A Dual Heterogeneous Graph Attention Network to Improve Long-Tail Performance for Shop Search in E-Commerce
Authors:
Open Access
- 20 August 2020
- conference paper
- Published by Association for Computing Machinery (ACM)
- p. 3405-3415
Abstract
Shop search has become an increasingly important service provided by Taobao, the China's largest e-commerce platform. By using shop search, a user can easily identify the desired shop that provides a full-scale of relevant items matching his information need. With the tremendous growth of users and shops, shop search faces several unique challenging problems: 1) many shop names do not fully express what they sell, i.e., the semantic gap between user query and shop name; 2) due to the lack of user interactions, it is difficult to deliver a good search result for the long-tail queries and retrieve long-tail shops that are highly relevant to a query. To address these two key challenges, we resort to graph neural networks (GNNs) which have various successful applications in arbitrarily structured graph data. Specifically, we propose a dual heterogeneous graph attention network (DHGAT) integrated with the two-tower architecture, using the user interaction data from both shop search and product search. At first, we build a heterogeneous graph in the context of shop search, by exploiting both the first-order and second-order proximity from user search behaviors, user click-through behaviors and user purchase records. Then, DHGAT is devised to attentively adopt heterogeneous and homogeneous neighbors of query and shop to enhance representations of themselves, which can help relieve the long-tail phenomenon. Besides, DHGAT enriches semantics of query text and shop name by compositing the titles of the relevant items to alleviate the semantic gap. Moreover, to enhance the graph representation learning, we augment DHGAT with a regularized neighbor proximity loss (NPL) to explicitly learn the graph topological structure and train whole framework in an end-to-end fashion. Compelling results from both offline evaluation and online A/B tests demonstrate the superiority of DHGAT over state-of-the-art methods, especially for long-tail queries and shops.Keywords
Funding Information
- National Natural Science Foundation of China (61872278)
This publication has 17 references indexed in Scilit:
- Graph Convolutional Neural Networks for Web-Scale Recommender Systems2018
- Learning and Transferring IDs Representation in E-commerce2018
- On Application of Learning to Rank for E-Commerce Search2017
- Learning Latent Vector Spaces for Product Search2016
- node2vec2016
- Mining Coordinated Intent Representation for Entity Search and Recommendation2015
- DeepWalk2014
- Learning deep structured semantic models for web search using clickthrough data2013
- A probabilistic mixture model for mining and analyzing product search log2013
- Multi-facet product information search and retrieval using semantically annotated product family ontologyInformation Processing & Management, 2010