# Heterogeneous Graph Neural Networks for Large-Scale Bid Keyword Matching

**Authors:** Zongtao Liu, Bin Ma, Quan Liu, Jian Xu, Bo Zheng  
**arXiv:** [2111.00926](https://arxiv.org/abs/2111.00926)  
**Date:** 2021-11-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2111.00926v1/

## Key points

- E-commerce relies heavily on digital advertising
- Taobao and Amazon use digital advertising as a critical component
- Advertiser-side services are becoming increasingly important
- Sponsored search is a fundamental service that provides marketing tools to advertisers
- Existing methods for keyword recommendation neglect rich heterogeneous interactions
- Researchers face challenges in learning enriched and robust embeddings from complex interactions among various types of objects, and conducting high-quality matching for new ads that usually lack sufficient data.
- HetMatch is a heterogeneous-graph-neural-network-based model for keyword matching developed by the authors to address these challenges.
- HetMatch extracts enriched and robust embeddings among rich relations by designing a hierarchical structure that fuses and enhances relevant neighborhood patterns both on the micro-and macro-levels.
- HetMatch involves more positive samples for cold-start ads through its multi-view framework.
- Experimental results on a large-scale industrial dataset as well as online AB tests exhibit the effectiveness of HetMatch.

## Summary

The world of e-commerce relies heavily on digital advertising, with platforms like Taobao and Amazon using it as a critical component. While consumer-focused issues such as click-through rate (CTR) and conversion rate (CVR) prediction have garnered attention in recent years, the advertiser side is becoming increasingly important. Sponsored search is a fundamental service that provides marketing tools to advertisers. The primary step in keyword recommendation for sponsored search is keyword matching which has been addressed by existing methods that model relevance based on a single type of relation among ads and keywords, such as query clicks or text similarity. However, these approaches neglect the rich heterogeneous interactions hidden behind them. To address this gap in knowledge, researchers face several challenges when it comes to keyword matching: how to learn enriched and robust embeddings from complex interactions among various types of objects; and how to conduct high-quality matching for new ads that usually lack sufficient data. To overcome these challenges, the authors developed HetMatch - a heterogeneous-graph-neural-network-based model for keyword matching - which has been deployed both online and offline at the core sponsored search platform of Alibaba Group. HetMatch extracts enriched and robust embeddings among rich relations by designing a hierarchical structure that fuses and enhances relevant neighborhood patterns both on the micro- and macro-levels. Additionally, by proposing a multi-view framework, HetMatch can involve more positive samples for cold-start ads. Experimental results on a large-scale industrial dataset as well as online AB tests exhibit the effectiveness of HetMatch. Overall, this paper highlights the importance of advertiser-side services in digital advertising platforms like Taobao and Amazon. It also sheds light on the challenges associated with keyword matching while introducing HetMatch - an effective solution that leverages heterogeneous graph neural networks to extract enriched embeddings from complex interactions among various types of objects.