Personalized outfit generation
Predicting user preference on a set of well-matched fashion items should address both item compatibility and personalisation incorporating outfit-item and user-outfit relationships. [Chen 19] connects user preferences regarding individual items/outfits with Transformer (encoder-decoder) architecture. Outfit item is generated step by step according to user preference signal from Per network and compatibility signal from Gen network
Through embedding propagation on a hierarchical graph with user-outfit-item relationships, item information is aggregated into an outfit representation which is then refined into a user’s representation via his/her historical outfits [Li 20]. Inter-relations between outfits (e.g., fashion items with same colors (black suit, white shirt) or outfits sharing similar styles (e.g., casual, high heels)) are ignored in learning the user-outfit preference predictor however.There is thus a need to go beyond user's preference over a single item - with little or no interaction.
New users usually only have few (less than 5) outfits available for learning. With such a limited number of training examples, it is challenging to model their preferences reliably. [Lu 21] uses a stacked self-attention mechanism to model high-order interactions among items. Items in an outfit are then embedded into a single compact representation within the outfit space. To accommodate a variety of users’ preferences, each user is characterized with a group of learnable latent vectors in outfit space that are representatives of their likes. A set of general anchors is also learned to model preference shared by all users.?
Capsule wardrobe generation is a complex combinatorial problem that requires understanding of how multiple visual items interact. The generative process often needs fashion experts to manually tease the combinations out, making it hard to scale. TensorNet [Chen 21] consists of two core modules: a Cross-Attention Message Passing module and a Wide&Deep Tensor Interaction module so as to characterize local region-based patterns as well as global compatibility of entire outfits.
Recommending pairs of items that the customer would like to wear together is still less studied as it involves learning a compatibility metric personalized to each customer. A new framework (PSA-Net [Taraviya 21]) is to learn compatibility that is personalized to the customer - a customer dependent subspace learning framework where attention weights of subspaces are learnt using customer representations.
领英推荐
Existing methods mainly concentrate thus on user and item entities, as well as their interactions, but ignore attribute entities, which contain rich semantics. [Guan 22] builds a heterogeneous graph to unify the three types of entities and their relations (i.e., user-item interactions, item-item matching relations, and item-attribute association relations). A multi-modal content-oriented user embedding module is designed to learn user representations by inheriting contents of interacted items. Meanwhile, user- and item-oriented metapaths are defined and graph learning enhances user/item embeddings -? with contrastive regularization for performance improvement finally.
[Taraviya 21] PSANet-subspace attention for personalized compatibility