Fuzzy matching on big-data: An illustration with scanner and crowd-sourced nutritional datasets

Abstract

Food retailers' scanner data provide unprecedented details on local consumption, provided that product identifiers allow a linkage with features of interest, such as nutritional information. In this paper, we enrich a large retailer dataset with nutritional information extracted from crowd-sourced and administrative nutritional datasets. To compensate for imperfect matching through the barcode, we develop a methodology to efficiently match short textual descriptions. After a preprocessing step to normalize short labels, we resort to fuzzy matching based on several tokenizers (including n-grams) by querying an ElasticSearch customized index and validate candidates echos as matches with a Levensthein edit-distance and an embedding-based similarity measure created from a siamese neural network model. The pipeline is composed of several steps successively relaxing constraints to find relevant matching candidates.