Outfit2Vec: Incorporating Clothing Hierarchical MetaData into Outfits’ Recommendation

Fashion Personalisation is emerging as a major service that online retailers and brands are competing to provide. They aim to deliver more tailored recommendations to increase revenues and satisfy customers by providing them options of similar items according to their purchase history. However, many online retailers still struggle with turning customers’ data into actionable and intelligent recommendations that reflect their personalised and preferred taste of style. On the other hand due to the ever increasing use of social media, fashion brands invest in influencers’ marketing to advertise their brands to reach a larger segment of customers who strongly trust their influencers’ choices. In this context the textual and visual analysis of social media can be used to extract semantic knowledge about customers’ preferences that can be further applied in generating tailored online shopping recommendations. As style lies in the details of outfits, recommendation models should leverage the fashion metadata ranging from clothing categories and subcategories to attributes such as materials and patterns to overall style description in order to generate fine-grained recommendations. Recently, several recommendation algorithms suggested to model the latent representations of items and users with neural word embeddings approaches which showed improved results. Inspired by Paragraph Vector neural embeddings model, we present Outfit2vec and PartialOutfit2vec in which we leverage the complex relationship between user’s fashion metadata while generating outfits’ embeddings. In this paper, we also describe a methodology to generate representative vectors of hierarchically-composed fashion outfits. We evaluate our models using different strategies in comparison to the paragraph embedding models on an extensively-annotated Instagram dataset on recommendation and multi-class style classification tasks. Our models achieve better results specially in whole outfits’ ranking evaluations with an average of 22% increase.