Fraud Detection in Financial Services using Graph Analysis and Machine Learning

Fraud Detection in Financial Services using
Graph Analysis and Machine Learning
Swiss Conference on Data Science 2019
Hans Viehmann
Product Manager EMEA
ORACLE Corporation
June 14, 2019
@SpatialHannes

Safe Harbor Statement
The following is intended to outline our general product direction. It is intended for
information purposes only, and may not be incorporated into any contract. It is not a
commitment to deliver any material, code, or functionality, and should not be relied upon
in making purchasing decisions. The development, release, and timing of any features or
functionality described for Oracle’s products remains at the sole discretion of Oracle.

Copyright © 2019 Oracle and/or its affiliates. All rights reserved. | 4
https://twitter.jeffprod.com
Following, no follow back
Follower, no follow back
Follow each other

Graph – an important growth area for data & analytics
Confidential – Oracle Internal/Restricted/Highly Restricted 5
Gartner identifies Top 10 Data and Analytics
Technology Trends for 2019 / No. 5
• Graph processing and Graph DBMS grow 100% annually through 2022.
• Adoption of graphs is limited by lack of skills in the market.
• Graphs are more practical to answer complex questions on
complex data – SQL is not always efficient.

Graph Analysis for Business Insight
9
Identify
Influencers
Discover Graph Patterns
in Big Data
Generate
Recommendations

• Customer profitability analysis
– Part of larger Hadoop/Big Data project
• Analysis of banking transactions
– Focus on corporate customers
• Identification of undesired behavioural
patterns, eg.
– Customers using other banks to make large
numbers of transactions
– Many of which flow back to Banco Galicia
• Increase fees, terminate contracts, or
move activities to Banco Galicia
10
Banco de Galicia

• Providing online payment solutions
– Real-time payments, e-Wallets
– 1bn revenue/yr
– 500000 payments/day
• Strong demand for fraud detection
– Only feasible with graph data
– In real-time, upon money movement
– During account creation
– In investigation, visualizing payment flows
• Analysis of payment flows
– Identifying suspicious patterns
11
Paysafe

Followed by interactive analysis with visualization and PGQL
Using graph algorithms for initial assessment

Moving towards graph analysis with machine learning
Rule Engine:
Takes decision
to process or
fail payment
Graph Query
Example: Is there fraudster in 3
payments distance?
Graph Query
Example: Do we have
linked by password
customer in 3 payments
distance?
Example: Pass fraud probability as
fact to the rule engine
Graph
Database
Machine
Learning

• Example: Finding anomalies in
healthcare billing data
– Medical providers and their operations
• Providers of the same specialty are
close to each other in the graph
– Closely connected by common services
• a provider vertex exceptionally close
to vertices of a different specialty
should be an anomaly
• Using closeness as a metric
– Eg. Hop-distance, ...
Anomaly Detection (without Machine Learning)
X
Doctors
900,000 HCPCS
6,000Edges
9,000,000

Using Personalized Pagerank to find outliers and anomalies
Variant of Page-rank algorithm that requires a set
of starting vertices*
– Random walks (with restart) from the starting vertices
– Computes a new probability of visiting each vertex in
the graph biased by the vertices on the starting set
– Personalized Page-rank score ➔ a natural relative
distance (or closeness) with respect to the vertices
from the starting set
– Algorithm generates regular pagerank values when
starting set contains all vertices in the graph
15
Starting set
of vertices

Anomaly Detection Procedure
1. Compute Pagerank (PR)
2. Select a Specialty S (e.g. Optometrist)
3. Find the set of doctors of the specialty (starting set)
4. Personalized Pagerank from the starting set (PPR)
5. Mark procedures that has PPR – PR > ɸ as specialty action
6. For every doctor X of other specialties W who has high PPR
value.
– Mark doctor X as anomalous
– If doctor X indeed performed any of specialty action
– And if W does NOT have more than δ% of such doctors
X
Doctors
900,000 HCPCS
6,000
Same specialty
(starting set)
Anomalous
(other specialty)
Edges
9,000,000
Specialty
Actions

Combining Graph Analytics and Machine Learning
Graph Analytics
• Compute graph metric(s)
• Explore graph or compute
new metrics using ML result
Machine Learning
• Build predictive model
using graph metric
• Build model(s) and
score or classify data
17
Add to
structured data
Add to graph

Encoding similarity for use in machine learning
• Graph captures fine-grained relationship between data entities
– As before, closeness can be defined and measured on the graph
– Providing numeric representation of your data that retains the distance information
18
Raw
Data
ML
Model
Graph
Representation
Numeric Representation
(N-dimensional vector)
x, y: data entity (represented as vertex in graph)
v(x), v(y): n-dimensionsal vector representation of x and y
x, y close in graph ➔ ∥ v(x) - v(y) ∥ small in n-dimensional vector space

Encoding similarity for use in machine learning
• Different approaches available
– eg. exploiting techniques from modern NLP
(natural language processing)
• Used Word2Vec in our example
– a ML technique that learns closeness between
words from large number of sentences
• Perform many random walks on the
graph
• Apply W2V technique on random walk
traces, treating vertices as words
19
KDD‘14

Deepwalk – Translate graphs to a vector space
20

Practical example – Student classification
• Can you predict a student’s major or
department just by looking at the
classmates in the course that (s)he is
taking?
• Very similar to customer segmentation
problem
– Student => Customer
– Course taking => Item or service purchase
– Department => Segment label
21
CS
ME
10.003
10.004
10.005
11.103
11.213
12.118
students courses

Evaluation – Comparison
1. CNN trained on “standard” features
(e.g., student age, courses taken, …)
2. Use PPR and predict the department of
the highest-scoring vertex
3. Train a CNN on vertex embeddings
extracted with DeepWalk
4. Add “standard” features beside graph
embeddings
22
CS
ME
10.003
10.004
10.005
11.103
11.213
12.118
students courses

• (Result #1) Graph-based prediction gives better
result than naïve application of ML (e.g. CNN) on
basic student features (e.g. age, gender,
background, …)
• (Result #2) Deep-Walk preserves information
from graph representation
• (Result #3) Deep-Walk allows to combined graph
data with other features
Results
CNN on Original Features
PPR (Graph Algorithm)
CNN on Extracted Graph Features
(from deep-walk)
CNN on Original + Graph Features

Summary
Graph capabilities in Oracle Products and Cloud Services
• Graph databases are powerful tools, complementing relational databases
– Especially strong for analysis of graph topology and connectedness
• Graph analytics offer new insight
– Especially relationships, dependencies and behavioural patterns
• Oracle Property Graph technology offers
– Comprehensive analytics through various APIs, integration with relational database
– Scaleable, parallel in-memory processing
– Secure and scaleable graph storage using Hadoop platform or Oracle Database
• Available both on-premise or in the Cloud already today
24

Appendix: Oracle Graph Analytics Architecture
Scalable and Persistent StorageScalable and Persistent Storage
Graph Storage ManagementGraph Storage Management
Graph Analytics
In-memory Analytic Engine
Graph Analytics
In-memory Analytic Engine
Blueprints & SolrCloud / Lucene
RESTWebService
Python,Perl,PHP,Ruby,
Javascript,…
Java APIs
Java APIs/JDBC/SQL/PLSQL
25
Visualization
RIntegration(OAAgraph)
Sparkintegration

Appendix: Graph Cloud Service
• “One-click” deployment: no installation, zero configuration
– Automated failure detection and recovery
• Automated graph modeler
– Easily convert your relational data into property graphs
• Pre-built Algorithms, Flows and SQL-like graph query language
– Java, Groovy
– Rest APIs
• Rich User Interface
– Low code / zero code features
– Notebook support and powerful data visualization features
Fully managed graph cloud service

Appendix: More information
• Oracle Property Graph Technologies OTN product page:
https://www.oracle.com/database/technologies/spatialandgraph/property-graph-features.html
– White papers, software downloads, documentation and videos
• Oracle Labs Tutorials
https://docs.oracle.com/cd/E56133_01/latest/tutorials/index.html
• Blog post series on setting up Graph Analysis on Oracle Cloud
https://blogs.oracle.com/oraclespatial/how-to-enable-oracle-database-cloud-service-with-property-graph-
capabilities
– Free cloud credits available on http://cloud.oracle.com
• Blog – examples, tips & tricks: blogs.oracle.com/bigdataspatialgraph
• @OracleBigData, @SpatialHannes, @JeanIhm Oracle Spatial and Graph Group
27

Fraud Detection in Financial Services using Graph Analysis and Machine Learning

More Related Content

Similar to Fraud Detection in Financial Services using Graph Analysis and Machine Learning (20)

More from Thomas Teske (8)

Recently uploaded (20)

Fraud Detection in Financial Services using Graph Analysis and Machine Learning