When you need to arrange your products by category, we refer to the Attribute Extraction; what style and category best match the product. Often times, textual data can be noisy and the taxonomy for products will change. In practical terms, a rough grouping of images would be created for an editor or stylist to view. By reducing the poor quality images, the editor can be presented with more closely identifiable and manageable sets of images and by merely checking yes/no and using the fine-grained skills of a stylist, a unique style library can be created.

Technically, a model would be trained on two labels of every image: category and style. A pre-trained model can be used to create the embeddings and from that, you can build two classifiers. In the final step, you can combine the objective function into one. It can be similar to a K-NN model using visual embedding and voting model to improve accuracy.

From the table above, it is clear that the best style accuracy comes from the model that learns both category and style and so, learning to differentiate categories helps to learn style.