SAGEL: smart address geocoding engine for supply-chain logistics

Abstract

With the recent explosion of e-commerce industry in India, the problem of address geocoding, that is, transforming textual address descriptions to geographic reference, such as latitude, longitude coordinates, has emerged as a core problem for supply chain management. Some of the major areas that rely on precise and accurate address geocoding are supply chain fulfilment, supply chain analytics and logistics. In this paper, we present some of the challenges faced in practice while building an address geocoding engine as a core capability at Flipkart. We discuss the unique challenges of building a geocoding engine for a rapidly developing country like India, such as, fuzzy region boundaries, dynamic topography and lack of convention in spellings of toponyms, to name a few. We motivate the need for building a reliable and precise address geocoding system from a business perspective and argue why some of the commercially available solutions do not suffice for our requirements. SAGEL has evolved through 3 cycles of solution prototypes and pilot experiments. We describe the learnings from each of these phases and how we incorporated them to get to the first production-ready version. We describe how we store and index map data on a SolrCloud cluster of Apache Solr, an open-source search platform, and the core algorithm for geocoding which works post-retrieval in order to determine the best matches among a set of candidate results. We give a brief description of the system architecture and provide accuracy results of our geocoding engine by measuring deviations of geocoded customer addresses across India, from verified latitude, longitude coordinates of those addresses, for a sizeable address set. We also measure and report our system's ability to geocode up to different region levels, like city, locality or building. We compare our results with those of the geocoding service provided by Google against a set of addresses for which we have verified latitude-longitude coordinates and show that our geocoding engine is almost as accurate as Google's, while having a higher coverage.