Are these the same pair of shoes? - Matching retail products at scale
Submitted by Nikhil Ketkar (@nikhilketkar) on Monday, 15 June 2015
Section: Full Talk Technical level: Intermediate
Matching identical products from different retail websites is one of the hardest and the most impactful problems in the space of product intelligence. This talk will cover the breadth of algorithms and models we use for matching products across customer catalogs. It will also cover some practical aspects of taking these algorithms and models to production.
Product matching is the problem of resolving product entities across e-commerce sites. This involves a complex sequence of tasks which include -
1) automatic extraction of key information regions from raw HTML (for example, product titles, UPCs etc.)
2) categorising products into a unified taxonomy
3) semantic parsing of product titles and specifications
4) standardization of attributes such as brands, colours etc.,
5) grouping products into clusters of matched products based on a similarity function or inferencing model. This is a challenging problem because unique and universally agreed upon identifiers are not always available and product details are noisy and often sparse. So we have to develop contextual understanding of product specifications, which are often expressed differently by retailers, merchants, aggregators etc.
To scale the matching problem to half a billion products, we also need to prune and bucket effectively while achieving good recall. Matches need to be highly precise since customers may use them for sensitive tasks such as price comparison, competitive analysis and catalog enrichment. We employ an ensemble of online and offline algorithms and models to perform matching at scale for a large number of stores, categories and brands.
Nikhil Ketkar leads the data science team at Indix (www.indix.com) which does the R&D around product categorization, standardization, matching, search relevance and ranking. He brings along a decade of experience in making data-driven decisions and building machine learning models.