Graph Data Science DEMO for fraud analysis
1 like633 views
This document provides an agenda and overview for a graph data science demo focusing on fraud analysis. The demo will review Neo4j's graph data science library and algorithms for pathfinding, centrality, community detection, and similarity. It will use sample bank transaction and customer data modeled as a graph to demonstrate PageRank, betweenness centrality, weakly connected components, Louvain modularity, and node similarity algorithms. The goal is to identify important nodes, communities, and similar entities to detect potential fraud.
Graph Data Science DEMO for fraud analysis
- 1. Graph Data Science Demo for Fraud Analysis Joe Depeau Sr. Presales Consultant, UK 28th April, 2020 @joedepeau http://linkedin.com/in/joedepeau
- 2. • Review of the Neo4j Graph Data Science Library • Demo Data Overview • Review of Graph Algorithms for Demo • Demo • Q&A 2 Agenda
- 3. Neo4j Graph Data Science Library 3
- 4. 4 Neo4j Graph Algorithms Overview Pathfinding & Search Centrality / Importance Community Detection Link Prediction Finds optimal paths or evaluates route availability and quality. Determines the importance of dis9nct nodes in the network. Detects group clustering or partition options. Evaluates how alike nodes are by neighbors and relationships. Estimates the likelihood of nodes forming a future relationship. Similarity
- 5. Graph and ML algorithms in Neo4j • Minimum Weight Spanning Tree • Shortest Path • Single Source Shortest Path • All Pairs Shortest Path • A* • Yen’s K-shortest Paths • Random Walk • Breadth First Search • Depth First Search • Degree Centrality • Closeness Centrality • Betweenness Centrality • PageRank • ArticleRank • Eigenvector Centrality • Triangle Count / Clustering Coefficient • Weakly Connected Components • Strongly Connected Components • Label PropagaJon • Louvain Modularity • K-1 Colouring • Modularity OpJmisaJon • Node Similarity • Approximate Nearest Neighbours • Cosine Similarity • Euclidean Similarity • Jaccard Similarity • Overlap Similarity • Pearson Similarity Pathfinding & Search Centrality / Importance Community Detec<on Similarity https://neo4j.com/docs/graph-data-science/1.0/ Link Prediction • Adamic Adar • Common Neighbours • PreferenJal ARachment • Resource AllocaJons • Same Community • Total Neighbours 5
- 7. 7 Some Examples of Typical Bank Data Event DataProduct and Services Data Customer DataOrganisational Data 3rd Party Data Documentation Employee Data Processes Systems and Databases KPIs and Reports Address Personal Data Documents Relationships Assets Documentation Processes Product / Service Details Product / Service Hierarchy Pricing Money Movements Web / App Activity Customer Contact Social Media Credit Rating Agencies Market Data Organisational Hierarchy Corporate Data
- 8. 8 Some Examples of Typical Bank Data Event DataProduct and Services Data Customer DataOrganisational Data 3rd Party Data Documentation Employee Data Processes Systems and Databases KPIs and Reports Address Personal Data Documents Relationships Assets Documentation Processes Product / Service Details Product / Service Hierarchy Pricing Money Movements Web / App Activity Customer Contact Social Media Credit Rating Agencies Market Data Organisational Hierarchy Corporate Data
- 9. 9 Our Graph Model Total Nodes : 3.4m Total Relationships : 10.2m
- 10. 10 Three ways a Client node can be Flagged Performed a transaction flagged as fraud Share a SSN with another Client Have more than one SSN on file
- 12. PageRank What: Finds important nodes based on their relationships. Why: Identify important or influential Client nodes by quantifying the flows of money towards them. Uses: - Fraud detection - Anti-money Laundering - Inform prioritization during analysis and investigation12
- 13. 13 The PageRank Algorithm PageRank: what nodes can be considered ‘important’ in our graph based on money flows ?
- 14. 14 The PageRank Algorithm PageRank: what nodes can be considered ‘important’ in our graph based on money flows ? Inputs .pagerank Property Output
- 15. Betweenness Centrality 15 What: The sum of the % shortest paths that pass through a node, calculated by pairs. Why: Identify well connected nodes within the graph, or bridges between communities. Uses: - Fraud detection - Anti-money Laundering - Inform prioritization during analysis and investigation
- 16. 16 The Betweenness Centrality Algorithm Betweenness Centrality: what nodes can be considered ‘important’ in our graph based on how many ‘shortest paths’ they are present in?
- 17. 17 The Betweenness Centrality Algorithm Betweenness Centrality: what nodes can be considered ‘important’ in our graph based on how many ‘shortest paths’ they are present in? Inputs .centrality Property Output
- 18. Weakly Connected Components What: Finds disconnected community subgraphs in our data. Why: Identify communities based on connections with shared pieces of identity. Uses: - Householding - Synthetic identities - Stolen identities 18
- 19. 19 The Weakly Connected Components Algorithm Weakly Connected Components: what communities exist in the data based on connections to pieces of identity ?
- 20. 20 The Weakly Connected Components Algorithm Weakly Connected Components: what communiFes exist in the data based on connecFons to pieces of idenFty ? .component_id Property Output Inputs
- 21. Louvain Modularity What: Finds communi?es in our graph who are connected. Can return intermediate results. Why: Useful for iden?fying communi?es based on transac?on behaviour rather than iden?ty. Uses: - Fraud ring detecHon - AnH-money Laundering 21
- 22. 22 The Louvain Algorithm Louvain: what communiFes of nodes transact amongst themselves ?
- 23. 23 The Louvain Algorithm Louvain: what communities of nodes transact amongst themselves ? Inputs .louvain_community Property Output
- 24. Node Similarity What: Similarity between nodes based on neighbours. Writes a new relationship to the graph. Why: Identify similar nodes who share common pieces of identity. Uses: - Entity Resolution - Synthetic identities - Stolen identities 24
- 25. 25 The Node Similarity Algorithm Node Similarity : how similar are two Client nodes based on the non- Merchant accounts they transact with ?
- 26. 26 The Node Similarity Algorithm Node Similarity : how similar are two Client nodes based on the non- Merchant accounts they transact with ? SIMILAR Relationship Output with .score property Inputs
- 27. Demo 27
- 28. 28 Thank you!