Article 26: Algorithmic System Integrity: Explainability (Part 3) - Complicated Processes

Spoken by a human version of this article.

TL;DR (TL;DL?)

• Algorithmic processes are often complicated by intricate data flows and transformations.
• Data flow diagrams and documentation can help make processes simpler.

About this podcast

A podcast for Financial Services leaders, where we discuss fairness and accuracy in the use of data, algorithms, and AI.

Hosted by Yusuf Moolla.
Produced by Risk Insights (riskinsights.com.au).

Episode Transcript

Yusuf (00:14):
This article is titled Algorithmic System Integrity
Explainability, and this is partthree of the series.
this one's about complicatedprocesses.
So the two long-term read isalgorithmic processes are often
complicated by intricate dataflows and transformations.
Data flow diagrams anddocumentation can help make
processes simpler.

(00:36):
In a previous article, weexplored the concept of
explainability, its importanceand for challenges.
And then we addressed the firstchallenge, which was complexity.
In this one, we explore somesolutions to the second
challenge, complicatedprocesses.
So to recap that, challengealgorithmic systems often
involve intricate workflowscombining multiple data sources.

(00:58):
Both internal and external andtransformations consider how
much work was involved inunpacking data flows for capital
adequacy projects, puzzle orsolvency, and the resulting
spaghetti.
These are both wide, so manysystems and deep multiple data
elements or variables.
Some of these data flows crossover with credit scoring

(01:20):
processes, pricing models, andclaims flows, so the same
complications apply.
Data lineage was challengingenough with simple algorithms,
and now the flows include morecomplex algorithms and directly
impact customers.
This poses several problems,trust.
Opaque or convoluted processesare not naturally trusted.

(01:41):
Compliance regulatory frameworksrequire clear explanations of
how data flows, especially fordecisions affecting customers,
and then errors, complicatedprocesses make it difficult to
identify where problems occur,increasing the risk of
undetected issues.
Here are a couple of solutions.

(02:01):
Banks and insurers can useseveral methods.
Of course.
Here are three sets of commonlyused methods which may look
familiar if you've had to dealwith capital adequacy projects
or similar.
The first of those is pictureit.
visual tools are tremendouslyhelpful.
One such solution is a data flowdiagram.
It provides a visualrepresentation of how data

(02:23):
enters and moves through yoursystems, what models or
algorithms are involved, and howthese all work together.
As an example, for an insuranceclaims process, a data flow
diagram might show how datamoves from the initial claims
submission through variousvalidation steps.
Risk assessment algorithms,fraud, triage, and then payment

(02:43):
processing.
The diagram can highlight whereexternal data enters the flow
and where automated decisionsoccur.
The second potential solution isto document it.
So documentation typically goesbeyond what even detailed
diagrams can capture, especiallyin explaining transformations,
systems and algorithms.

(03:04):
In documenting, we consider howoften the process changes in
deciding what depth to go todeep in the context of frequent
changes can be difficult tomaintain in the long term.
But automated documentationtools can help with this.
Documentation commonly includesdata provenance tracking, so
that's origins andtransformations, explanations

(03:27):
for each transformation step inplain language if appropriate,
and details about each system,model, and algorithm involved.
As an example, in bankingdocumentation for a credit
scoring system might track how acustomer's data is captured,
normalized.
Combined with credit bureauinformation and put through the

(03:47):
scoring model, this can helpshow how the fields that
influence the final decision areprocessed at each step.
And then the third potentialsolution is to simplify it.
Traditional processes oftencontain redundant
transformations and unnecessarycomplexity.
You may have data formatting orstandardization at various

(04:09):
stages.
Moving these to early in theprocess can make the flow
simpler.
It's not uncommon to findtransformations that reverse
earlier steps.
As system flows evolve overtime, You may be able to
eliminate both and get to thesame result while making the
whole process simpler.
Simplifying processes can alsoreduce operational costs,

(04:32):
improve system performance, andmake maintenance easier.
As an example, a bank's customeronboarding process might be
simplified by checking forredundant identity verification
steps and reducing these, ratherthan running separate
validations for address IDdocuments and personal
information across differentsystems, a unified verification

(04:54):
system could handle all thesechecks while maintaining a clear
explainable process.
In the next episode, the nextarticle in the series focuses on
the third challenge.
That's privacy andconfidentiality, exploring how
to balance transparency withsafeguarding sensitive data.
Thanks for listening.
I.