Computational Advertising Conversion Rate Prediction Based On Integrated Learing

The rapid development of the Internet has boosted Internet advertising.Advertising effect is usually evaluated by the click-through rate,whereas advertisers prefer conversion rate as it reflects revenue more directly.Using both rates as indicators of advertising effect can greatly help with advertising pricing and accuracy.Thus,research on the advertising conversion rate is vital for both advertisers and advertising media.The following work was done for the modeling of the advertising conversion rate:(1)Data analysis: First,the training and the test set were analyzed to ensure the accuracy of model evaluation.Second,the data presentation of features in different data sets was studied to understand the data sets.Third,the correlation between the features and results were studied and key factors were sought to prepare for feature construction.(2)Feature engineering: It is primarily about feature construction and selection.First,since the construction of new features was based on advertising conversion,five new categories of features were extracted and conversed from the original ones.Second,since not all features could be used for model building,they were further selected by using the chi-square test,online and offline consistency method and integration approach to get the most appropriate features for modeling.(3)Model building and optimization: Model optimization was mainly based on feature selection and parameter adjustment.First,with the constructed and the original features used as the dataset,LightGBM was used to train the model and predict the test set;the logloss was 0.09643.Second,features that scored low in importance or relevance were deleted,and the model was retrained to get a logloss of 0.09634.Finally,the optimum parameters for the model were sought according to the key parameters guide of LightGBM,and the model was trained again to get a logloss of 0.09632.