Research On Advertising Click Prediction Based On Deep Learning

With the rapid development of the mobile Internet,the importance of online adver-tising is getting higher.As a core step in the advertising system,advertising click-through rate prediction(CTR)refers to predicting the probability of a user’s click on a display ad-vertisement under a given scenario.An effective advertising CTR model can bring huge profits to the company.How to establish an accurate,efficient and stable CTR model is a hot topic in the academic and industrial circles.The CTR method based on traditional machine learning has been widely used in the industry,but it is difficult to construct implicit and high-level combined features,and a lot of manual feature engineering is required.The current CTR algorithm based on deep learning uses factorization machines,MLP and other structures to perform low-order and deep-order feature intersections,extract high-level combined features,and establish an end-to-end CTR model.Therefore,this article studies the CTR model based on deep learning.The main work content is divided into the following two aspects:(1)Aiming at the current CTR model for element-level cross features,only partial feature combination information is obtained and the relationship between features is not clear.This paper proposes a parallel advertising CTR based on vector-wise cross fea-tures Model(GFICNN).The low-order structure of the GFICNN model is a factorization machine model.The deep structure uses a multi-layer convolutional neural network to perform global vector-wise cross features to obtain deep high-order combined features.Then through the global average pooling layer,the combined features obtained from the deep and low-order models are merged to obtain the final feature representation of the ad-vertisement log,and then the click-through rate is estimated.In this paper,Criteo’s public logs are processed as an experimental data set for horizontal comparison experiments.The experimental results show that compared with the xDeepFM model,the GFICNN model has a 1.1%increase in AUC,a 1.3%reduction in Log Loss,and a 11%reduction in train-ing time.And further studied the role of the GAP layer on multiple classic CTR models,and found that the use of the GAP layer can bring a certain improvement to the parallel CTR model.(2)Aiming at the situation that the amount of new advertising data is not enough to learn effective embedding representations in the research of cold-start problem,this paper proposes the CTR model ICN and WICN based on attribute interaction.The core idea is to use the features of sufficient exposure Attribute enriches the embedded representation of new ads.The ICN model is generally divided into three parts:the embedding layer,the attribute interaction layer and the feature cross layer.Based on ICN,WICN adds weights to the attribute interactions.The model learns the attributes of users and advertisements to obtain the interaction characteristics,which are used to enrich the embedding representa-tions of the new advertisement.Finally the high-level feature combinations are extracted through the feature cross layer,which are learned to predict the CTR.This article uses the log data provided by AVAZU to construct a data set and conduct comparative experiments to verify the new model on the test set with 10%,20%,and 30%of new ads.The ICN model and the three constructed by WICN The AUC on the data set has been improved to a certain extent,indicating that the attribute interaction features can learn more accurate and effective embedding representations of new advertisements.