With the rapid growth of fashion-focused social networks and online shopping,intelligent fashion recommendation is now in great need. We design algorithmswhich automatically suggest users outfits (e.g. a shirt, together with a skirtand a pair of high-heel shoes), that fit their personal fashion preferences.Recommending sets, each of which is composed of multiple interacted items, isrelatively new to recommender systems, which usually recommend individual itemsto users. We explore the use of deep networks for this challenging task. Oursystem, dubbed FashionNet, consists of two components, a feature network forfeature extraction and a matching network for compatibility computation. Theformer is achieved through a deep convolutional network. And for the latter, weadopt a multi-layer fully-connected network structure. We design and comparethree alternative architectures for FashionNet. To achieve personalizedrecommendation, we develop a two-stage training strategy, which uses thefine-tuning technique to transfer a general compatibility model to a model thatembeds personal preference. Experiments on a large scale data set collectedfrom a popular fashion-focused social network validate the effectiveness of theproposed networks.