Trust & Safety

Tree-based anomaly isolation achieving O(n log n) complexity; the industry standard for fraud detection.

Introduced Local Outlier Factor assigning continuous 'degree of outlierness' for variable-density anomaly detection.

Comprehensive taxonomy covering classification, nearest-neighbor, clustering, and statistical approaches; 8,000+ citations.

Extended journal version with theoretical analysis; handles high-dimensional masking and swamping effects.

One-class SVM for novelty detection; foundational method for fraud detection when only normal data is available.

Comprehensive survey of deep learning anomaly detection; covers autoencoders, GANs, and self-supervised approaches.

Survey comparing neural networks, genetic algorithms, and expert systems for credit card fraud detection.

Addresses realistic fraud challenges: extreme class imbalance, concept drift, and delayed feedback loops.

Comprehensive review analyzing ML techniques for health insurance fraud over two decades.

Network-based approach using statistically validated networks to detect coordinated fraud rings.

Field experiment showing supervised and unsupervised methods are complements, not substitutes.

Detects anomalies in weighted graphs using egonet features; foundational for fraud ring detection.

Detects dense subgraphs even when fraudsters add random edges to camouflage; handles lockstep behavior.

GEM model using heterogeneous graphs (users, devices, transactions) for Alipay fraud detection.

Dynamic network embeddings for streaming anomaly detection; updates in O(1) per edge.

Relational GCN exploiting follower/friend graphs achieves SOTA on bot detection benchmarks.

Foundational Botometer paper; random forest on 1,000+ features estimating 9-15% of Twitter accounts are bots.

Seminal paper defining social bots, detection challenges, and policy implications for platform manipulation.

Practitioner guide for Botometer v4 with CAP scores, threshold selection, and case study methodology.

Addresses cross-dataset generalization using ensemble of specialized classifiers.

First large-scale study of fake reviews; 16% of Yelp reviews flagged as fake, increasing with competition.

Compares TripAdvisor vs Expedia reviews; finds review manipulation concentrated among independent hotels.

Comprehensive survey covering linguistic, behavioral, and graph-based fake review detection methods.

Foundational 3-class dataset (hate/offensive/neither) with 24K labeled tweets; standard benchmark.

Technical architecture behind Google Jigsaw's Perspective API; handles obfuscation, code-switching, multilingual toxicity.

Develops methods for measuring unintended identity-term bias in toxicity classifiers; foundational for fair ML.

First benchmark with rationale annotations for explainable hate speech detection across 20K posts.

Multi-label cyberbullying detection with fine-grained categories; benchmark on Dutch social media.

Studies how subreddit norms shape behavior; users posting in banned communities post less toxic content after ban.

Google study on account hijacking vectors; SMS verification blocks 100% of automated bots and 96% of bulk phishing.

Analyzes RBA deployments at Google, Microsoft, Amazon; develops measurement framework.

Graph neural networks modeling account-device-transaction relationships for ATO detection.

Deep learning approach combining behavioral biometrics with session features for continuous authentication.

Maps the underground economy of stolen accounts; traces supply chain from compromise to monetization.

Characterizes Russian IRA troll activity across platforms; develops detection methodology.

Framework for understanding information operations as collaborative work; case study of 2016 election.

Analysis of 14M tweets by IRA; identifies patterns distinguishing troll engagement from organic users.

Comprehensive taxonomy of manipulation techniques across platforms and actor types.

Yahoo system combining n-grams, syntactic, and semantic features; production-scale abuse detection.

CNN and RNN architectures for harassment detection; analyzes temporal patterns in abuse.

Wikipedia personal attack corpus with 100K+ labeled comments; crowdsourcing methodology for abuse annotation.

RNN with attention for comment moderation; deployed at Greek news organization.

Experiments showing people value privacy but underestimate risks; foundational behavioral privacy study.

JEL survey covering market structure, price discrimination, and welfare effects of privacy regulation.

Foundational paper defining differential privacy; mathematical framework for privacy-preserving computation.

Studies tradeoff between ad intrusiveness and platform revenue; privacy implications of targeting.