Clothing we wear reveals our personal style - wealth, occupation, religion, location and socio-identity. Shopper’s aesthetic preferences thus, and not simply suggesting popular items from a category (e.g., shirts), influences purchasing decision in a lifestyle marketplace. Given the image of a fashion item (e.g., blouse), recommending complementary matches(e.g., skirt) is a challenge. Users’ taste evolves over time and depends on persona (e.g., dressing for a party). Humans relate objects (e.g., substitutes) based on their appearance; nonvisual factors of lifestyle merchandise (e.g., durability) further complicates recommendation task. Composing outfits in addition necessitates constituent items (e.g., pants, shoes) to be compatible - similar in some but different in other aspects.
Tutorial addresses various techniques for fashion recommendation which in turn enhance conventional data mining approaches like collaborative filtering and matrix factorization. For a few such models and methods (e.g., StickBreaking Process, Mixture of Experts), we will outline results using real-world clothing data from various, online shopping platforms (e.g., Alibaba, JD.com, Polyvore, fashionbeans). Recent advances in deep learning applications (e.g., encoderdecoder, Recurrent Neural Network (RNN), Generative Adversarial Networks (GAN)) are presented for compatibility modelling, learning-to-rank and explainable recommendation.
Director Engineering Programs : Data Science - ML, Flipkart
Dept. of CSE, Indian Institute of Technology, Kharagpur