A Statistical Approach to Inferring Business Locations Based on Purchase Behavior

TL;DR

This paper introduces a statistical method to infer physical business locations from transaction descriptions by leveraging customer sharing graphs and known seed locations, enabling spatial mapping of merchants.

Contribution

It presents a novel approach combining graph modeling and maximum likelihood estimation to recover business locations from transaction data, which was previously unavailable.

Findings

Effective in multiple cities with low displacement error

Utilizes a small set of seed locations for accurate inference

Demonstrates robustness on real-world transaction data

Abstract

Transaction data obtained by Personal Financial Management (PFM) services from financial institutes such as banks and credit card companies contain a description string from which the merchant, and an encoded store identifier may be parsed. However, the physical location of the purchase is absent from this description. In this paper we present a method designed to recover this valuable spatial information and map merchant and identifier tuples to physical map locations. We begin by constructing a graph of customer sharing between businesses, and based on a small set of known "seed" locations we formulate this task as a maximum likelihood problem based on a model of customer sharing between nearby businesses. We test our method extensively on real world data and provide statistics on the displacement error in many cities.