computing advertising has become an emerging important discipline.The advertisement click-through rate prediction algorithm is an important part in the calculation of the advertisement delivery system.Improving the accuracy of the advertisement click prediction is of vital importance to improving the revenue of the advertisement delivery system.The difficulty in advertising click-through rate prediction is that its features are mostly discrete and feature sparsity is high.Traditional machine learning classification algorithms such as logistic regression require a lot of feature engineering to solve this problem.The factorization machine model can learn second-order cross features,and deep learning is good at implicitly learning cross features.Some models of deep learning and factor decomposition machine fusion have achieved good results.Integrating different deep models can effectively improve the effect of ad click rate prediction.This paper studies a deep learning algorithm based on multiple sets of features,designs an improved hybrid expert model for ensemble learning and verifies the effect on a real data set,and finally designs an ad click rate prediction system to optimize system performance.The main contributions of this research are as follows:(1)Research on the data set of advertisement click rate and construct multiple sets of feature engineering.The study found that multiple sets of features perform better on deep learning models such as DeepFM and xDeepFM,which is the theoretical basis for ensemble learning.The research on single model found that the model jointly trained by deep learning and factor decomposition machine model performed well,and provided a theoretical basis for the subsequent research on hybrid models.(2)In order to be able to dig deep into multiple sets of feature information,multiple models are comprehensively applied.This paper studies related ensemble learning methods and designs an improved hybrid expert model for ensemble learning.According to the experimental results of multiple groups of characteristics,the structure of the variable was determined by the experimental method of controlled variables.The final experimental results show that the mixed expert model can increase the AUC index by 1.5%compared to the single model without increasing training time on a large scale.(3)This article builds an advertisement click-through rate prediction system and tunes its online prediction performance,and introduces a cold-start mechanism to increase the robustness of the system.Aiming at the relatively slow running speed of Python language,using Libtorch framework and offline features to build prediction modules in C++,the online prediction rate was increased by 3.5 times.