Ad Click Prediction by using Domain Adaptation Neural Network

Abstract

Online display advertising is one of the largest businesses for Internet companies and growing each year. Predict- ing the probability of ad click is important for buyers to value bid requests in Real Time Bidding (RTB) setting. Most of previous research use ad impressions, which happened only when the buyer wins in the auction, as a train- ing data for click prediction. Such click prediction models, however, predict the probability of click not only for impression but all bid requests in the real product. This gap suggests that the click prediction model trained with impression data is suffered from selection bias. In this paper, we propose a new click prediction model that uses domain adaptation neural network (DANN) to mitigate this problem. DANN can train both a label predictor and the invariant features between domains at once. Experimental result shows that our proposed method improves the accuracy of click prediction.