Identifying Non-Intentional Ad Traffic on the Demand-Side in Display Advertising

Abstract

The ad traffic from fraudulent or invalid activities costs advertisers a significant proportion of their ad spending. For advertisers, the ad traffic from fraudulent or invalid activities is non-intentional, and this non-intentional ad traffic should not be considered for ad delivery. In this paper, we would like to safeguard the interests of advertisers by identifying the non-intentional ad traffic from the perspective of the Demand-Side Platform (DSP), which serves the advertisers by managing their advertising budget and delivering ads to the right audience in display advertising. Then, DSPs could filter out the identified non-intentional ad traffic to avoid ad spending on and ad delivery of this traffic. To identify the non-intentional ad traffic, our approach is based on Positive-Unlabeled (PU) learning. In particular, we first extract the features which represent the corresponding access behavior, and label the partial non-intentional ad traffic instances we confirmed. Then, given the labeled non-intentional ad traffic instances and the unlabeled ad traffic instances, we build a model to infer the degree of non-intention for each incoming ad request based on our feature space. Our experimental results show that our approach outperforms the baselines on various metrics on one real dataset.