Client Time Series Model: a Multi-Target Recommender System based on Temporally-Masked Encoders

Abstract

Stitch Fix, an online personal shopping and styling service, creates a personalized shopping experience to meet any purchase occasion across multiple platforms. For example, a client who wants more one-on-one support in shopping for an outfit or look can request a stylist to curate a ‘Fix’, an assortment of 5 items; or they can browse their own personalized shop and make direct purchases in our ‘Freestyle’ experience. We know that personal style changes and evolves over time, so in order to provide the client with the most personalized and dynamic experience across platforms, it is important to recommend items based on our holistic and real-time understanding of their style across all of our platforms. This work introduces the Client Time Series Model (CTSM), a scalable and efficient recommender system based on Temporally-Masked Encoders (TME) that learns one client embedding across all platforms, yet is able to provide distinctive recommendations depending on the platform. An A/B test showed that our model outperformed the baseline model by 5.8% in terms of expected revenue.