Apprentissage automatique sur graphes pour la fraude et les réseaux

Ce qu'elle couvre

Ce programme de niveau praticien enseigne aux analystes fraude et aux ingénieurs ML à modéliser des données relationnelles sous forme de graphes, à extraire des features pertinentes et à entraîner des architectures de Graph Neural Networks pour la détection de fraude et de risques. Les participants travaillent en pratique avec Neo4j pour le stockage et l'interrogation des graphes, PyTorch Geometric pour la construction de modèles, et apprennent les patterns de déploiement pour une inférence à faible latence en production. Les sessions combinent des bases théoriques en théorie des graphes avec des ateliers de codage en direct sur des jeux de données de fraude réels. À l'issue du programme, les participants sont capables de concevoir une solution graph ML de bout en bout — de la résolution d'entités jusqu'au monitoring des modèles.

À l'issue, vous saurez

• Construct a fraud knowledge graph in Neo4j from raw transactional data using entity resolution techniques
• Engineer graph-level, node-level, and edge-level features that outperform tabular baselines on fraud detection benchmarks
• Implement and tune a GraphSAGE or GAT model in PyTorch Geometric for node classification on imbalanced fraud datasets
• Deploy a trained GNN model behind a REST API with sub-100ms inference latency
• Design a monitoring pipeline that detects graph distribution shift and model performance degradation in production

Sujets abordés

• Graph theory fundamentals: nodes, edges, properties, and heterogeneous graphs
• Entity resolution and record linkage for building fraud graphs
• Graph feature engineering: degree centrality, PageRank, community detection, motifs
• Graph Neural Network architectures: GCN, GraphSAGE, GAT for node and edge classification
• Neo4j data modelling and Cypher queries for fraud graph construction
• PyTorch Geometric: building, training, and evaluating GNN models
• Imbalanced learning strategies for rare fraud events on graphs
• Production deployment: serving GNN models, latency optimisation, and drift monitoring

Modalité

Delivered as a blended programme over 3–5 days (on-site or virtual instructor-led), with approximately 60% hands-on lab time. Participants receive a pre-configured cloud environment with Neo4j, PyTorch Geometric, and sample fraud datasets. Each module pairs a 30-minute concept session with a 90-minute coding lab. A capstone project — building a full pipeline on a synthetic payments fraud graph — anchors the final day. Remote delivery uses breakout rooms for pair-programming; in-person delivery is preferred for groups larger than 10.

Ce qui fait que ça marche

• Start with a clear entity resolution strategy before any ML — garbage graphs produce garbage models
• Validate graph features against business-defined fraud rings before investing in GNN complexity
• Involve production engineers from day one to align graph schema and model serving architecture early
• Establish graph drift monitoring alongside standard model drift monitoring to catch structural data changes

Erreurs fréquentes

• Treating fraud detection as purely a tabular problem and ignoring relational signals between accounts, devices, and merchants
• Building graph features in offline batch mode only, making real-time scoring impossible without re-architecting
• Using homogeneous GNN architectures on heterogeneous fraud graphs, losing important semantic edge-type information
• Neglecting class imbalance strategies specific to graph data, leading to models that learn majority-class structure instead of fraud patterns

Fournisseurs à considérer

• Neo4j GraphAcademygraphacademy.neo4j.com →
• DataScientestdatascientest.com →
• DeepLearning.AIwww.deeplearning.ai →
• Jedha Bootcampwww.jedha.co →

Sources

• PyTorch Geometric Documentation — Graph Neural Networks for Fraud Detection →
• Neo4j Graph Data Science Library Documentation →