GeoCUTS: Geographic Clustering Using Travel Statistics

Web-based services often run experiments to improve their products. To carry out an effective experiment and evaluate the results appropriately, there must be a control group and at least one treatment group. Ideally all of these groups are disjoint, so that each user is given a specific treatment. Using geographical locations as units of experimentation is desirable because this does not require tracking individual users or browser cookies. With the popularity of mobile devices, a user may issue queries from multiple geographical locations. Hence, to be used as units of experimentation, geographical partitions should be chosen in a way that reduces transit between regions. The strategy of clustering users by region is common in advertising. Designated marketing areas (DMAs) are specifically designed for this purpose. However, DMAs are restricted to the US and their granularity is inflexible (there are around two hundred in total). Moreover, they are built based on population density - one DMA per metropolitan area - rather than mobile movement patterns. In this paper, we present GeoCUTS, an algorithm that forms geographical clusters to minimize movement between clusters while preserving rough balance in cluster size. We use a random sample of anonymized mobile user traffic to form a graph representing user movements, then construct a geographically coherent clustering of the graph. We propose a statistical framework to measure the effectiveness of clusterings and perform empirical evaluations showing that the performance of GeoCUTS is comparable to hand-crafted DMAs with respect to both novel and existing metrics. GeoCUTS offers a general and flexible framework for conducting geo-based experiments in any part of the world.