Research On Ad Click-through Rate Prediction Based On Big Data Analysis

With the rapid development of the global online advertising industry,computational advertising came into being.As an ad serving mode that combines computation with online ad serving decisions,it selects the ad that matches the most from the candidate ad library,and pushes it to the target audience in view of the context of the serving request and the audience's characteristics.In the field of computational advertising,ad click-through rate prediction is the most relevant core technology in the entire online advertising industrial chain.This thesis studies the ad click-through rate prediction methods on the basis of actual industrial advertising logs.We systematically explored the data preprocessing methods,feature extraction and representation approaches and the ad click-through rate prediction algorithms through the big data analysis.The major works of this thesis are as follows:(1)In terms of log preprocessing,we propose an abnormal user detection method which is based on the power law distribution.This method takes the click distribution of all users in the dataset into account,and detects abnormal users from the perspective of the statistical analysis,thus it has a good physical meaning and interpretability.Besides,we verify the effectiveness of this method through the experimental design and analysis.(2)As for the feature extraction,we first extract category features from the user,ad and context aspects,then design a framework to construct statistical features.We verify the effectiveness of the framework by analyzing the results in various models.(3)In order to make effective use of the constructed features,we propose an ad click-through rate prediction hybrid model which contains the feature selection.Experimental results show that this model effectively improves the accuracy of the prediction.Based on the above study,we design and implement a big data analysis platform for ad prediction.We deploy the distributed processing framework on a computer cluster,and provide the front-end display interface for the feature construction and the ad prediction.