top of page
Citrus

About Behavioural Data Science

This page provides a brief summary of Behavioural Data Science

Behavioral Data Science as a Logical Evolution of Behavioral Science

Behavioral Data Science is a new, emerging, interdisciplinary field, which combines techniques from the behavioral sciences, such as psychology, economics, sociology, and business, with computational approaches from computer science, statistics, data-centric engineering, information systems research and mathematics, all in order to better model, understand and predict behavior.

Behavioral Data Science lies at the interface of all these disciplines (and a growing list of others) — all interested in combining deep knowledge about the questions underlying human, algorithmic, and systems behavior with increasing quantities of data. The kinds of questions this field engages are not only exciting and challenging, but also timely, such as:

Behavioral Data Science is capable of addressing all these issues (and many more) partly because of the availability of new data sources and partly due to the emergence of new (hybrid) models, which merge behavioral science and data science models. The main advantage of these models is that they expand machine learning techniques, operating, essentially, as black boxes, to fully tractable, and explainable upgrades. Specifically, while a deep learning model can generate accurate prediction of why people select one product or brand over the other, it will not tell you what exactly drives people’s preferences; whereas hybrid models, such as anthropomorphic learning, will be able to provide this insight.

Behavioral Data Science Is NOT Behavioral Analytics

It is important to understand, that behavioral data science addresses not only consumer behavior (which is the realm of behavioral analytics — a field often confused with behavioral data science). Behavioral Data Science incorporates 3 important strands: human behavior, algorithmic behavior and systems behavior.

  • The human behavior strand provides a range of methodological tools, which originate from research in psychology, decision theory, and behavioral science in order to show how standard methods used in these fields can be enriched by data science techniques to explain human behavior utilizing large datasets.

  • The algorithmic behavior strand, on the one hand, combines a range of algorithmic methods from statistics, computer science, mathematics, as well as other sciences, which, on the one hand, can be used to explain and predict behavior; and, on the other hand, deals with machine behavior and behavior of algorithms, as machines and algorithms also exhibit behavioral regularities and biases.

  • Finally, the systems behavior incorporates methods which allow modelling complex systems, networks, markets, cultural differences and culture in a wide variety of contexts.

Behavioral Data Science “In the Wild”

 

Behavioral data science emerges as a direct response to the need for studying behavior “in the wild”, outside the “sterile” laboratory setting and controlled environments. This task is especially important when we consider interactions between humans and technology. Decision support systems, suggestion systems, automation, etc. — all these technologically intense aspects of human life require accurate predictions of what people like, what people prefer, and where people need help of automated agents and/or algorithms. Further, we need to better understand how humans and algorithms can harmoniously co-exist in a system as well as how to make these systems resilient to change.

Think of the recent shortages of goods in shops we have all experienced due to COVID-19. These shortages highlighted how bad our prediction models are and how they fail to deal with sporadic, lumpy or abnormal demand patterns (such as panic-buying behaviour) which is made worse though manual planning corrections done by the retailers.

In fact, much of the current strategy in retail supply chain management (SCM) is defined by:

  1. Demand forecasting (to capture the demand dynamics); and

  2. Stock & replenishment (to buffer for demand variations, forecast error or to deal with promotions).

Current demand forecasting focuses on individual stock keeping units (SKUs) and relies on smoothing, traditional time series analysis and machine learning techniques. While these methods are effective for predictable demand, it has a number of disadvantages for sporadic, lumpy or abnormal demand dynamics:

  • Traditional forecasting techniques generally deal very poorly with exceptional demand signals, leading to stock-outs or waste due to excess stock which is particularly problematic for perishable goods. This is often made worse when planners try to correct it manually;

  • Planning stock and replenishment generally fail to deal with sudden shocks due to various externalities (e.g.: riots, environmental disasters such as bushfires, COVID19, etc.).

Under these circumstances, recent advances in computer science, statistics, and mathematics offer several methods which try to model human behaviour. Specifically, the methodology of machine learning and, more recently, deep learning allows us to generate predictions useful for many different facets of human life. Yet, there are many aspects of human life and decision making where machine learning and deep learning fail to provide reliable and accurate results. One of the most notorious examples is suggestion systems: many of us regularly shop online using different platforms (such as Amazon) and receive suggestions for future purchases. Yet, very few of us find these suggestions helpful.

One of the reasons why AI fails in many cases to correctly anticipate human behaviour is that AI algorithms tend to ignore existing insights from decision theory and behavioural science. And this is where Behavioral Data Science becomes very helpful. By combining behavioural science models with AI algorithms, we are able to significantly improve and simplify predictions of human behaviour in a wide variety of contexts.

For example, anthropomorphic learning allows us to develop more functional systems and algorithms, which better understand humans. This methodology is explainable, traceable, requires smaller training sets and, in retails context often outperforms existing algorithms by generating more accurate predictions. Since not only data, but also behavior propagates through the entire supply chain, Behavioral Data Science has a real potential to revolutionize our predictive capabilities and make supply chains resilient and agile.

So, What Is the Future?

Advances in artificial intelligence make possible complex socio-technical systems that have enormous potential to promote human wellbeing. Yet rising public anxiety concerning their implications, including, for example, job security, societal cohesion, the integrity of democratic processes and the opaque and largely unaccountable flows of information and power between institutions across the globe, point to a crisis of trust arising from their development and use.

Optimists foresee a world of ‘super-abundance’ with machines satisfying humanity’s basic needs while solving our most serious social challenges, such as malnutrition, food insecurity, poverty, disease, disability and climate change. Pessimists, by contrast, envisage human communities enslaved by machine overlords as the rule of technology replaces the rule of law. Irrespective of where one stands in this debate, policy-makers everywhere now recognise the urgent need to understand and respond appropriately to the rapid proliferation of AI systems to ensure that AI will promote, rather than undermine, our individual and collective well-being. This is the core global challenge to which Behavioral Data Science aims to respond in the future.

Behavioral Data Science of the future will also identify, map and explain the interactions between society and AI systems with a view to establishing a robust evidence base that can inform policy responses. For example, it will understand how various social groups perceive AI, their vulnerability to its attendant risks, and their assessment of the trade-offs between competing values embedded into AI systems. These might include questions such as:

  • How should the pursuit of higher profits by AI systems be constrained by risks to values that may undermine user trust?

  • How do (and should) AI systems shape our attitudes, beliefs, personality and perceptions?

  • How are labour markets likely to be affected by AI and at what cost?

  • What are the implications of AI systems for developing countries?

The field’s ambition is to identify ways to embed human values into the heart and operation of AI systems, establishing methods to verify their integrity, accountability and resilience thereby ensuring that they, and the data which feeds them, ultimately operate in the service of successful, democratic, digitally empowered yet human-centred communities. This can only be achieved through rigorous, problem-oriented research, which goes hand-in-hand with practice.

Project Lead

Behavioural Data Science is a project by Ganna Pogrebna

ganna-pogrebna.jpg

Ganna Pogrebna

Executive Director, AI and Cyber Futures Institute, Australia

Honourary Professor of Behavioural Business Analytics and Data Science, University of Sydney, Australia

Lead for Behavioural Data Science at the Alan Turing Institute, UK

to stay in touch, join the Behavioural Data Science LinkedIn Group

bottom of page