Research map / Label-efficient and robust learning / Self-supervised and contrastive
Pre-trained Behavioral Model for Malicious User Prediction on Social Platform
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 MaP, a three-stage contrastive and masked-reconstruction pre-training framework for BERT-based user behavior models, evaluated on malicious user detection and classification on the Weixin platform.
Problem
Malicious user prediction on social platforms needs labeled data that is scarce, and existing self-supervised behavior pre-training methods focus on single malicious types, rely on manual rules/features, and fail to capture repetitive and sporadic malicious behavior patterns.
Approach
A three-stage pre-training scheme: masked behavior reconstruction, then contrastive learning with behavior consistency augmentation (same user across periods as positives) and local disruption augmentation (shuffled local segments as positives), then a triplet-loss stage over pseudo-malicious users sampled by report/block signals.
Key results
- On malicious user detection, MaP reaches AUC 0.8600 and KS 0.7201, versus 0.8398 AUC for UserBERT and 0.8392 for TS2Vec.
- Macro F1 for malicious user classification is 0.7056 for MaP versus 0.6999 for TS2Vec (best of baselines).
- In an online A/B test on Weixin over one billion users, MaP improved AUC by 4.4% and KS by 11.1% over the deployed UserBERT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MaP pre-training rather than UserBERT | BERT · Malicious user detection | AUC: 0.8398 -> 0.8600 | Helped |
| Use MaP pre-training rather than UserBERT | BERT · Malicious user detection | KS: 0.6797 -> 0.7201 | Helped |
| Use MaP rather than TS2Vec | BERT4Rec · Malicious user classification | Macro Average Precision: 0.7653 -> 0.7608 | Hurt |
| Use MaP rather than TS2Vec | BERT4Rec · Malicious user classification | Macro Average Recall: 0.6583 -> 0.6678 | Helped |
| Use MaP rather than TS2Vec | BERT4Rec · Malicious user classification | Macro Average F1-Score: 0.6999 -> 0.7056 | Helped |
| Do not apply PUS during MBR pre-training (ablation) | MaP · Malicious user detection | AUC: 0.8384 -> 0.8600 | Helped |
| Add FL4PMB (ablation) | MaP · Malicious user detection | AUC: 0.8427 -> 0.8600 | Helped |
| Add FL4MBP (ablation) | MaP · Malicious user detection | AUC: 0.8377 -> 0.8600 | Helped |
| Add PUS (ablation) | MaP · Malicious user detection | AUC: 0.8449 -> 0.8600 | Helped |
| Add BCA (ablation) | MaP · Malicious user detection | AUC: 0.8216 -> 0.8600 | Helped |
| Add LDA (ablation) | MaP · Malicious user detection | AUC: 0.8269 -> 0.8600 | Helped |
Surprising findings
- Applying the pseudo-malicious user sampling strategy during the masked reconstruction stage degraded performance, so the paper skips it in that stage.
Limitations
- Relies solely on user behavior sequences without social network or content signals, as the authors acknowledge.
- Evaluated only on proprietary Weixin data, so generality to other platforms is unverified.
- Pseudo-malicious labels derived from reports and blocks are noisy proxies for true malicious users.
Experimental setup
Datasets: Weixin user behavior sequences (proprietary) Weixin malicious user detection dataset Weixin malicious user classification dataset
Benchmarks: Weixin malicious user detection dataset Weixin malicious user classification dataset Weixin online A/B test
Metrics: AUC KS macro average precision macro average recall macro average F1-score
Baselines: BERT w/o pre-training BERT4Rec UserBERT MSDP SubSeq TS2Vec
Builds on: BERT BERT4Rec SiT Sentence-BERT ConSERT
Closest papers by meaning
- PANTHER: Generative Pretraining Beyond Language for Sequential User Behavior Modeling (NeurIPS 2025 · similarity 0.86)
- SeGA: Preference-Aware Self-Contrastive Learning with Prompts for Anomalous User Detection on Twitter (AAAI 2024 · similarity 0.83)
- BehaviorBench: Modeling Real-World User Decisions from Behavioral Traces (NeurIPS 2026 · similarity 0.82)
- Federated Behavioural Planes: Explaining the Evolution of Client Behaviour in Federated Learning (NeurIPS 2024 · similarity 0.81)
- Defense Against Model Stealing Based on Account-Aware Distribution Discrepancy (AAAI 2025 · similarity 0.81)
- T-NET: Weakly Supervised Graph Learning for Combatting Human Trafficking (AAAI 2024 · 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.