[2504.08169] On the Practice of Deep Hierarchical Ensemble Network for Ad Conversion Rate Prediction

View the PDF file from the paper entitled the practice of the deep hierarchical band to predict the advertising rate, by Jinfeng Zuang and 16 other authors

PDF HTML (experimental) view

a summary:Click Predictions with a CTR rate and the conversion rate (CVR) play a decisive role in the success of advertising recommendation systems. The Deep Hierarch Network (DHEN) was suggested to integrate multiple features crossing units and achieved great success in the clicking bank to paper. However, her performance of CVR forecast is not clear in the preparation of transfer ads, as it offers ads to the possibility of user procedures outside the site on a third -party website or application, including purchase, in addition to the cart, etc. 2) How deep and wide DHEN is to achieve the best comparison between efficiency and effectiveness? 3) What are the excessive parameters of choice in each unit intersection of features? Purcent with the structure of the model, the input customization features also greatly affect the performance of the model with a high degree of freedom. In this paper, we attack this problem and make our contributions biased to the aspect of applied data science, including:

First, we suggest a multi -tasking educational framework with DHEN as the structure of the individual spine model to predict all CVR tasks, with a detailed study on how to make DHEN actively work in practice; Second, we build both the user’s behavior sequences in actual time, the conversion serials outside the site for the prediction purposes, and the study of detection on its importance; Last but not another, we suggest losing self -supervising assistance to predict future procedures in the input sequence, to help solve the problem of posters thinking about the prediction of CVR.

Our way of modern performance is achieved compared to the previous individual features crossing with the pre -trained user customization features.