Hyperlocal: inferring location of IP addresses in real-time bid requests for mobile ads

Abstract

To conduct a successful targeting campaign in mobile advertising, one needs to have reliable location information from real-time bid requests. However, many real-time bid requests do not include fine-grained location information (such as latitude and longitude) because (1) the device or the application did not collect that information or (2) some components of the real-time bid ecosystem did not forward that information. In this paper, we present a three-step approach that takes as input hashed public IP addresses in real-time bid requests and (1) creates a weighted heterogenous network, (2) applies network-inference techniques to infer fine-grain (but possibly noisy) location information for the hashed public IPs, and (3) uses k-nearest neighbor and census data to assign census block group IDs to those hashed public IPs. Our experiments on two large real-world datasets show the accuracy of our approach to be over 74% for hashed IPs (regardless of their type: mobile or non-mobile) when basing the inference on only hashed public mobile IPs. This is notable since our inference is over 212K possibilities.