Research And Improvement Of The Image Advertising Click-through Rate Prediction

In recent years, with rapid development of internet technology, online advertising has become the main profit way in most internet companies. So the research on imporving the user's satisfaction and the advertisers' profit becomes more and more important.In which, the common research is around on using the mass of user clicks logs to mine user personalized informations. Among all of these research, the most important problem is to improve the ad click through rate(CTR) prediction algorithm's accuracy.Current common advertising CTR prediction methods need user historical behavior data to learn patterns and make predictions. However, the user behavior data sparse and user attribute deficiency problems provide a great challenge to exist advertising CTR predict method. To ease all these problems, this paper introduced better advertising image features to improve existing image advertising CTR method. The improved method's core idea is that better advertising image features can help new ads find corresponding similar ads, then using the historical information of similar ads can improve the advertising CTR prediction accuracy of new ads.In this paper, we first analyzed commonly used image advertising CTR prediction method, and the principle of related technology. Then we research on current image advertising CTR prediction method's image feature extraction and user portrait infer stage. In which, image visual feature have a problem of limited represent ability and task independented, user portrait infer have a problem of low predicted accuracy. To solve these problems, this paper designed a new advertising image feature learning architecture based the convolution neural network model for its amazing feature learning ability. And then this paper use the designed architecture learned advertising image high-level abstract features from the original images and user click feedbacks. Features extracted in this way is more flexible for CTR prediction task and needn't artificial selection and combinations. This paper improved the existing advertising CTR prediction method through the introduction of advertising image high-level abstract features. And the combination of image high-level abstract features and advertising image prediction model eased the new ads' low prediction accuracy rate problem caused by the lack of user behavior data.The experiments of this paper use about 1G image ads click log data published by Tencent in the 2014 CCF technology innovation contest. Experimental results verified that the improved image advertising CTR estimation method can significantly improve the new ads' prediction accuracy. The improvement use the proposed ad image features learning architecture to extract the meaningful feature, which make the new ads' prediction accuracy improved.