Knowledge-aware Complementary Product Representation Learning

Abstract

Learning product representations that reflect complementary relationship plays a central role in e-commerce recommender system. In the absence of the product relationships graph, which existing methods rely on, there is a need to detect the complementary relationships directly from noisy and sparse customer purchase activities. Furthermore, unlike simple relationships such as similarity, complementariness is asymmetric and non-transitive. Standard usage of representation learning emphasizes on only one set of embedding, which is problematic for modelling such properties of complementariness.We propose using knowledge-aware learning with dual product embedding to solve the above challenges. We encode contextual knowledge into product representation by multi-task learning, to alleviate the sparsity issue. By explicitly modelling with user bias terms, we separate the noise of customer-specific preferences from the complementariness. Furthermore, we adopt the dual embedding framework to capture the intrinsic properties of complementariness and provide geometric interpretation motivated by the classic separating hyperplane theory. Finally, we propose a Bayesian network structure that unifies all the components, which also concludes several popular models as special cases.The proposed method compares favourably to state-of-art methods, in downstream classification and recommendation tasks. We also develop an implementation that scales efficiently to a dataset with millions of items and customers.

Theoretical Understandings of Product Embedding for E-commerce Machine Learning

“…the outcome of random walks (MetaPath2vec [9], ProdNode2vec [28]), or by the follow-up sampling steps, e.g. sampling the products from a given context window (CompProd2vec [29]). Therefore, the realization of neighborhood varies for different input data structures and problem settings (Figure 1), and the resulting co-occurrence statistics are random variables as well.…”

“…Several recent papers have challenged the state-of-the-art deep learning recommendation algorithms against the vanilla collaborative filtering, ending up finding worse performances from deep learning on various benchmark datasets [8,21]. All the above concerns motivate our exploration of the theoretical perspective of product embeddings -the cornerstone for a considerable amount of machine learning models in e-commerce [4,12,25,[28][29][30].…”

“…The incompatibility issue is common for industrial problems, since the product features are often mixtures of quantitative and categorical variables that expand a huge irregular space that few modern machine learning models are suitable. Different approaches have been proposed to train product embeddings using SGNS with problem-specific modification, including Item2vec [4], Prod2vec [12,25], Triple2vec [27], MetaPath2vec [9], CompProd2vec (complementary product embedding) [29] and product knowledge graph [30]. Despite the different formulations of the input data structure and regularization, the core component remains to be the SGNS induced by the input co-occurrence statistic (see Section 2 for details).…”

Variational Inference for Category Recommendation in E-Commerce platforms

“…Works that exploit the graphical structure of the user-item interaction data in e-Commerce settings involve works like [44], which introduces a new representation learning approach to leverage complementary items, user-item compatibility, and user loyalty. Other works in this domain include those that use knowledge-aware learning with dual product embedding to detect complementary product relationships from noisy and sparse user purchase activities [47], etc. The Metapath2Vec algorithm used in our work was introduced by Dong et al in [13], where they formalize metapath based random walks to construct the heterogeneous neighborhood of a node and then leverage a heterogeneous skip-gram model to perform node embeddings.…”

GAN-based Recommendation with Positive-Unlabeled Sampling

“…Recommender systems have been prevalent in recent decades across multiple domains in e-Commerce [38], content streaming (YouTube) [7], and business service industries (Yelp) [34], due to their success in filtering or retrieving relevant information from user profiles and behaviors. Traditional collaborative filtering methods [10,24] and matrix factorization methods [22,27,32] are the most popular and effective set of methods of recommender systems for many years.…”