Skip to main content
SAGE
Search form
  • 00:05

    Developing New Anthropomorphic Learning Modelsto Study Decision Theory

  • 00:09

    GANNA POGREBNA: Hi, I'm Ganna Pogrebna.I'm a professor of behavioral economics and datascience at the University of Birmingham.And I'm also an ESRC Alan Turing Fellowat the Alan Turing Institute.[Ganna POGREBNA, Professor of Behavioral Economics and DataScience, Alan Turing Fellow]My work is interdisciplinary.So I work in decision theory and data science.And so my work kind of gets more and more data science-ish

  • 00:33

    GANNA POGREBNA [continued]: as I go along.Mostly, I'm interested in topics like artificial intelligence,data governance, cyber security, and howwe can use data science to have better servitizationand customization of products.And I'm also interested in business modelsthat are associated with our understanding of personal data.

  • 00:54

    GANNA POGREBNA [continued]: [How can anthropomorphic learning be used,and how is it different from other methods?]My work on anthropomorphic learningessentially tries to merge decision theoryliterature with machine learning literature.So I'm trying to see how decision theory could inform

  • 01:14

    GANNA POGREBNA [continued]: machine learning algorithms to make them betterand to better predict behavior and better understand behavior.Yes, indeed, we currently have a very interesting methodologythat is developed in data science--machine learning, deep learning.So why do we need yet another concept,

  • 01:36

    GANNA POGREBNA [continued]: anthropomorphic learning?So there are several reasons for this.Let's just take a very simple example.So when you go on Amazon and you buy something,normally you see suggestions.And these suggestions say, people wholiked this, also liked this.Or people who bought this item, also bought this item.

  • 02:01

    GANNA POGREBNA [continued]: So these types of suggestion systemsuse what they call collaborative filtering approach.And the way it works is, I would have some itemsthat I bought in the past.And a friend of mine, or someone I don't know, bought items.And if we have some similarities in the items that we bought,

  • 02:25

    GANNA POGREBNA [continued]: then this person would be consideredto be something "like me" or something close to me.And basically, I would be offeredthe objects that this person boughtor a number of people that bought similar things bought.But the problem with this approach

  • 02:46

    GANNA POGREBNA [continued]: is that the assumption is that everything that you buy,you like.And that's not necessarily the case.So this is where decision science becomes extremelyuseful, because in decision science,we have models we call stochastic choice models.Now, let me explain with a simple example

  • 03:09

    GANNA POGREBNA [continued]: how these models work.Imagine that I offer you a choice between tea or coffee.And I do this several times.And imagine that you prefer coffee to tea.So most of the time, you will choose coffee.But every once in a while, you will choose tea.So these models are designed precisely to predict why,

  • 03:33

    GANNA POGREBNA [continued]: on which occasions, in what environment, in what contextyou are choosing tea, even though you might notlike it necessarily.So essentially, the way we could approach this collaborativefiltering model is that we can say that whateveris in your choice set, or whatever

  • 03:55

    GANNA POGREBNA [continued]: you've chosen in the past, is not actually deterministic.It is stochastic.So in a sense, there is noise component in this choice set.For example, it could be somethingthat you don't like but your mother likes or your friendlikes--so things like this.And by looking at historical data,

  • 04:16

    GANNA POGREBNA [continued]: we can actually deconstruct the noise componentand the actual preference component.And that way, we can have a better predictionof what you actually like.Now, if we do it at scale for many different people,we have a much better prediction of what other people likeand what are the actual synergies and compatibility

  • 04:39

    GANNA POGREBNA [continued]: and likeness between people.So that way, we can predict a lot betterwhat people are really in your networkor what people really prefer similar things.So in that way, we can enhance learning algorithms.And this is what they call anthropomorphic learning.

  • 04:59

    GANNA POGREBNA [continued]: So this is one example.There are many other examples.But this is one example of how you can use it.So effectively, it's taking shortcutsin understanding what the preferences are.A lot of people ask, can deep learning do the same?

  • 05:21

    GANNA POGREBNA [continued]: Deep learning is a very powerful tool.And I'm a big fan of deep learning.The advantage of using anthropomorphic learning,in comparison with deep learning,is that anthropomorphic learning cando the same thing but quicker, because we alreadyknow how people are making decisions in many situations.

  • 05:44

    GANNA POGREBNA [continued]: And we can use this knowledge to inform the model.So we do not need to go through all the permutations.So we can do it quicker.We can do it cheaper.And we can construct a model that is a lot more informative.But can learning, in principle, do the same?It probably could.But as a scientist, I'm curious to compare these methods.

  • 06:07

    GANNA POGREBNA [continued]: And this is what I'm working on right now.[How do you combine the decision theory with the machinelearning models?]In economics or psychology, we oftentalk about rational decision maker.And a lot of models assume that people are rational.

  • 06:29

    GANNA POGREBNA [continued]: But recent work in decision theoryshows that people are not necessarily rational.And in fact, they can apply different decision rules.So if we compare our knowledge to this rational individualparadigm, at the moment, we have a lot of alternatives.One of the well-known alternatives

  • 06:51

    GANNA POGREBNA [continued]: is prospect theory, for example.[prospect theory]So prospect theory is not assumingthat people are irrational.This is the usual misconception.So it's just a more general model than the rational model--expected model.But prospect theory will also have more than a dozen

  • 07:13

    GANNA POGREBNA [continued]: different alternatives.And all these alternatives have a deterministic versionand stochastic version.Deterministic version means that if Ioffer you a choice between A and B,I can always predict what you are going to choose--so A or B. And I expect that you will always

  • 07:33

    GANNA POGREBNA [continued]: make the same choice.So you will always choose A, or you will always choose B.[stochastic choice models]A stochastic version means that on some occasions,I predict that you will choose A, and on other occasions,I will predict that you choose B. So these models--like I said, there are a number of them.And we know decision situations in which they work.

  • 07:57

    GANNA POGREBNA [continued]: So for example, we know how a decision setcould be constructed so that peoplechoose a particular thing.This is kind of the classical nudging example--so when we are trying to nudge peopleto make more rational decision or more,

  • 08:18

    GANNA POGREBNA [continued]: say, more green decision.And people who work in nudging use these techniquesall the time.So now all these models are mathematical constructs.And they are based on mathematical rules.So now we can take these mathematical modelsand embed them into machine learning algorithmand allow machine learning algorithm to run at scale.

  • 08:42

    GANNA POGREBNA [continued]: So it can be machine learning algorithm,or it can be deep learning algorithm.It doesn't really matter.So the only thing that matters is that the decision rulethat the algorithm is based on will be essentially comingfrom decision theory.So in terms of machine learning algorithms,

  • 09:06

    GANNA POGREBNA [continued]: you probably know about supervised and unsupervised.So for example, if we take the simplest thing--a supervised algorithm when you can set the categories and thenyou're trying to see to what extent the algorithm canproduce these categories--so what factors predict better?

  • 09:27

    GANNA POGREBNA [continued]: So the decision theory actually allowsyou to set up the rules in such a waythat they will follow a particular decision theoryprediction.Another thing to remember is that a lot of machine learningalgorithms are black boxes.

  • 09:49

    GANNA POGREBNA [continued]: So you know that you can predict a certain eventwith a certain probability--I don't know, 90% or something.But you don't know which factors exactlyare the determining factors.And this is, again, where decision theory could help,because with combination models, you can actually--

  • 10:13

    GANNA POGREBNA [continued]: with these hybrid models--you can actually tell which factors are important factors.One of the first of these hybrid modelswas attribute-based modeling.And these models essentially were testing environments where

  • 10:35

    GANNA POGREBNA [continued]: you had a lot of attributes.So for example, you're making a choice between two cars.A car has thousands, probably, of attributes.And what they found, using these hybrid models,was that you will always weigh a factor on which the two

  • 10:57

    GANNA POGREBNA [continued]: objects are most different more than other factors.So let me give you a specific example, because I was probablyquite technical.Imagine that you have two cars that have a lot of attributes.But on all of these attributes, these two carsare very, very similar.But the one car is red, and the other one is blue.It's very likely that your decision

  • 11:19

    GANNA POGREBNA [continued]: will hinge on just the color of the car,because this is the attribute where these two objects aremost different.So this is kind of an example of early methodology.Now, we have more deeper methodology.And so we have more results, particularly using

  • 11:41

    GANNA POGREBNA [continued]: complex systems and using these collaborative filteringmodels that I was talking to you about.We have conducted some experiments.So one of the experiments, we conducted in Singapore.And it was in the context of a shopping center--so shopping mall.And within the shopping mall, we have different types

  • 12:01

    GANNA POGREBNA [continued]: of customers.And what we do is, we give them a Tinder-like app,except this app shows them offers from different shopsin the supermarket.And so we split these people into three groups.One group just gets the usual collaborative filtering

  • 12:23

    GANNA POGREBNA [continued]: procedure in which they get these offers.The other groups just is a baseline,so they just get the random assignment.And then the third group gets this anthropomorphic learningtype of model.

  • 12:45

    GANNA POGREBNA [continued]: So effectively, what we find is that when we actuallyask people how helpful they find different offers--so when we use a standard collaborative filtering model,only 20% to 30% find it helpful.But when we use anthropomorphic learning,we can get satisfaction rate of up to 85%,

  • 13:09

    GANNA POGREBNA [continued]: because people find this information more relevant.So that's a specific example of how these models couldpotentially be useful.[What are the limitations with this model?]One of the advantages of using anthropomorphic learning

  • 13:30

    GANNA POGREBNA [continued]: is that you don't need a very large learning dataset.So if you are talking about deep learning algorithms,they need a lot of data to learn.Anthropomorphic learning algorithmsdo not need that much data.However, in order to understand the limitations

  • 13:52

    GANNA POGREBNA [continued]: of these algorithms, we are stillin very early stages of developing them.So to me, the first task would beto really test these algorithms against, say, deep learningalgorithms or machine algorithms and understand to what extentand in which contexts these algorithms arereally beneficial.

  • 14:12

    GANNA POGREBNA [continued]: So we see that in some context of consumer choice,they are really interesting and beneficial.But of course, there could be contexts in which deep learningalgorithms do a lot better.So in a sense--so the collaborative filtering approach is only one example.

  • 14:33

    GANNA POGREBNA [continued]: So there are a lot of other versions and variants of this.So to me, my task for the next three yearsis to understand to what extent these models are useful.I think they are useful definitely in some contexts.I'd like to know more about how they compare with deep learningalgorithms.

  • 14:53

    GANNA POGREBNA [continued]: And this is what I'm going to do.[Have you used any traditional methods in developing this newmodel?]I think in order to understand how valuable these algorithmsare, experimental [INAUDIBLE] is actually not bad,even though it relies on relatively small samples.

  • 15:16

    GANNA POGREBNA [continued]: It would be, of course, good to deploy these typesof algorithms at larger scale.But then the problem with non-experimental methodologyis, by applying algorithm, you will change behavior.So in any case, it would have to be an experimental setting,even if it's an experiment at a large scale.

  • 15:37

    GANNA POGREBNA [continued]: So you would need to measure behavior before and measurebehavior after, or satisfaction before and satisfaction after,because if you just start to kindof deploy these algorithms in the field,you will see a lot of adaptation.And maybe what we might see--I'm just speculating here--

  • 15:57

    GANNA POGREBNA [continued]: is that people might have high expectations of, say,predictive models and suggestion systems than they currentlyhave.For example, at the moment, I have yetto find the person who told me that, say, Amazon suggestionsystem really gives them something that they like.

  • 16:21

    GANNA POGREBNA [continued]: I'm sure there are these people.I just haven't met them.But if we deploy, say, anthropomorphic learningalgorithm and they use it, say, on Amazon--just, this is a suggestion and complete speculation--they use it on Amazon system, then

  • 16:42

    GANNA POGREBNA [continued]: people might expect that it should always be accurate.And then it could actually damage the brand.So you have to be very careful with practical applications.So I think, for me, at the moment,the experimental application is a good one.And I'd like to see how far it could go.With practice, I have to be very careful,because you're talking about actual business

  • 17:04

    GANNA POGREBNA [continued]: reputation at stake.[Have there been any challenges in developing this model?]So yes, there is basically a steep learning curvewith trying to learn all the existing algorithms,because algorithms appear every day.

  • 17:25

    GANNA POGREBNA [continued]: We have Google that regularly publishes new things.And we also have DeepMind, who iskind of at the forefront of this research.And they constantly publish interesting new algorithms.To me, it's all about trying to learn this informationand trying to distill it and see to what extent

  • 17:45

    GANNA POGREBNA [continued]: this new method can contribute to existing algorithms.But it's mostly information availability and availabilityof new algorithms that is a limiting factor,because you don't want to miss somethingthat is really, really good and potentially couldbe better performing than the models that I'm testing.

  • 18:07

    GANNA POGREBNA [continued]: [What are the benefits of applying social scienceto computer and data science?]Definitely the biggest problem at this stageis that a lot of people do research on behavior,but in different fields.People do research on behavior in data science.

  • 18:29

    GANNA POGREBNA [continued]: People do research on behavior in engineering.People do research on behavior in mathematics.And all these people lie in silos.So social science can contribute.It can provide the link to combine these silos into oneand potentially have--

  • 18:51

    GANNA POGREBNA [continued]: as a method synergy, it can provide a better understandingof behavior.In terms of other sciences-- so for example,I now have a PhD student who is a physicist by training.And so we are working on synergiesbetween physics and data science--

  • 19:13

    GANNA POGREBNA [continued]: so also very exciting.And so we can go beyond social science and data science.And so it's not like we're trying to developa theory of everything.But we're trying to see to what extentthe foundations of physics and some of the fundamental science

  • 19:35

    GANNA POGREBNA [continued]: could contribute to this and to what extentwe can derive better.[What advice would you give to students?]We're talking about a potential student and what adviceI could give them, I just say, try to talkto a person not in your field--

  • 19:56

    GANNA POGREBNA [continued]: someone you're not even thinking about.I can give you a specific example.So now we're doing a project on Banksy.We're trying to understand whether he sets the trendsor he just reflects the trends.So the way we do it is we obtained all his artwork

  • 20:17

    GANNA POGREBNA [continued]: that we could actually get our hands on.And then we're trying to see to what extentthe keywords that are associated with this artworkwere already existing in the mediabefore the work was published and after.So this project came out from my conversations

  • 20:38

    GANNA POGREBNA [continued]: with philosophers.So people who are in qualitative field, they don't do--they have completely different methodology.But it's a very interesting social science slash datascience project.And I think it's very, very importantto talk to a person who is not in your field.And it's very likely that this person

  • 20:59

    GANNA POGREBNA [continued]: will be thinking outside the box of usingall the imagination they have.And they will be asking really interesting question.So I think that is very important.And by this ability to listen, youcan actually combine different types of fields.And you can generate new insights

  • 21:21

    GANNA POGREBNA [continued]: for practice and for research.[Conclusion]Data science and the decision sciencehave interesting synergies.And people should not--people who are maybe young researchers, early career,

  • 21:41

    GANNA POGREBNA [continued]: or PhD students--they should not be afraid of trying thingsin the other field.As decision scientists, maybe some peoplethink that the grass is greener on the other side.Maybe data scientists have better insights, better data.But I think what is important here is to think creatively

  • 22:05

    GANNA POGREBNA [continued]: about the methodology, because when you have a data scienceproject, you think, oh, if I have great data,then I have great project.Well, we need to think from a different perspective on thisand see, what can we do with available data, which maybeisn't perfect, but how can we develop the methods thatcan tell us more about this limited data that we have?

Abstract

Ganna Pogrebna, PhD, Professor of Behavioral Economics and Data Science at the University of Birmingham and ESRC Alan Turing fellow at the Alan Turing Institute, discusses her research developing anthropomorphic learning models to study decision theory, including ways anthropomorphic learning can be used and how it differs from other models, combining decision theory with machine learning, limitations of the model, traditional methods used in development of the model, challenges faced, benefits of applying social science to computer and data science, and advice for students.

Looks like you do not have access to this content.

Developing New Anthropomorphic Learning Models to Study Decision Theory

Ganna Pogrebna, PhD, Professor of Behavioral Economics and Data Science at the University of Birmingham and ESRC Alan Turing fellow at the Alan Turing Institute, discusses her research developing anthropomorphic learning models to study decision theory, including ways anthropomorphic learning can be used and how it differs from other models, combining decision theory with machine learning, limitations of the model, traditional methods used in development of the model, challenges faced, benefits of applying social science to computer and data science, and advice for students.

Copy and paste the following HTML into your website