Research map / Label-efficient and robust learning / Recognition and long-tail
SeGA: Preference-Aware Self-Contrastive Learning with Prompts for Anomalous User Detection on Twitter
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SeGA, a preference-aware self-contrastive pre-training with LLM-generated topic/emotion prompts over user-list networks to detect bots and trolls (anomalous users) on Twitter, evaluated on Twibot-22-derived data.
Problem
Detecting anomalous Twitter users, including bots and trolls that mimic normal users, where existing bot-detection methods fail to capture subtle distinctions between users.
Approach
Self-contrastive pre-training that learns user preference embeddings from topic and emotion descriptions of tweets (generated via prompts to an LLM) over the user-list graph, before fine-tuning with other features for classification of normal users, bots, and trolls.
Key results
- SeGA achieves F1 of 64.65 on partial Twibot-22 without troll labels versus 62.57 for RGT.
- Troll-class F1 improves from 46.62 (RGT) to 53.95 (SeGA) per the rebuttal.
- Bot-class F1 improves from 32.79 (RGT) to 33.33 (SeGA) per the rebuttal.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SeGA instead of RGT | Anomalous user detection · Original partial Twibot-22 | F1: 62.57 -> 64.65 | Helped |
| Use only topics in preference pre-training (ablation) | SeGA · Twibot-22 | performance | No clear effect |
| Use only emotions in preference pre-training (ablation) | SeGA · Twibot-22 | performance | No clear effect |
Limitations
- Trolls labeled automatically via Bot Sentinel threshold 0.5, with moderate annotator agreement (Cohen's kappa about 0.7).
- Pre-training relies mainly on textual features, ignoring other metadata during pre-training.
- Reported gains in bot-class F1 are small.
Experimental setup
Datasets: Twibot-22 Bot Sentinel-labeled trolls
Benchmarks: Twibot-22
Metrics: F1 precision recall accuracy
Baselines: RGT
Builds on: RGT Twibot-22
Closest papers by meaning
- Pre-trained Behavioral Model for Malicious User Prediction on Social Platform (AAAI 2025 · similarity 0.83)
- A Multi-Task Benchmark for Abusive Language Detection in Low-Resource Settings (NeurIPS 2025 · similarity 0.82)
- Beyond Memorization: Violating Privacy via Inference with Large Language Models (ICLR 2024 · similarity 0.81)
- Distributional Preference Learning: Understanding and Accounting for Hidden Context in RLHF (ICLR 2024 · similarity 0.81)
- VLMGuard: Bootstrapping Malicious Prompt Detectors from Unlabeled Vision-Language Prompts in the Wild (NeurIPS 2026 · similarity 0.81)
- Do LLMs Recognize Your Preferences? Evaluating Personalized Preference Following in LLMs (ICLR 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.