Optimally Connected Deep Belief Net for Click Through Rate Prediction in Online Advertising

Abstract

Many researches of machine learning aim to improve the click prediction of online advertisement (ads). One important method is to investigate the pairwise relevance among instances on impression data and the global interaction among the key features of instances. However, the feature extraction ability is not effective for large amounts of variables in prediction process. In this paper, we propose a novel model named optimally connected deep belief net (OCDBN) for click prediction with rotation codes whitening technology based on optimal mean removal. According to what we have learned, OCDBN is the first method that tries to utilize optimal mean removal technology to improve the effectiveness of click through the rate prediction. OCDBN is capable of extracting global primary features of input instances with various elements, which can be implemented for single and sequential ads impression. Extensive experiments have been performed to show the effectiveness of our architecture by modeling different types of input instances. After implementing comprehensive experiments in a 16-nodes cluster with 32 vCPUs on Amazon EC2, the results also show that the architecture significantly outperforms the existing models in accuracy, coefficient of determination, sparsity, and perplexity of click prediction.