Using supervised learning to classify clothing brand styles

Abstract

Machine learning techniques have the potential to alter the highly competitive online fashion retail industry by improving customer service through personalized recommendations. A fashion style classification system can improve the customer search functionality and provide a more personalized experience for the user. Supervised learning techniques with fashion based applications face the problem of developing quantitative measures for describing fashion products which are subjective in nature. To address this issue the authors asked fashion experts to assist in the assembly of a training set of brand-style associations. Quantitative measures were attributed to each brand in the training set by applying natural language processing, text mining, and eBay query results. This data set was used to train a support vector machine which classified the approximately 8000 remaining brands into style categories. The prospective classifier model was assessed based on its positive predictive values which yielded a 56.25% success rate. Given that there are eight different styles to choose from, a baseline for the percentage is only 12.5%. The SVM thus adds significant value to the classification of fashion brands. The final style categorization was integrated as a new filter feature that allows the user to narrow down their searches and access relevant results.