Prediction Of Online Advertising Conversion Rate Based On Multi-model Integration

With the advent of the big data era,the transformation of traditional industries towards the internet has no long been a voguish catchword.By proactively embracing“Internet+”,a company can not only enhance its competitive edge,but also gain access to a greater amount of development opportunities.For the advertising industry,mastering the internet and big data technologies to predict the conversion rate of online advertisements in various scenarios,and therefore place them precisely,can improve users’ experience of the platform,allowing them to quickly identify the product they are most willing to use.Moreover,it also improves the input-output ratio of advertisers and tap their potential customers.Yet,it is not an easy task to accurately predict the conversion rate of online advertisements.The main reason lies in the high sparsity of user behavior data of online advertisements and the unbalanced distribution of positive and negative samples,which lead to unsatisfactory prediction results by traditional machine learning(ML)models.In light of the above context,this dissertation establishes a feature-engineeringbased multi-dimensional feature framework for advertisements.A Light GBM+XGBoost+LR(Stacking/Blending)model is proposed by combining the method of ensemble learning,aiming to improve the prediction accuracy of online advertising conversion rate.The specific research contents are as follows:In terms of feature engineering,exploratory analysis and feature visualization are first carried out on the dataset from the perspective of advertisement and user features,with the features’ impact on conversion rate discussed.Next,data preprocessing procedures such as data fusion,data cleaning,one-hot encoding of discrete features,discretization of continuous features,etc.At last,feature construction is performed from aspects including basic features,sparse features,sequence features,conversion rate features,and combinatorial features to form a set of online advertisement feature framework with a total of 836 dimensions,which support subsequent investigation of conversion rate prediction.In terms of model construction,taking XGBoost and Light GBM as individual learners and LR as a secondary learner,this study integrates the models through stacking and blending respectively.Experimental results validate the superiority of the proposed ensemble model over traditional ML models.K-means clustering is further performed on the user and advertisement features to evaluate the model’s prediction performance in groups based on the clustering result,demonstrating better performance of the proposed ensemble model on clustered samples.In sheer contrast to conventional modelling approaches,the proposed algorithm fully exploits the advantages of linear models and tree models.It effectively reduces model overfitting through several models with significant inter-model structural discrepancies,thereby enhancing the generalization and accuracy of the proposed modelling.This dissertation is intended to provide a more general machine learning method for predicting the conversion rate of online advertisements,which shall provide both theoretical and practical value for related research and application.