Fraud Detection Class Slides

Graphs in Fraud Detection
Max De Marzi
Field Engineer, Neo4j
@maxdemarzi

About Me
• Max De Marzi - Neo4j Field Engineer
• My Blog: http://maxdemarzi.com
• Find me on Twitter: @maxdemarzi
• Email me: maxdemarzi@gmail.com
• GitHub: http://github.com/maxdemarzi

Overview
Types of Fraud
• Credit Card Fraud
• First-Party Fraud
• Synthetic Identities and Fraud Rings
• Insurance Fraud
Types of Analysis
• Traditional Analysis
• Graph-Based Analysis
Fraud Detection and Prevention
Common Questions

…but before we get into that …
• What isn’t Fraud?

I don’t know, but I know who does
• Alex Beutel, CMU
• Leman Akoglu, Stony Brook
• Christos Faloutsos, CMU
• Graph-Based User Behavior Modeling: From Prediction to
Fraud Detection
• http://www.cs.cmu.edu/~abeutel/kdd2015_tutorial/

User Behavior Challenges
• How can we understand
normal user behavior?

• How can we find
suspicious behavior?

• How can we find
suspicious behavior?
• How can we distinguish
the two?

Does your little girl like Rambo?

Understanding our Users
• What do we know about them?

MATCH (u:User)-[r:RATED]->(m:Movie) 
RETURN u.gender, u.age,  
COUNT (DISTINCT u) AS user_cnt,  
COUNT (DISTINCT m) AS mov_cnt,  
COUNT(r) AS rtg_cnt

MATCH (me:User {id:1}) -[r1:RATED]-> (m:Movie)  
<-[r2:RATED]- (similar_users:User) 
WHERE ABS(r1.stars-r2.stars) <= 1  
RETURN similar_users.gender,  
similar_users.age,  
COUNT(DISTINCT similar_users) AS user_cnt,  
COUNT(r2) AS rtg_cnt

Little Girls like Movies other Little Girls Like

What do Little Girls Like?
WHERE u.age = 1 AND u.gender = "F" AND r.stars > 3 
RETURN m.title, COUNT(r) AS cnt 
ORDER BY cnt DESC 
LIMIT 10

What do Men 25-34 Like?
WHERE u.age = 25 AND u.gender = "M" AND r.stars > 3 
RETURN m.title, COUNT(r) AS cnt 
ORDER BY cnt DESC 
LIMIT 10

Modeling “Normal” Behavior
• Predict Edges 
(Similar Users)

• Predict Edges 
(Movies I should Watch)

Recommendation Engine with Neo4j
Recommendation

Content Based Recommendations
• Step 1: Collect Item Characteristics
• Step 2: Find similar Items
• Step 3: Recommend Similar Items
• Example: Similar Movie Genres

There is more to life than Romantic Zombie-coms

Collaborative Filtering Recommendations
• Step 1: Collect User Behavior
• Step 2: Find similar Users
• Step 3: Recommend Behavior taken by similar users
• Example: People with similar musical tastes

Using Relationships for Recommendations
Content-based filtering
Recommend items based on what users
have liked in the past
Collaborative filtering
Predict what users like based on the
similarity of their behaviors, activities
and preferences to others
Movie
Person
Person
RATED
SIMILARITY
rating: 7
value: .92

Hybrid Recommendations
• Combine the two for
better results
• Like Peanut Butter and
Jelly

Cypher Query: Movie Recommendation
MATCH (watched:Movie {title:"Toy Story”}) <-[r1:RATED]- () -[r2:RATED]-> (unseen:Movie)
WHERE r1.rating > 7 AND r2.rating > 7
AND watched.genres = unseen.genres
AND NOT( (:Person {username:”maxdemarzi"}) -[:RATED]-> (unseen) )
RETURN unseen.title, COUNT(*)
ORDER BY COUNT(*) DESC
LIMIT 25
What are the Top 25 Movies
• that I haven't seen
• with the same genres as Toy Story
• given high ratings
• by people who liked Toy Story

Let’s try k-nearest neighbors (k-NN)
Cosine Similarity

Cypher Query: Ratings of Two Users
MATCH (p1:Person {name:'Michael Sherman’}) -[r1:RATED]-> (m:Movie),
(p2:Person {name:'Michael Hunger’}) -[r2:RATED]-> (m:Movie)
RETURN m.name AS Movie,  
r1.rating AS `M. Sherman's Rating`,
r2.rating AS `M. Hunger's Rating`
What are the Movies these 2 users have both rated

Cypher Query: Ratings of Two Users
Calculating Cosine Similarity

Cypher Query: Cosine Similarity
MATCH (p1:Person) -[x:RATED]-> (m:Movie) <-[y:RATED]- (p2:Person)
WITH SUM(x.rating * y.rating) AS xyDotProduct,
SQRT(REDUCE(xDot = 0.0, a IN COLLECT(x.rating) | xDot + a^2)) AS xLength,
SQRT(REDUCE(yDot = 0.0, b IN COLLECT(y.rating) | yDot + b^2)) AS yLength,
p1, p2
MERGE (p1)-[s:SIMILARITY]-(p2)
SET s.similarity = xyDotProduct / (xLength * yLength)
Calculate it for all Person nodes with at least one Movie between them

Cypher Query: Your nearest neighbors
MATCH (p1:Person {name:'Grace Andrews’}) -[s:SIMILARITY]- (p2:Person)
WITH p2, s.score AS sim
RETURN p2.name AS Neighbor, sim AS Similarity
ORDER BY sim DESC
LIMIT 5
Who are the
• top 5 Persons and their similarity score
• ordered by similarity in descending order
• for Grace Andrews

Cypher Query: k-NN Recommendation
MATCH (m:Movie) <-[r:RATED]- (b:Person) -[s:SIMILARITY]- (p:Person {name:'Zoltan Varju'})
WHERE NOT( (p) -[:RATED]-> (m) )
WITH m, s.similarity AS similarity, r.rating AS rating
ORDER BY m.name, similarity DESC
WITH m.name AS movie, COLLECT(rating)[0..3] AS ratings
WITH movie, REDUCE(s = 0, i IN ratings | s + i)*1.0 / LENGTH(ratings) AS recommendation
ORDER BY recommendation DESC
RETURN movie, recommendation 
LIMIT 25
What are the Top 25 Movies
• that Zoltan Varju has not seen
• using the average rating
• by my top 3 neighbors

• Predict Edges
• Predict Node Attributes 
(Age, Gender, etc)
Age: 35
Age: ?

• Predict Edges
• Predict Node Attributes
• Predict Edge Attributes 
(Rating)

What Rating should I give 101 Dalmatians?
MATCH (me:User {id:1})-[r1:RATED]->(m:Movie) 
<-[r2:RATED]-(:User)-[r3:RATED]-> 
(m2:Movie {title:”101 Dalmatians”}) 
WHERE ABS(r1.stars-r2.stars) <=1 
RETURN AVG(r3.stars)

• Predict Edges
• Predict Edge Attributes
• Clustering and
Community Detection

Predict a Star Rating purely on Demographics
MATCH (u:User)-[r:RATED]->(m:Movie {title:”Toy Story”}) 
WHERE u.age = 1 AND u.gender = "F"  
RETURN AVG(r.stars)

• Predict Edges
• Predict Edge Attributes
• Clustering and
Community Detection
• Fraud Detection

Two Sides of the Same Coin
Recommendations
• Add the relationship
that does not exist
Fraud Detection
• Find the relationships
that should not exist

Modeling User Behavior
• Modeling normal users
and detecting anomalies
are two sides of
understanding user
behavior

are two sides of
understanding user
behavior
• Rough Model of normal
vs outlier

are two sides of
understanding user
behavior.
• Fine grained models can
find more subtle outliers

are two sides of
understanding user
behavior
• Complex models can
capture normal and
abnormal patterns

are two sides of
understanding user
behavior
• Known fraudulent
patterns can be searched
for directly

Find the Nodes
ArrayList<Node> nodes = new ArrayList<Node>(); 
nodes.add(db.findNode(Labels.CC, “number”, card));
nodes.add(db.findNode(Labels.Phone, “number”, phone));
nodes.add(db.findNode(Labels.Email, “address”, address));
nodes.add(db.findNode(Labels.IP, “address”, ip));

Add the Crosses
for(Node node : nodes){
HashMap<String, AtomicInteger> crosses = new HashMap<String, AtomicInteger>();
crosses.put("CCs", new AtomicInteger(0));
crosses.put("Phones", new AtomicInteger(0));
crosses.put("Emails", new AtomicInteger(0));
crosses.put("IPs", new AtomicInteger(0));
for ( Relationship relationship : node.getRelationships(RELATED, Direction.BOTH) ){
Node thing = relationship.getOtherNode(node);
String type = thing.getLabels().iterator().next().name() + "s";
crosses.get(type).getAndIncrement();
}
results.add(crosses);
}

Examine Results
[{"ips":4,"emails":7,"ccs":0,"phones":4}, -- cc returned 4 ips, 7
emails, and 3 phones.
{"ips":1,"emails":1,"ccs":1,"phones":0}, -- phone returned just 1
item for each cross reference check.
{"ips":2,"emails":0,"ccs":4,"phones":3}, -- email returned 2 ips, 4
credit cards and 3 phones.
{"ips":0,"emails":1,"ccs":3,"phones":2}] -- ip returned 3 credit
cards and 2 phones.

What is a subgraph?
KDD 2015 2

What is a subgraph?
KDD 2015 3
• A Subset of nodes and
the edges between them

What are some useful subgraphs?
Largest dense subgraph
(Greatest average
degree)

E
Ego-network:
the subgraph
among a node and
its neighbors

Graph queries:
find subgraphs of
particular pattern

Graph queries:
find subgraphs of
particular pattern
MATCH (a)--(b)--(c)--(a) 
RETURN *

Graph queries:
find subgraphs of
particular pattern
MATCH (a)—(b)—(c)—
(d)—(a)—(c), (d)—(b) 
RETURN *

Ego-net Patterns
Ni: number of neighbors of ego i
Ei: number of edges in egonet i
Wi: total weight of egonet i
λw,i: principal eigenvalue of the
weighted adjacency matrix of egonet i

Power Law Density
slope=2
slope=1
slope=1.35

Find Groups within Ego-Nets
Link

First-Party Fraud
• Fraudster’s aim: apply for lines of credit, act normally, extend credit,
then…run off with it
• Fabricate a network of synthetic IDs, aggregate smaller lines of credit
into substantial value
• Often a hidden problem since only banks are hit
• Whereas third-party fraud involves customers whose identities are stolen
• More on that later…

So what?
• $10’s billions lost by banks every year
• 25% of the total consumer credit write-offs in the USA
• Around 20% of unsecured bad debt in E.U. and N.A. is misclassified
• In reality it is first-party fraud

Then the fraud happens…
• Revolving doors strategy
• Money moves from account to account to provide legitimate transaction
history
• Banks duly increase credit lines
• Observed responsible credit behavior
• Fraudsters max out all lines of credit and then bust out

… and the Bank loses
• Collections process ensues
• Real addresses are visited
• Fraudsters deny all knowledge of synthetic IDs
• Bank writes off debt
• Two fraudsters can easily rack up $80k
• Well organized crime rings can rack up many times that

Discrete Analysis Fails to predict…

…and Makes it Hard to React
• When the bust out starts to happen, how do you know what to cancel?
• And how do you do it faster then the fraudster to limit your losses?
• A graph, that’s how!

Probably Non-Fraudulent Cohabiters

Probable Cohabiters Query
MATCH (p1:Person)-[:HOLDS|LIVES_AT*]->() 
<-[:HOLDS|LIVES_AT*]-(p2:Person)
WHERE p1 <> p2
RETURN DISTINCT p1

Risky People
MATCH (p1:Person)-[:HOLDS|LIVES_AT]->()
<-[:HOLDS|LIVES_AT]-(p2:Person)
-[:HOLDS|LIVES_AT]->()
WHERE p1 <> p2 AND p2 <> p3 AND p3 <> p1
WITH collect (p1.name) + collect(p2.name) +
collect(p3.name) AS names
UNWIND names AS fraudster
RETURN DISTINCT fraudster

Pretty quick…
Number of people: [5163]
Number of fraudsters: [40]
Time taken: [100] ms

Localize the focus
MATCH (p1:Person {name:'Sol'})-[:HOLDS|LIVES_AT]-()…
Number of fraudsters: [5]
Time taken: [13] ms

Quickly Revoke Cards in Bust-Out
MATCH (p1:Person)-[:HOLDS|LIVES_AT]->()
-[:HOLDS|LIVES_AT]->()
WHERE p1 <> p2 AND p2 <> p3 AND p3 <> p1
WITH collect (p1) + collect(p2)+ collect(p3) 
AS names
UNWIND names AS fraudster
MATCH (fraudster)-[o:OWNS]->(card:CreditCard)
DELETE o, card

Whiplash
http://georgia-clinic.com/blog/wp-content/uploads/2013/10/whiplash.jpg

Whiplash for Cash
http://georgia-clinic.com/blog/wp-content/uploads/2013/10/whiplash.jpg http://cdn2.holytaco.com/wp-content/uploads/2012/06/lottery-winner.jpg

Whiplash for Cash Example
Accidents
Cars
Doctor Attorney
People
Drives
Is Passenger
Drivers 
Passengers 
Witnesses

Risk
• $80,000,000,000 annually on auto insurance fraud and growing
• Even small % reductions are worthwhile!
• British policyholders pay ~£100 per year to cover fraud
• US drivers pay $200-$300 per year according to US National Insurance
Crime Bureau

Regular Drivers Query
MATCH (p:Person)-[:DRIVES]->(c:Car)
WHERE NOT (p)<-[:BRIEFED]-(:Lawyer)
AND NOT (p)<-[:EXAMINED]-(:Doctor)
AND NOT (p)-[:WITNESSED]->(:Car)
AND NOT (p)-[:PASSENGER_IN]->(:Car)
RETURN p,c LIMIT 100

Genuine Claimants Query
MATCH (p:Person)-[:DRIVES]->(:Car),
(p)<-[:BRIEFED]-(:Lawyer),
(p)<-[:EXAMINED]-(:Doctor)
OPTIONAL MATCH (p)-[w:WITNESSED]->(:Car),
(p)-[pi:PASSENGER_IN]->(:Car)
RETURN p, count(w) AS noWitnessed, 
count(pi) as noPassengerIn

Fraudsters
MATCH (p:Person)-[:DRIVES]->(:Car),
(p)<-[:BRIEFED]-(:Lawyer),
(p)<-[:EXAMINED]-(:Doctor),
(p)-[w:WITNESSED]->(:Car),
(p)-[pi:PASSENGER_IN]->(:Car)
WITH p, count(w) AS noWitnessed,  
count(pi) as noPassengerIn
WHERE noWitnessed > 1 OR noPassengerIn > 1
RETURN p

Auto-fraud Graph
• Once you have the fraudsters, finding their support team is easy.
• (fraudster)<-[:EXAMINED]-(d:Doctor)
• (fraudster)<-[:BRIEFED]-(l:Lawyer)
• And it’s also easy to find their passengers
• (fraudster)-[:DRIVES]->(:Car)<-[:PASSENGER_IN]-(p)
• And easy to find other places where they’ve maybe committed fraud
• (fraudster)-[:WITNESSED]->(:Car)
• (fraudster)-[:PASSENGER_IN]->(:Car)
• And you can see this in milliseconds!

Online Payments Fraud (First-Party)
• Stealing credentials is commonplace
• Phishing, malware, simple naïve users
• Buying stolen credit card numbers is easy
• How should one protect against seemingly fine credentials?
• And valid credit card numbers?

We are all little stars
• Username and passwords
• Two-factor auth
• IP addresses, cookies
• Credit card, paypal account
• Some gaming sites already do some of this
• Arts and Crafts platform Etsy already embraced the idea of graph of
identity

An Individual Identity Subgraph
128.240.229.18
fred@rbs.co.uk
1234LOL

Other Specific
Considerations
Specific Weighted Identity Query
MATCH (u:User {username:'Jim', password: 'secret'})
OPTIONAL MATCH
(u) -[cookie:PROVIDED]->(:Cookie {id:'1234'})
OPTIONAL MATCH
(u)-[address:FROM]->(:IP {network:'128.240.0.0'})
RETURN SUM(cookie.weighting) + SUM(address.weighting)
AS score
Bare
Minimum
Other Specific
Considerations
Final
Decision

General Weighted Identity Query
MATCH (u:User {username:'Jim', password: 'secret'})
OPTIONAL MATCH (u)-[rel]->()
WHERE has(rel.weighting)
RETURN SUM(rel.weighting) AS score
Bare
Minimum
All Available
Weightings
Final
Decision

An Individual Login History
fred@rbs.co.uk
1234LOL

From 1st to 3rd Party
• The 1st party identity graph can easily be extended to 3rd party fraud
• Like in the bank fraud ring, fraudsters can mix-n-match claims
• Start with a few phished accounts and expand from there!

Shared Connections
128.240.229.18
fred@rbs.co.uk
1234LOL nick@bearings.com
Ca$hMon£y

Graphing Shared Connections
Hmm….

Scan for Potential Fraudsters
MATCH (u1:User)--(x)--(u2:User)
WHERE u1 <> u2 AND NOT (x:IP)
RETURN x
Network in
common is OK

Stop specific fraudster network, quickly
MATCH path =  
(u1:User {username: 'Jim'})-[*]-(x)-[*]-(u2:User)
WHERE u1<>u2 AND NOT (x:IP) AND NOT (x:User)
RETURN path

How do these fit with traditional fraud prevention?
http://www.gartner.com/newsroom/id/1695014
Gartner’s Layered Fraud Prevention Approach

Bank Fraud
http://gist.neo4j.org/?dfdfbddfdc63f4858f80

Credit Card Fraud Detection
http://gist.neo4j.org/?3ad4cb2e3187ab21416b

Whiplash for Cash
http://gist.neo4j.org/?6bae1e799484267e3c60

Ask for help if you get stuck
• Online training - http://neo4j.com/graphacademy/
• Videos - http://vimeo.com/neo4j/videos
• Use cases - http://www.neotechnology.com/industries-and-use-cases/
• Meetups
• Books to get your started
• http://www.graphdatabases.com
• http://neo4j.com/book-learning-neo4j/

Deep Neural Networks for Bank Fraud
https://www.youtube.com/watch?v=TAer-PeIypI
Fraud Detection starts about half-way (after intro)

Thanks for listening
@maxdemarzi

Fraud Detection Class Slides

More Related Content

Viewers also liked (13)

Similar to Fraud Detection Class Slides (20)

More from Max De Marzi (20)

Recently uploaded (20)

Fraud Detection Class Slides