Fashion recommendations through cross-media information retrieval

Abstract

Fashion recommendation has attracted much attention given its ready applications to e-commerce. Traditional methods usually recommend clothing products to users on the basis of their textual descriptions. Product images, although covering a large resource of information, are often ignored in the recommendation processes. In this study, we propose a novel fashion product recommendation method based on both text and image mining techniques. Our model facilitates two kinds of fashion recommendation, namely, similar product and mix-and-match, by leveraging text-based product attributes and image features. To suggest similar products, we construct a new similarity measure to compare the image colour and texture descriptors. For mix-and-match recommendation, we firstly adopt convolutional neural network (CNN) to classify fine-grained clothing categories and fine-grained clothing attributes from product images. Algorithm is developed to make mix-and-match recommendations by integrating the image extracted categories and attributes information are with text-based product attributes. Our comprehensive experimental work on a real-life online dataset has demonstrated the effectiveness of the proposed method.