| This dissertation focuses on learning robust and discriminative feature representation for fashion and social media data analysis. Deep learning based methods show huge improvement compared to traditional handcraft features in many fields such as classification, generation, recommendation and retrieval. However, conventional deep learning methods may not fully address the challenges in fashion and social media data analysis. This dissertation explores four specific tasks applying regularized deep learning algorithms including auto-encoder, convolutional neural networks, and generative adversarial network. First, this dissertation addresses on the style classification problem (e.g., Baroque and Gothic architecture style), which is grabbing increasing attention in many fields such as fashion, architecture, and manga. The spread out phenomenon in style classification is pointed out. It means that visually less representative images in a style class are usually very diverse and easily getting misclassified, and they are named as weak style images. A Consensus Style Centralizing Auto-Encoder (CSCAE) for learning robust style features representation is presented, especially for weak style classification. Second, this dissertation works on the cross-domain clothing retrieval problem. A deep cross-triplet embedding convolutional neural network is presented, considering the discrepancy (e.g., background, pose, illumination) between street domain and shop domain clothing. Third, this dissertation presents an end-to-end feed-forward neural network for fashion style generation, considering both global and patch based regularizers. Given a basic clothing image and a fashion style image (e.g., leopard print), the algorithm generates a clothing image with the certain style in real time with a neural fashion style generator. Fourth, this dissertation focuses on heterogeneous recommendation on social media. Real-world recommender usually makes use of heterogeneous types of user feedbacks---for example, binary ratings such as likes and dislikes and numerical rating such as 5-star grades. This dissertation presents a Deep Low-rank Sparse Collective Factorization (DLSCF) to transfer knowledge from binary ratings to numerical ratings, facing more serious data sparsity problem. |
