The Danish Business Authority (DBA) uses machine learning and a graph database to analyze data from various sources like the Danish Business Registry and tax authorities. The goal is to detect potential fraud in near real-time by profiling businesses and monitoring their lifecycles. A risk score is calculated based on 224 parameters from raw data, processed data, and model outputs. If suspicious activity is detected, like a sudden increase in the risk score, a VAT control may be triggered. The system aims to continuously improve by learning from the data and evaluating model performance.
Reassessing the Bedrock of Clinical Function Models: An Examination of Large ...
GraphTour 2020 - Danish Business Authority: First line of Defence
1. Marius Hartmann, The Danish Business Authority
First Line of Defence
03. March 2020
2. What is the DBA?
Erhvervsstyrelsen 2
Ministry of Industry, Business and Financial Affairs
https://www.linkedin.com/in/mariushartmann/
The Danish Business Authority
The Danish Business Registry(CVR)
part of
runs
ML Lab
part of
works at
Marius
3. Reduce the yearly VAT deficit by 1 billion DKK
Erhvervsstyrelsen 3
Mission
*approx. 133 million EUR
5. Tactical choices
Erhvervsstyrelsen 5
Batch
Profiling
Lease domain experts
Tech first
Supercomputer
Black box
Precision first
Redundant systems
Manual processes
Automated decision making
Unsupervised
Running models without governance
Limited traceability
Complexity
Rigid
Concentration
Proprietary
POC
Monolith
-Real time
Prevent fraud
Retain and build capability
Business driven
Fast response times
Explainability
Data ethics
Reusable components
Automation
Decision support
Supervised
Continous improvement
Single source of truth
Simplicity
Adaptive
Distribution
Open source
Productionization
Micro services
+
15. Machine learning
controls all identity
papers for foreign
business actors
ML controls that
fictional assets are
not inserted
‘Weaponize’
unstructured data
concerning
complaince
Control new
businesses for
concerns of fraud
Identity
Assets
Audits
1.st line
Model to model
Erhvervsstyrelsen 15
20. 0 10050
0
Erhvervsstyrelsen 20
6 7
64%
:Person
:Business
:Business status
:Business type
:Business classification
:Address
A business starts
Months
Risc
score
21. 0 10050
0
Erhvervsstyrelsen 21
6 7 11
64%
:Person
:Business
:Business status
:Business type
:Business classification
:Address
A business starts
Months
Risc
score
22. 0 10050
0
Erhvervsstyrelsen 22
6 7 1211
72%
:Person
:Business
:Business status
:Business type
:Business classification
:Address
A business starts
Months
Risc
score
23. 0 6 7 12 17
VAT control!
Possible fraud
Erhvervsstyrelsen 23
:Person
:Business
:Business status
:Business type
:Business classification
:Address
:VAT control
11
A business starts
Months
24. The principles
▪ Uses data from other authorities
▪ Event driven so we can react in (near) real time
▪ Graph adoption to contextualize business lifecycles
▪ Meta data strategy to produce data from data
▪ ML enriched automation so we may adopt machine generated insight
▪ Monitor and trace usage so we can explain
▪ Evaluate and improve continuously
Erhvervsstyrelsen 24
25. ▪ All companies controlled in near real time
▪ Continuously
▪ Full transparency and traceability
Erhvervsstyrelsen 25
Effect of graph + ML