The autoencoder is a fundamental deep learning concept used for compression of typically unquantifiable data – music, images, sentences, the list goes on. By training an autoencoder, the model can start to pick up features of the data as vectors that can be manipulated just like normal matrices. In this project, a special type of autoencoder called a disentangled variational autoencoder, or β-VAE, is used on images of shoes and tops. These specific autoencoders are called disentangled and variational because the autoencoder can slowly pick up on features and directly manipulate them since the data is transformed into data both understandable by computers and humans. For example, the autoencoder can change something as specific as the visibility of a logo on a shoe or change the colorway.
By manipulating these values within a given vector, the machine can then correlate these vectors to prices, sales, or other metrics, leading to a model through which a company can tweak certain elements of a product and immediately quantify customer satisfaction. This model shown is with shoes, but works with all kinds of fashion – tops, bottoms, handbags, and watches. It has yet to be tested in larger markets, but this has the potential to advise companies on how customers will react to new products before A/B-testing even occurs so companies can make better decisions on their strategy.