Data modalities, consumer attributes and recommendation performance in the fashion industry

Abstract

This paper investigates determinants of recommendation systems’ performance in an online experiment in a large European Internet footwear store. By combining transactional data and archival customer records, a unique database was compiled from which proxy variables were extracted to represent dimensions of consumer loyalty and shopping involvement. These variables were combined in regression analysis with technical characteristics of two types of algorithms employed for generating recommendations: the EMDE algorithm, relying on the LSH method, and the industry-standard CF-RS. Statistical analysis reveals that recommendations are more successful when visual data modality is combined with behavioural data. Better recommendation performance was found to be associated with lower levels of consumer involvement in shopping, as well as higher levels of trust and engagement with the vendor. Experience with the vendor showed a negative correlation with recommendation performance through both its main effect and by its interactions with other consumer-related variables.