# BotTriNet: A Unified and Efficient Embedding for Social Bots Detection via Metric Learning

**Authors:** Jun Wu, Xuesong Ye, Man Yan Yuet  
**arXiv:** [2304.03144](https://arxiv.org/abs/2304.03144)  
**Date:** 2023-04-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.03144v1/

## Key points

- Bot accounts in online social networks are a persistent concern for genuine users who are often subjected to harassment and invasion by these malicious entities.
- The proposed solution is a unified content embedding framework called BOTTRINET that utilizes textual content posted by accounts for bot detection.
- The approach assumes that contexts naturally reveal account personalities and habits, making the extraction of bot-related information using embedding techniques efficient and valuable.
- Compared to the side information approach commonly used in bot detection, our content approach model can be deployed in each content-related user action such as posting, commenting, or sending messages. This allows for faster detection without waiting an extended period to collect data.
- The method makes several contributions to the field of bot detection including applying Triplet Learning method from computer vision to the bot detection field and designing a simple and easy-to-use embedding framework called BOTTRINET that represents raw embeddings into more efficient account embeddings.
- The system was evaluated on a real-world dataset CRESCI2017 comprising three bot account categories and five bot sample sets. It achieved the highest average accuracy of 98.34% and f1score of 97.99% on two content-intensive bot sets outperforming previous work and becoming state-of-the-art.
- In conclusion, the proposed BOTTRINET framework provides a reliable solution for detecting bots quickly and accurately in online social networks using textual content posted by accounts.

## Summary

The detection of bot accounts in online social networks is a persistent concern for genuine users who are often subjected to harassment and invasion by these malicious entities. To address this issue, we propose a unified content embedding framework called BOTTRINET that utilizes textual content posted by accounts for bot detection. Our approach assumes that contexts naturally reveal account personalities and habits, making the extraction of bot-related information using embedding techniques efficient and valuable. Compared to the side information approach commonly used in bot detection, our content approach model can be deployed in each content-related user action such as posting, commenting, or sending messages. This allows for faster detection without waiting an extended period to collect data. We also introduce the first practice of applying metric learning techniques in the bot detection field, which has never been considered before. Our method makes several contributions to the field of bot detection. Firstly, we apply Triplet Learning method from computer vision to the bot detection field by abstracting an account's entire lifetime as an Anchor and capturing its intrinsic nature as an independent entity of behavior and content. This breakthrough significantly improves bot detection performance. Secondly, we design a simple and easy-to-use embedding framework called BOTTRINET that represents raw embeddings into more efficient account embeddings, greatly enhancing the separability between bot and genuine accounts. We evaluate our system on a real-world dataset CRESCI2017 comprising three bot account categories and five bot sample sets. Our method achieves the highest average accuracy of 98.34% and f1score of 97.99% on two content-intensive bot sets outperforming previous work and becoming state-of-the-art. It also makes a breakthrough on four content-less bot sets with an average accuracy improvement of 11.52% and an average f1score increase of 16.70%. In conclusion, our proposed BOTTRINET framework provides a reliable solution for detecting bots quickly and accurately in online social networks using textual content posted by accounts.