# TwHIN: Embedding the Twitter Heterogeneous Information Network for Personalized Recommendation

**Authors:** Ahmed El-Kishky, Thomas Markovich, Serim Park, Chetan Verma, Baekjin Kim, Ramy Eskander, Yury Malkov, Frank Portman, Sofía Samaniego, Ying Xiao, Aria Haghighi  
**arXiv:** [2202.05387](https://arxiv.org/abs/2202.05387)  
**Date:** 2022-02-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2202.05387v1/

## Key points

- Study focuses on Twitter as a heterogeneous information network (HIN)
- Nodes in the network represent entities like users, content, and advertisers
- Importance of leveraging interactions from multiple relation types within the HIN
- Research explores knowledge-graph embeddings for entities in TwHIN
- Effectiveness of embeddings shown in improving performance in tasks like personalized ads rankings and offensive content detection
- Practical challenges in deploying industry-scale HIN embeddings discussed
- Valuable insights provided for optimizing performance in similar domains

## Summary

The study "TwHIN: Embedding the Twitter Heterogeneous Information Network for Personalized Recommendation" by Ahmed El-Kishky et al. delves into the realm of social networks with a focus on Twitter as a heterogeneous information network (HIN). In this network, nodes represent various domain entities such as users, content, and advertisers while edges signify different types of interactions between these entities. The authors highlight the significance of leveraging interactions from multiple relation types within the HIN to extract valuable insights about social network entities. Their research revolves around exploring knowledge-graph embeddings for entities within the TwHIN and demonstrates their effectiveness in improving performance across various downstream tasks such as personalized ads rankings and offensive content detection mechanisms. Furthermore, they discuss practical challenges associated with deploying industry-scale HIN embeddings and offer valuable insights for optimizing performance in similar domains.