The document explores the relationship between data science and traditional scientific methods, highlighting potential pitfalls and the role of social sciences and humanities. It presents various data science lifecycle models, computational mathematics, and agent-based modeling examples, illustrating how these frameworks can inform decision-making in complex systems. Additionally, it emphasizes the need for a philosophical perspective on algorithms to address ethical considerations in data science practices.

1. Varied Encounters
With Data Science
How does data science relate to traditional scientific and
computational approaches? What can we learn about data science
pitfalls from these approaches? Is there a role for the social sciences
and humanities?
A partial exposition by example.
Gilbert Peffer
gilbert@kubify.co
25 August 2020

2. About
A quick overview of what we are
covering in our discussions today.
★ Method
★ A data science lens
★ Computational mathematics
★ Agent-based models
★ Quantified self
★ Learning analytics
★ Business intelligence
★ Philosophy of algorithms

3. Method
Quizzing projects through a data
science lens.
Poking the data science lens
through project work.
Source: https://jobs-to-be-done.com/what-is-jobs-to-be-done-fea59c8e39eb

4. A Data Science
Lens
Three data science work process
and lifecycle models to motivate
the discussion and questions.
Source: https://www.flickr.com/photos/marc_smith/6792756018

5. Predictive Analytics Workflow Analytics platform developer
https://www.mathworks.com/discovery/predictive-analytics.html
Data Science Workflow AI consultant
http://sadiestlawrence.com/home/data-science-workflow
Data Science Lifecycle Business solution provider
https://docs.microsoft.com/en-us/azure/machine-learning/team-data
-science-process/lifecycle-data
Data Science Lifecycle Models

6. Predictive Analytics Workflow
Data Science Lifecycle Models
Source: https://www.mathworks.com/discovery/predictive-analytics.html

7. Example - Interaction between two research fields
Data Science Lifecycle Models -> Predictive Analytics Workflow

8. Data Science Workflow
Data Science Lifecycle Models
Linear (?) Nonlinear Problem definition vs. analysis
Source: http://sadiestlawrence.com/home/data-science-workflow

9. Example - Cooum River and Environs in Chennai, India
Data Science Lifecycle Models -> Data Science Workflow
Checkland’s Soft System Methodology Expressing the problem situation (by participants)
Source: Bunch, M. J. (2003). Soft systems methodology and the ecosystem approach: a system study of the Cooum River and environs in Chennai, India. Environmental Management, 31(2), 0182-0197.

10. Data Science Lifecycle
Data Science Lifecycle Models
Source: https://docs.microsoft.com/en-us/azure/machine-learning/team-data-science-process/overview

11. Example - The Hubspot Flywheel
Data Science Lifecycle Models -> Data Science Lifecycle
Data-based insights are
secondary
Data-based insights are
a necessity
Source: https://www.webdew.com/hubspot/demo
Data hooks

12. Examples of problem areas
Data Science Lifecycle Models
Theory, Model, World
View, Ideology,
Concept
Assumptions
Evidence
explicit
Assumptions that
are thought to be
well-understood
FUEL ENGINE WORK
- Scientific jobs
- Engineering jobs
- Policy jobs
- Thinking vs. acting jobs
Unintended
consequences
Intended
outcomes
implicit
Assumptions that
are well-
understood
Jobs done
Assumptions that we
aren’t aware of
Definitional struggles
Boundary struggles

13. Computational
Mathematics
Equations, spaces, errors, and
data.
Source: https://commons.wikimedia.org/wiki/File:Ray_40MTet.png

14. Equations - Turbulent gas flow
Computational Mathematics (1992)
Leonardo da Vinci
Cray 2 Supercomputer
Navier-Stokes Equations
k- 𝛆 Turbulence Model
Finite Volume Method
Source: https://dayintechhistory.com/tag/cray-2/
Source: https://journals.physiology.org/doi/pdf/10.1152/ajplung.00378.2016

15. Equations - Turbulent gas flow
Computational Mathematics (1992)
● Business understanding? => Physics.
● Modelling? => PDEs; well-researched.
● Data? => Initial & boundary conditions;
geometry.
● Deployment? => Super computer; not for
production.
Caution: Engineering perspective tends to
take business understanding for granted.

16. Spaces - Adaptive mesh generators
Computational Mathematics (1990)
Laminar compressible flow Finite Element 2-step Taylor-Galerkin
Mesh generation
Supersonic shock waves

17. Spaces - Adaptive mesh generators
Computational Mathematics (1990)
● Problem? => Set by the customer (ESA).
● Validate? => Standard cases (e.g. wedge).
● Hypothesis? => Sub vs. supersonic.
● Model? => PDE + mesh.
● Deployment? => Computer; not for
production.
Caution: Well-defined data science workflow,
but mesh dependency illustrates effect of
problem-external dependency.

18. Errors - Solving (near-)hyperbolic PDEs
Computational Mathematics (1994)
Diffusion term 𝛆 = 0.1 Diffusion term 𝛆 = 0.000001
=>
Convection-diffusion PDE

19. Errors - Solving hyperbolic PDEs
Computational Mathematics (1994)
● Problem? => Abstract and stylised.
● Validate? => Standard cases (a step BC).
● Model? => Simpler physics but worse
numerics. Method selection crucial.
● Deployment? => Academic journals.
Caution: Simple problem but counterintutive
behaviour. Error measures for well-behaved
situations don’t work. Go back to the math
drawing board.

20. Data - Stochastic financial models
Computational Mathematics (90’s)
Interest rate models
Vasicek model
Two-factor Hull and White model
Complex financial market interlinkages
Sources: https://www.ecb.europa.eu/pub/pdf/other/financialstabilityreview200806en.pdf,
https://www.bankofengland.co.uk/-/media/boe/files/financial-stability-report/2007/october-2007.pdf,
https://www.newyorkfed.org/medialibrary/media/research/staff_reports/sr458.pdf
Source: https://www.investopedia.com/articles/economics/08/yield-curve.asp

21. Data - Stochastic financial models
Computational Mathematics
● Data? => Single yield curve = OK; financial
system = NOT OK.
● Processing? => Vast computing capacities and
capabilities in financial firms. Regulators?
● Models? => Very sophisticated, but data hungry
for complex securities. Even more so for
systemic risk.
● Integration? => Typically extensive in financial
firms.
Caution: Firm-level: Availability of sophisticated
models can hide problems with the data. CDOs
etc.@ GFC. Regulator-level: Data on global
counterparty exposures often not available where
most needed (e.g. large hedge fund positions).

22. Agent-Based
Models
Online business models and
systemic risk.
Source: https://www.economist.com/finance-and-economics/2010/07/22/agents-of-change

23. Business models - SimWeb
Agent-Based Models
Stylised consumer-provider model
Partners: 6
Countries: 5
Duration: 2002-2005
Budget: 1,85 M€
Stylised business model and environment

24. Business models - SimWeb
Agent-Based Models
Small data & No data
● Business understanding? => Emerging;
understanding vs. performing the market.
● Models? => Agent-models, but constraint by
limited business understanding.
● Data? => No understanding, no model,... what
data?
Caution: Not understanding the problem situation
makes model choice very tough, unless
‘understanding’ is replaced by ‘performing’. Sound
judgement and expertise are crucial. Also, it is
‘negotiating’ rather than ‘understanding’ => bottom
up vs. top-down.

25. Systemic Risk in Economics and Finance (I)
Agent-Based Models
Bank of England risk transmission map
Source:
https://link.springer.com/chapter/10.1007/9
78-3-319-23947-7_12
Commercial bank
network
Transmission
modes
Source:
https://check-risk.com/network-risk-system/
Source:
https://www.bankofengland.co.uk/-/media/boe/files/financial-stability-report/2
007/october-2007.pdf

26. Systemic Risk in Economics and Finance (I)
Agent-Based Models
● Business understanding? => Emerging
understanding of financial networks. But less
appreciation about interactions with other
systems such as health, the environment,
politics, etc.
● Models? => Model ecosystems are needed.
● Data? => limited to none.
● Politics => Subsumed in ‘understanding’, but
really a key dimension.
Caution: In complex systems, system feedback can
lead to large-impact consequences. The silos of
science are not well-equiped to deal with this. E.g.
economy vs. health controversy in Covid-19. Also,
‘Understaning’, ‘Models’, ‘Data’ all have a political
dimension.

27. Systemic Risk in Economics and Finance (II)
Agent-Based Models
Source: https://limn.it/haldane-complexity/
Macro-Networks Micro-Mechanisms
Source: Peffer, G., & Llacay, B. Exploiting the Full Potential of Multiagent‐Based Simulation in Finance: Principles and Methods.

28. Systemic Risk in Economics and Finance (II)
Agent-Based Models
● Access data? => Multi-scale means huge
amounts of diverse data. 1bn contract pages
just for a single financial securities category!
● Data processing? => Crazy.
● Models? => New territory. Multi-sectorial /
-system models.
● Integrative Analytics => For whom and for what
purpose?
Caution: Idem.

29. Quantified Self
Understanding the emotional
antecedents of the disposition
bias. Managing the bias through
emotion regulation.
Source: https://tinyurl.com/ycdfs8c8

30. Emotions, cognitive biases, and learning - xDelia
Quantified Self
xDelia learning pathway
Emotion regulation modes
Partners: 7
Countries: 6
Duration: 2007-2013
Budget: 3,1 M€
ER and trader performanceDisposition effect
Source: https://www.cairn.info/revue-finance-2009-1-page-51.htm#
Source: Fenton-O’Creevy, M., Vohra, S.
(2011). Emotion regulation and trader
performance.
Source: Peffer, G. and Fenton-O’Creevy, M, Eds. xDelia –
Emotion-centred financial decision-making and learning
Source: Gross, J.J., & Thompson, R.A. 2007. Emotion
regulation: Conceptual foundations.

31. Emotions, cognitive biases, and learning - xDelia
Quantified Self
● Access data? => Sensors to measure heart rate
=> work intensive and intrusive. Financial data
=> easy-ish.
● Data processing? => Small data sets.
● Models? => Proto-models. Emerging field of
emotion regulation. Qualitative fit with literature.
● Integrative Analytics => The xDelia learning
interventions (diagnostics; training; transfer).
Caution: In naturalistic environments, data
collection from people who go about their daily jobs
is a challenge. Data gathering needs to be ongoing
and learning interventions should aid in getting
more relevant data to validate the (proto-)model.

32. Learning
Analytics
Informal, organisational, and
network learning.
Source: https://www.flickr.com/photos/gforsythe/20543912596/in/photostream/

33. Informal learning - Layers
Learning Analytics
A) Learning approaches
B) Locus of learning
Partners: 18
Countries: 8
Duration: 2012-2016
Budget: 9.9 M€
Source:
https://elearning.adobe.com/2019/07/classic-learning-research-practice-leading-strategy/
Source: http://results.learning-layers.eu/scenarios/
Current learning analytics
Source: https://learninganalytics.ubc.ca/about-the-project/why-learning-analytics/

34. Informal learning - Layers
Learning Analytics
What are the challenges of learning analytics here?
● Generally => few operational models, sparse data.
● Informal learning (individual)?
● Learning in a community of practice?
● Organisational learning?
● Cross-organisational learning? => knowledge, learning, and innvoation in industry clusters
Caution: The standard learning analytics won’t work in these cases. They measure the wrong thing
the wrong way for the wrong target group.

35. Business
Intelligence
How data science impacts
business practice.
Source: https://atonce.be/blog/business-intelligence-is-a-must/

36. Scalable AI factory Karim Lakhani, HBS
Business Intelligence
Ant Financial uses AI to drive digital learning Generalisable AI factory layer
Source: https://youtu.be/O5242n_W9vA

37. The quantified pig farm
Business Intelligence
Source: https://youtu.be/O5242n_W9vA

38. Scalable AI factory = Scalable ethics problems
Business Intelligence
“In an analogue world you can basically have
your bias limited. In a digital world, you can
bias at scale.
“We have to think about ethics at scale.”
Karim Lakhani, Harvard Business School

39. Philosophy of
Algorithms
The need for a social sciences
and humanities approach in the
data sciences.
Source: https://www.gapingvoid.com/blog/2007/08/20/complicated/

40. The critical and integrative force of philosophy
Philosophy of Algorithms
1951 (34) 1980 (63) 1987 (70)

41. System decisions and boundary judgements
Philosophy of Algorithms
Werner Ulrich’s critical systems heuristics
Economics
frame
Social justice
frame
Source: W. Ulrich (2000). Reflective practice in the civil society: the contribution of
critically systemic thinking.
The interplay between system boundaries, facts, and values Example: transport policy

42. Accountability of algorithms
Philosophy of Algorithms
In Cloud Ethics Louise Amoore examines how machine learning
algorithms are transforming the ethics and politics of contemporary
society. Conceptualizing algorithms as ethicopolitical entities that are
entangled with the data attributes of people, Amoore outlines how
algorithms give incomplete accounts of themselves, learn through
relationships with human practices, and exist in the world in ways that
exceed their source code. In these ways, algorithms and their
relations to people cannot be understood by simply examining their
code, nor can ethics be encoded into algorithms. Instead, Amoore
locates the ethical responsibility of algorithms in the conditions of
partiality and opacity that haunt both human and algorithmic decisions.
To this end, she proposes what she calls cloud ethics—an approach to
holding algorithms accountable by engaging with the social and
technical conditions under which they emerge and operate.
For a discussion, see: “Review of Louise Amoore, Cloud Ethics: Algorithms and the Attributes of Ourselves
and Others.” at https://www.theoryculturesociety.org/review-amoore-cloud-ethics/

43. Thank you!
Gilbert Peffer
gilbert@kubify.co
25 August 2020