Data-driven clustering for new garment forecasting

Abstract

The ability to detect patterns early in the design process is critical for fashion firms to make decisions, particularly given the speed at which new garments are introduced. Traditionally, most garment defining features were only used by designers and buyers since the data was intractable for a computer: shape, color, fit, etc. By using natural language processing (NLP) techniques that preserve semantics, in combination with traditional data-mining, we unlock the potential to use these garment characteristic and embed them in a numerical space that's tractable. By using this novel approach to fashion data, this thesis develops two custom algorithms to forecasting the size-curve distribution of a new garment. This task is achieved by automatically finding a set of comparables of previous garments and leveraging the know results to make predictions. We develop and implement two main algorithms: \textit{Cluster-While Regress} (CWR) and \textit{k-Nearest Neighbours} (kNN) and show that with enough data the algorithms should achieve human-level accuracy and automate the comparables-finding process.