Abstract
Personalized outfit recommendation, which aims to recommend the outfits to a given user according to his/her preference, has gained increasing research attention due to its economic value. Nevertheless, the majority of existing methods mainly focus on improving the recommendation effectiveness, while overlooking the recommendation efficiency. Inspired by this, we devise a novel bi-directional heterogeneous graph hashing scheme, called BiHGH, towards efficient personalized outfit recommendation. In particular, this scheme consists of three key components: heterogeneous graph node initialization, bi-directional sequential graph convolution, and hash code learning. We first unify four types of entities (i.e., users, outfits, items, and attributes) and their relations via a heterogeneous four-partite graph. To perform graph learning, we then creatively devise a bi-directional graph convolution algorithm to sequentially transfer knowledge via repeating upwards and downwards convolution, whereby we divide the four-partite graph into three subgraphs and each subgraph only involves two adjacent entity types. We ultimately adopt the Bayesian Personalized Ranking loss for the user preference learning and design the dual similarity preserving regularization to prevent the information loss during hash learning. Extensive experiments on the benchmark dataset demonstrate the superiority of BiHGH.
Framework
Dataset
In fact, there are three popular datasets for personalized outfit recommendation: Polyvore-630, Polyvore-519, and IQON-550. However, the former two datasets lack the attribute information of items. Therefore, we adopt the dataset IQON550 that has rich attribute labels for evaluating our method in the context of POR. Derived from the dataset IQON3000, this dataset consists of 550 users with 57,750 positive outfits, where each user historically interacted with 120 positive outfits. For each user, the 120 positive outfits are divided into three parts: 85, 15, and 20 outfits for training, validation, and testing, respectively. It is worth noting that we further split the training set into 2 chunks: 50% for building the whole heterogeneous graph, and 50% for training the model. For each positive pair, the original IQON-550 also provides 10 negative outfits, randomly sampled from the other users’ positive outfits. In this work, we randomly select one from the 10 negative samples, and constitute a triplet. Ultimately, we have 23,650 training and 8,250 validation triplets. As for testing, we use the original testing set provided by IQON-550, where for each user, there are 20 positive outfits together with 200 negative ones for ranking. Each outfit in IQON-550 involves 3∼8 items from different categories.