Value Driven Data Science: Boost your impact. Earn what you’re worth. Rewrite your career algorithm.

00:00:00 Dr Genevieve Hayes
Hello and welcome to value driven data science, brought to you by Genevieve Hayes Consulting.
00:00:07 Dr Genevieve Hayes
I'm your host doctor Genevieve Hayes and today I'm joined by guest Doctor Fei Huang to talk about fairness and anti discrimination in machine learning.
00:00:17 Dr Genevieve Hayes
Doctor Fei Fang is a senior lecturer in the School of Risk and Actuarial Studies at the University of NSW and has won awards for both her teaching and her research.
00:00:29 Dr Genevieve Hayes
Her main research interest is predictive modelling and data analytics, and she has recently been focusing on insurance discrimination.
00:00:37 Dr Genevieve Hayes
And pricing fairness.
00:00:39 Dr Genevieve Hayes
Faye, welcome to the show.
00:00:42 Dr Fei Huang
Thanks so much, Genevieve. It's definitely my, you know, pleasure to join you and I'm so happy.
00:00:47 Dr Fei Huang
To be here.
00:00:48 Dr Genevieve Hayes
It's great to have you here.
00:00:52 Dr Genevieve Hayes
For our listeners, I first met Faye at the actuary summit earlier this year where she presented a paper on fairness and anti discrimination, which I'll include link to in the show notes.
00:01:05 Dr Genevieve Hayes
And this is an excellent paper and Fey actually won a prize for it and well deserved prize.
00:01:13 Dr Genevieve Hayes
And there are a lot of interesting ideas in this paper, which I think which I thought would be of interest to anyone working with data science and machine learning, not just in the insurance sector.
00:01:24 Dr Genevieve Hayes
And that's why I thought they would be a perfect guest for this podcast. So I'm extremely happy that.
00:01:32 Dr Genevieve Hayes
They accepted my invitation to appear.
00:01:36 Dr Fei Huang
Thanks, Stanley.
00:01:38 Dr Genevieve Hayes
But before we get into the topic properly, how about we start with your background? So Fey, how did you first become interested in modelling fairness and anti discrimination?
00:01:49 Dr Fei Huang
Yeah, that that's a good question actually around. Let me think about maybe two or three years ago I I had some interesting conversations with Professor Jeffries which who is a professor in both Wisconsin, Madison and also Australian national universe.
00:02:05 Dr Fei Huang
City, yeah. Back then we have some conversations, the importance of understanding the, you know, the discrimination, fairness and in particular because we are both educators and we need to, you know, educate future actuaries and actually parts of the syllabus.
00:02:25 Dr Fei Huang
Requirements in both the US and you know UK, Australia around the regulations or understanding the regulation so that the actress need know how to obey the rules and then do the right things when it comes to pricing, underwriting and lots of insurance practise.
00:02:44 Dr Fei Huang
So, uhm, we, we both sort of, you know, agree that this is a very important area and also Jed was affiliated with, you know, or have some, you know, voluntary work in the karate actual society in the United States. And so we sort of think it's a great idea to write a paper on discrimination.
00:03:05 Dr Fei Huang
Actually, the the very first.
00:03:07 Dr Fei Huang
Motivation is from the education perspective. We saw that there is currently a gap about, you know how to what his insurance discrimination, what are the, you know, social, economic considerations there and how to how can future actuaries understand this and also.
00:03:27 Dr Fei Huang
Incorporate them in their business and pricing practise.
00:03:31 Dr Genevieve Hayes
And everything that you've learned about discrimination with pricing, I'm sure it can be applied to a broader area at broader audience, yeah.
00:03:43 Dr Genevieve Hayes
So, so with fairness and discrimination, I mean, we all know what it means for a human to discriminate against another human. I mean, it's probably happened. We've probably been on the receiving end of it.
00:03:56 Dr Genevieve Hayes
And for all.
00:03:56 Dr Genevieve Hayes
Of us.
00:03:57 Dr Genevieve Hayes
But the concept of a predictive model or an artificial intelligence discriminating against us as a person, and that's a relatively new concept. I mean, it's like something out of a science fiction movie, no? With the matrix, with the.
00:04:14 Dr Genevieve Hayes
Machines being the overlords of the humans and using us as batteries. So what does it mean for a model or an AI to discriminate against the person?
00:04:26 Dr Fei Huang
Yeah, I I agree with you that, you know, with the advent of AI and big data, this topic is just getting more and more, you know, attention from both regulators, practitioners and also every.
00:04:37 Dr Fei Huang
Body but I I have to say that the concept itself is actually not new. I mean even go back to that very olden ages wears for example when it comes to just like thinking about insurance pricing as example, even the insurers have very limited data sets to work with and just just several rating.
00:04:57 Dr Fei Huang
Factors. There could also potentially be, you know, discriminate.
00:05:01 Dr Fei Huang
Mission discrimination related issues. For example, there's some, you know, some example of right lining. I'm not sure if you heard of that.
00:05:10 Dr Fei Huang
It's basically, say some banks or insurers or some, you know, financial services, they could, you know, write lines or you know, circle out some geographic.
00:05:22 Dr Fei Huang
Areas that they do not want to ensure, right. So these sort of things actually also highly related to discrimination because those geographic locations could be highly related to some rays or some, you know, ethnicity related protected information. So this is actually not a new topic but with the advent.
00:05:42 Dr Fei Huang
Of and big data. It's just becoming.
00:05:44 Dr Fei Huang
More and more.
00:05:45 Dr Fei Huang
Obvious more and more important, because nowadays people can have access to lots of information. And then and then also some opaque algorithm.
00:05:56 Dr Fei Huang
So for example, in the olden days you just do a regression or a very simple models which is transparent and then you know what is happening there, you can interpret the results. But nowadays machine learning, artificial intelligence, that's actually just means that some very complex models that.
00:06:16 Dr Fei Huang
People find it very hard to explain, you know, explain the outcome or explain why it's generating this outcome. So that's just make the problem more complicated. And then we added the impact of big data.
00:06:29 Dr Fei Huang
Right. So big data just means that you have more information, maybe social media information, maybe your personal habits or your banking details.
00:06:38 Dr Fei Huang
So this sort of a great volume of consumer states available, potentially available to financial services then it, it could also mean that.
00:06:49 Dr Fei Huang
You know and plus the opaque algorithm they all added together just make the problem more pressing and more important to it.
00:06:56 Dr Genevieve Hayes
Yes. So, So what I'm hearing from you is so because we've got more data and these black box models, that's leading to discrimination, which might not have happened in the past when we had less data and more transparent models where people could see if the discrimination was actually occurring.
00:07:18 Dr Fei Huang
Yeah, it's just the more opportunities. So it's it's just the the say firms would have more opportunities. Sometimes they do not even know they are discriminating because it's there could be you know interact or proxy discriminations that it's basically not their intention but it just happened.
00:07:38 Dr Fei Huang
Unfortunately because of the black box models and the.
00:07:41 Dr Fei Huang
Big data.
00:07:42 Dr Genevieve Hayes
So, so how can that happen? How can you end?
00:07:44 Dr Genevieve Hayes
Up with a data scientist building a model that discriminates without them actually realising they.
00:07:51 Dr Genevieve Hayes
Were doing it.
00:07:51 Dr Fei Huang
Yeah, the reason is that different datasets they are associated.
00:07:56 Dr Fei Huang
So when we say discrimination is usually related to some protected attributes, that's specifically regulated in in, in laws.
00:08:05 Dr Fei Huang
Or the axe in different jurisdictions. For example, very commonly used protected attributes are, say, gender.
00:08:13 Dr Fei Huang
Uhm, uhm, raise disability?
00:08:16 Dr Fei Huang
The UM, so this this sort of things are protected information for of people. But you know the variables, the datasets, they are correlated with each other and.
00:08:29 Dr Fei Huang
So when you collect a huge amount of data, for example, maybe females tend to purchase cars with smaller engines and the males tend to purchase cars with bigger engines. And just some simple examples.
00:08:42 Dr Fei Huang
So even if you do not collect the gender you say, I just collect the the size of the engine and want to use that.
00:08:48 Dr Fei Huang
For pricing then you indirectly have disparate impact.
00:08:53 Dr Fei Huang
On the different genders.
00:08:54 Dr Fei Huang
So and then with the bigger data set, you know there are all sorts of potential correlations and associations of those datasets with protected attributes.
00:09:03 Dr Fei Huang
Some of them you do not even know because lots of the firm they do not collect those information to not collect your race or nationality or, you know, disability.
00:09:14 Dr Fei Huang
So they do not correct. So they they actually.
00:09:16 Dr Fei Huang
Are not aware.
00:09:16 Dr Fei Huang
Of that, but it's just, you know, hidden there in the data sets.
00:09:21 Dr Genevieve Hayes
One of the things I read about was that model that Amazon built to sort resumes in order to identify people who were most suited to a particular job.
00:09:34 Dr Genevieve Hayes
Initially they built the model and they found that it was discriminating against women because historically Amazon, in their engineering type roles had hired a lot of men.
00:09:44 Dr Genevieve Hayes
So there was a bias in favour of men, so they deliberately excluded gender from any of the data that went.
00:09:54 Dr Genevieve Hayes
In and it was not included as a rating factor in the model.
00:09:59 Dr Genevieve Hayes
But what they found was that the model then started to learn things that allowed it to discriminate against women and, for example, it started identifying the universities in America that had a very high proportion of women and penalising people who went to those universities.
00:10:19 Dr Genevieve Hayes
And people who had.
00:10:21 Dr Genevieve Hayes
Participated in sports that were commonly participated in by women and they tended to get discriminated against. So it found all these proxies for someone being a woman and discriminated against them, and Amazon was unable to eliminate any discrimination against women and.
00:10:41 Dr Genevieve Hayes
Ultimately they completely shelled the model because they could not find a way to make it non discriminatory.
00:10:48 Dr Fei Huang
Yes, yes, absolutely. I mean this, this, this is a very good example of showing you know the power of the proxy information that could have on the impact of the protected information state gender there.
00:11:00 Dr Fei Huang
Yeah, I think.
00:11:01 Dr Fei Huang
I think this sort of things are happening, you know. If there are lots of applications related to this, then it's related to the top.
00:11:08 Dr Fei Huang
Pickle directly discrimination and indirect discrimination. So if you do not use the information, say gender in your data set, you simply avoid directly discrimination.
00:11:19 Dr Fei Huang
But it's not guaranteed that you also avoid indirect discrimination, actually, as long as the data set you have, say, the universities.
00:11:29 Dr Fei Huang
Or some other attributes that related to the people. As long as they are related to gender, then it's it's basically you know.
00:11:38 Dr Fei Huang
So it's it's almost.
00:11:40 Dr Fei Huang
Guaranteed that you know your outcome would have impact on the genders even if you do not use it.
00:11:45 Dr Fei Huang
So then the problem would be, you know.
00:11:47 Dr Fei Huang
How to mitigate?
00:11:49 Dr Genevieve Hayes
Yeah. So how would you mitigate this?
00:11:52 Dr Fei Huang
Yeah, this is a very complex problem I would say. Actually it's also application.
00:11:58 Dr Fei Huang
Pacific and the jurisdiction specific, line of business specific. So we must you know take them case by case. The reason is that actually the regulation on you know in terms of these interactive.
00:12:13 Dr Fei Huang
Nation can be different for different line of business. For example for some jurisdictions they say what, what what's the you know, what we want to achieve is at the group level fairness. That means say our average that females and males they should get the same.
00:12:33 Dr Fei Huang
Opportunity or they should have. You know they should not have this part. There should not be disparate impact affecting on the two groups on average.
00:12:42 Dr Fei Huang
So this is a group level fairness, but for some other lines of business people would, you know, want to achieve individual level fairness that is subject to the.
00:12:52 Dr Fei Huang
Their other attributes or say for example it's insurance pricing, then it's subject to their risk levels or the risk related rating factors.
00:13:02 Dr Fei Huang
The males and females should get the same prices for the same insurance products. So then that could be individual level fairness and and it's.
00:13:12 Dr Fei Huang
It's very complex and I personally believe it's there are still some vagueness there. In terms of regulation it's still not crystal clear, but especially for some line for business, what is a specific fairness that practitioners should incorporate in their practise?
00:13:31 Dr Genevieve Hayes
So, so I'm trying to get my head around this whole individual versus group fairness. And so I think I understand it in the insurance context, but I'm trying to apply that, think about how that would apply to other contexts. Yeah, so.
00:13:50 Dr Genevieve Hayes
This is my thoughts about translating it into, say, the job recruitment example, like with Amazon, like we're talking about.
00:13:58 Dr Genevieve Hayes
So individual fairness would be two people, a man and a woman, who had identical resumes except for one man and one woman up.
00:14:09 Dr Genevieve Hayes
If it was individually fair, that would mean that those two people would be treated exactly the same is.
00:14:15 Dr Fei Huang
Yes, that's correct. Yeah. Yeah.
00:14:17 Dr Genevieve Hayes
Whereas great fairness would mean that.
00:14:21 Dr Genevieve Hayes
Those two individuals might not necessarily be treated the same, but on average males are treated are just as likely to.
00:14:32 Dr Genevieve Hayes
Get an engineering position at Amazon as females are. Is that?
00:14:37 Dr Fei Huang
Right. Yeah, that's right.
00:14:39 Dr Genevieve Hayes
OK.
00:14:41 Dr Genevieve Hayes
Is it possible for both of those to exist simultaneously? That you're fair at the individual level and your fair at the group level?
00:14:50 Dr Fei Huang
It's possible, but it's it requires some constraints or some assumptions. So, but it's not, it's these two actually in most of the time I would say they do not, you know, consistent with each other. So.
00:15:03 Dr Fei Huang
That means they achieving individual fairness may not, you know, lead to achieving group fairness only under certain scenarios. That these two are you know would be consistent with with with each other.
00:15:16 Dr Fei Huang
That means you can achieve both for example in and also when it comes to individual fairness and all group fairness.
00:15:24 Dr Fei Huang
There are also different fairness criteria there. So for example there are demographic parity, equalised dolls, equal opportunity. So there are different group level fairness as well. And for individual fairness it can also list A, you know, a list of individual fairness. Seiko consulting.
00:15:42 Dr Fei Huang
But it or it's just uh, they proxy avoid proxy discrimination or you know they they they are all we see each category there are also different fairness criteria there. So in terms of you know which fairness criteria are talking about is another complex issue here.
00:16:03 Dr Genevieve Hayes
So you mentioned, I think about half a dozen different fairness criteria is now, before we started this episode, I actually looked up what does fair mean in the dictionary, because it's just such a it's a word that we all have a sense of what fair is in.
00:16:22 Dr Genevieve Hayes
The sense of our own selves. But I just wanted to know what does fair actually mean? And according to the dictionary I looked up, fair means unbiased and I think that creates another set of problems because.
00:16:40 Dr Genevieve Hayes
One of the things I learned when I was studying machine learning is that every machine learning model has to have bias in it, because otherwise it would be just selecting people at random, which doesn't work. So you have to have, you know, in our.
00:16:58 Dr Genevieve Hayes
Job search model for example, we want the model to be biassed in favour of people who are actually good at doing their jobs.
00:17:07 Dr Genevieve Hayes
We just don't want it to be biassed based on protected attributes of people.
00:17:16 Dr Genevieve Hayes
Obviously in the machine learning sense, we can't have fair means completely unbiased, otherwise we'd get useless models.
00:17:25 Dr Genevieve Hayes
So can you give me one or two examples of some of these fairness criteria that you just mentioned? Absolutely.
00:17:35 Dr Fei Huang
I think this is a very good question, a very important one. So actually I would say that the fairness concept as you said, it's really differ.
00:17:45 Dr Fei Huang
You know if if you're talking about different situations.
00:17:49 Dr Fei Huang
I'll give you 2.
00:17:50 Dr Fei Huang
Simple examples 'cause I've been working on interest pricing for so long, I just, you know, just too easy for me to grab the insurance examples.
00:17:58
Let's let's sync up.
00:17:59 Dr Fei Huang
2 examples. One is that say we everyone.
00:18:02 Dr Fei Huang
Purchase car insurance, right?
00:18:04
And then you.
00:18:05 Dr Fei Huang
You would have, you know, face two kinds of car insurance. One is compulsory car insurance that is regulated by the by each jurisdiction you must purchase the, say the third party compulsory insurance.
00:18:19 Dr Fei Huang
Otherwise, a voluntary insurance that you can choose to purchase or not purchase. And you can choose comprehensive cover or some you know some other covers, right? So you can choose.
00:18:29 Dr Fei Huang
The coverage so like these.
00:18:31 Dr Fei Huang
Two type of car insurance they the fairness concepts are different. So for the compulsory third party insurance for example in in in a CT Australia it's community rating that means if you purchase the compulsory third party insurance from.
00:18:50 Dr Fei Huang
Ensure they gave.
00:18:50 Dr Fei Huang
You exactly the same price. They do not even think about your risk profile, your gender, your age, your car type.
00:18:58 Dr Fei Huang
Nothing. Everyone got the same price from the same insurer that's based in a city and then in say NSW is highly regulated.
00:19:07 Dr Fei Huang
It does not. Everyone may be may get different prices, but insurers are highly regulated to use only specific rating factors in order to do the pricing for your compulsory current.
00:19:20 Dr Fei Huang
So this is compulsory one. But when it comes to voluntary car insurance then there's much less regulation. That means the insurers can use a whole bunch of rating factors to do the pricing.
00:19:30 Dr Fei Huang
And for example in asked if you purchase the voluntary car insurance then you need to submit say your age, gender, card type, where you, where you, where is the location of your car.
00:19:41 Dr Fei Huang
We do leave right this sort of information should be provided because they are the rating factor that.
00:19:46 Dr Fei Huang
Ensures used to pray.
00:19:47 Dr Fei Huang
Your car insurance. So as you can see these two very simple example already showed you the different notions of fairness, why we are having like these different regulations.
00:19:58 Dr Fei Huang
And if I look.
00:19:59 Dr Fei Huang
You know, more specifically into these two examples, I can say that in the first example, the compulsory third party insurance.
00:20:07 Dr Fei Huang
In a city community rating, I can classified into group various because I can prove that it's actually satisfy the demographic parity group fairness.
00:20:16 Dr Fei Huang
But for the voluntary 1.
00:20:17 Dr Fei Huang
I would tend to classify it into individual fairness 'cause they are using individual rating factors and people are getting different prices subject to risk.
00:20:26 Dr Fei Huang
Files. So the reason of the difference is that these two line of business or these two different products are serving for different needs of the society for the compulsory third party.
00:20:39 Dr Fei Huang
We can treat that as a social safety net because everybody have to have access to compulsory car insurance.
00:20:46 Dr Fei Huang
As a protection to them, yeah, because this is the government or the society want to give everybody such a protection.
00:20:54 Dr Fei Huang
So when then we tend to allow for cross subsidy, right. And then so we tend to give everyone the same price so that everyone can have access.
00:21:03 Dr Fei Huang
To the to the product and also get the protection.
00:21:06 Dr Fei Huang
But for voluntary.
00:21:08 Dr Fei Huang
Then it's it's more for economic purposes.
00:21:11 Dr Fei Huang
So we allow for the insurers to to classify their policyholders give subject to the different risk profiles. This is more of a consideration of economic.
00:21:23 Dr Fei Huang
Efficiency because we want, because if you allow for this then it could potentially avoid adverse selection.
00:21:31 Dr Fei Huang
Moral hazard, this sort of economic related issues, it would, you know, help with the economy and also, you know, achieve more economic efficiency. So this is because of the different considerations and it's it's actually acceptable.
00:21:48 Dr Fei Huang
To the society as well.
00:21:50 Dr Fei Huang
So and also another example say health insurance in Australia is also community rating because.
00:21:55 Dr Fei Huang
It's it's it's more serve as a social safety net or social good we can treat that as a.
00:22:00 Dr Fei Huang
Social goods but.
00:22:01 Dr Fei Huang
For some, insurance is more, uh, economic product or economic good.
00:22:04 Dr Fei Huang
So there's sometimes it could also some somewhere in between, right? So these different considerations lead to the different notion of fairness.
00:22:14 Dr Genevieve Hayes
And I assume that which definition you choose is going to depend on what's considered to be socially acceptable in that particular set of circumstances, yes.
00:22:27 Dr Genevieve Hayes
So what I'm thinking, I was trying to translate your example into a non insurance example. If we just think about going down the supermarket, if you go down the supermarket and buy a can of coke.
00:22:41 Dr Genevieve Hayes
Every single person who buys that can of coke will be charged exactly the same amount, and we would consider it to be.
00:22:48 Dr Genevieve Hayes
Unfair if we charged one group of people more for the count of Cokes than another group of people in that same supermarket.
00:22:57 Dr Genevieve Hayes
But if you go to the movies, they have all different ticket prices. You know, they have the seniors price, the child price, the adults price, students price, all of those.
00:23:09 Dr Genevieve Hayes
So there's discrimination going on there. We're discriminating based on which group of people you belong to, but we consider it to be fair to charge.
00:23:18 Dr Genevieve Hayes
Some groups more than others, because some are considered to be more able to afford.
00:23:24 Dr Genevieve Hayes
Got it. And in that set of circumstances, society is clearly deemed at fair to charge adults more than children, for example.
00:23:34 Dr Fei Huang
Yeah, yeah. Actually in the economics, there are three types of price discrimination, their first degree, second degree and 3rd degree.
00:23:44 Dr Fei Huang
Actually, you know, charging a different price to different consumer groups belong to the third degree price discrimination.
00:23:51 Dr Genevieve Hayes
This is actually.
00:23:51 Dr Genevieve Hayes
Reminded me of a cartoon that I saw which I think did the rounds of the Internet a while ago.
00:23:56 Dr Genevieve Hayes
Yeah. And it's an illustration of equity versus equality. Have you seen it's the one with the three children?
00:24:04 Dr Genevieve Hayes
No. OK, so it's got three children and there are each of different heights, and they're trying to watch a baseball game over a fence, and only the tallest child is tall enough to actually see over the fence without any assistance.
00:24:20 Dr Genevieve Hayes
And one of the panels in the cartoon is labelled equality, and in that panel all of the children are given a single box to stand on.
00:24:31 Dr Genevieve Hayes
So that would be treating all three children equally, but in that scenario, the tallest child, well, they could see over the fence already without the box.
00:24:40 Dr Genevieve Hayes
The middle child can now see over the fence, but the smallest child still can't see over the fence.
00:24:48 Dr Genevieve Hayes
And then in the second panel, you've got equity, and in that case the shortest child is given 2 boxes, the middle child given one box, and the tallest child doesn't need a box.
00:25:01 Dr Genevieve Hayes
And then all of the three children can see over the fence.
00:25:06 Dr Genevieve Hayes
And I think what that's trying to illustrate.
00:25:11 Dr Genevieve Hayes
You can either treat everyone equally, but that doesn't necessarily lead to an equal outcome, or you can try and aim for an equal outcome, which might not necessarily be the result of equal inputs, and so that would be two different versions of.
00:25:31 Dr Genevieve Hayes
Dennis and.
00:25:33 Dr Genevieve Hayes
Each of those scenarios could be considered fair or unfair depending on the way you.
00:25:37 Dr Genevieve Hayes
Look at.
00:25:38 Dr Genevieve Hayes
It. So I would imagine that that's the sort of thing that people are looking at when they're evaluating the fairness of various pricing models and predictive models that right.
00:25:50 Dr Fei Huang
Yeah, absolutely. Yeah, absolutely. Like ECHO I I think equity is sort of related to equal treatment and then equality related to say equal outcome.
00:26:00 Dr Fei Huang
Right. And then actually in the United States, there are two similar legal concepts. One is called disparate treatment, the other is called disparate impact.
00:26:11 Dr Fei Huang
And so this further treatment.
00:26:13 Dr Fei Huang
Is sort of related to.
00:26:14 Dr Fei Huang
Like fairness through an awareness. For example if you don't, do not use that and then you do not use the protected attributes for example.
00:26:21 Dr Fei Huang
This sort of.
00:26:21 Dr Fei Huang
Belief that you are treating everyone you know.
00:26:24 Dr Fei Huang
It's it's like uh achieve that it this part, there's no disparate treatment.
00:26:30 Dr Fei Huang
But there could still lead to this part you know outcome or disparate impact on the on the outcome. So and then people would say maybe some more.
00:26:40 Dr Fei Huang
Uhm. Fairness criteria or more mitigation efforts should be made in order to achieve, you know, equality in the end and.
00:26:50 Dr Fei Huang
Again, it will link to the different, you know, specific application or specific scenario and see what is more acceptable to the society.
00:27:00 Dr Genevieve Hayes
So which fairness criteria are most commonly used in practise?
00:27:05 Dr Fei Huang
In practise it it also depends on which long business, but I would say fairness through an awareness is most commonly used this basically I do not collect any protected attributes and then but but people currently I think most people are well understood that it cannot achieve like.
00:27:25 Dr Fei Huang
Disparte achieve equal outcome, for example. At least the subject to some fairness criteria, yeah.
00:27:34 Dr Genevieve Hayes
So you need to have.
00:27:36 Dr Genevieve Hayes
If if you want to achieve a clap comes then you need to be aware of the fact that people are starting from different starting points and make concessions for people who are starting behind like the short kid and assist them so you can get those equal outcomes.
00:27:56 Dr Fei Huang
Yeah, yeah, I I think that's, that's more like it again for for some line of business especially I think this is especially related to the lab business that we tend to treat them as social good.
00:28:08 Dr Fei Huang
Especially for those.
00:28:09 Dr Fei Huang
Lines we treat my social good, then we should aim for like, you know, equal outcome, right? You know, in addition to, you know, not equal treatment, but more equal outcome.
00:28:19 Dr Fei Huang
Uh, another thing is that currently people are talking about, say, a climate related.
00:28:24 Dr Fei Huang
Uhm, you know, impacts on people with Steve, you know, living in different areas. That's also sort of related examples that, you know, people could think about what what's the real outcome that we want to achieve?
00:28:38 Dr Genevieve Hayes
So can you give me an example of those climate related examples?
00:28:42 Dr Fei Huang
Yeah, I think recently there's the the government have a reinsurance cyclone reinsurance pool and and also recently there are some you know reports things from affinity consulting on the.
00:28:56 Dr Fei Huang
Come on the you know the social, economic, you know, social economic differences related to the home insurance.
00:29:05 Dr Fei Huang
I think, I think that's, I mean for for example people living in the areas that's that's more vulnerable related to climate change or you know the say.
00:29:17 Dr Fei Huang
Climate related disasters then they are facing a higher premiums right and and then some for some areas could even do not have access to insurance and then I don't what's what should we do right and then what is.
00:29:32 Dr Fei Huang
So what's the fairness should be applied in these kinds of scenarios? I think it it's a really interesting problem.
00:29:39 Dr Genevieve Hayes
And I think that came up with the Lismore floods in Australia last. Was it last year or at the start of this year where there were some people who were uninsured because even before the floods occurred, the insurance premiums were 10s of thousands of dollars per year. Yeah. And you know, is it fair for them to have to pay higher premiums?
00:30:01 Dr Genevieve Hayes
Or is it fair?
00:30:04 Dr Genevieve Hayes
For them to be cross subsidised by people who are in safer, drier areas.
00:30:09 Dr Fei Huang
Yeah, yeah.
00:30:10 Dr Fei Huang
Yeah, that that's an interesting problem to think about. Actually, one of my research recently is just looking at the impact of the different fairness criteria from the economic perspective. So what's the, what's the impact on both consumer welfare and the firm's profits?
00:30:29 Dr Fei Huang
If we apply different fairness criteria, so that could be a, you know, an interesting thing that for the regulators or the, you know the general society to to to think about.
00:30:40 Dr Genevieve Hayes
And and it's interesting because we know from different areas where they have tried to make concessions to adjust for historic discrimination against particular groups, you know, sometimes.
00:30:56 Dr Genevieve Hayes
They're treated fair and no one seems to have any problems with them. Like, uh, no one has an issue with wheelchair access to buildings.
00:31:05 Dr Genevieve Hayes
Whereas other times, like gender quotas, gender quotas end up with all sorts of debate. Some people are in favour of them, some people are against them. I'm not opening that can of worms here.
00:31:19 Dr Fei Huang
Yeah, yeah, absolutely. Yeah.
00:31:21 Dr Genevieve Hayes
Yeah, So what? What we've said here is that it's possible for a organisation to end up accidentally discriminating against someone.
00:31:30 Dr Genevieve Hayes
Or without intending to do it that the the best the way people try and and.
00:31:38 Dr Genevieve Hayes
And deal with fairness is by trying to be blind to particular protected attributes, but even doing that they can end up accidentally discriminating due to proxies.
00:31:54 Dr Genevieve Hayes
So these well meaning organisations could inadvertently end up breaching anti discrimination laws that exist in particular jurisdictions.
00:32:06 Dr Genevieve Hayes
Has there any ever been a case where an organisation has been prosecuted for breaching anti discrimination laws when they were deliberately taking action to try and avoid discrimination?
00:32:21 Dr Fei Huang
I think there currently this is still an emerging area and the regulators are also trying to provide more clear guidance on what, what, what's the, you know, what the firm should do.
00:32:34 Dr Fei Huang
So I think as far as I know in Australia there is, it's, it's it's basically very, very limited examples in terms of the.
00:32:42 Dr Fei Huang
In terms of indirect discrimination related?
00:32:46 Dr Fei Huang
You know, breaching the laws.
00:32:47 Dr Fei Huang
Uhm, uhm. And but for the United States, I think it's it's always under a heat debates. I think the case laws are also there.
00:32:57 Dr Fei Huang
I think they are still very limited in terms of indirect discrimination because I'm I'm mainly focusing on the insurance lines. So I'm, I'm, I'm, I'm saying that for the insurance related indirect discrimination.
00:33:09 Dr Fei Huang
Cases there's very limited that currently we could find.
00:33:13 Dr Fei Huang
But it's definitely under heat discussions and and then the regular hurts are providing trying to provide more clear guidance in terms of of what should we do and what should not.
00:33:26 Dr Fei Huang
Be done, yeah.
00:33:27 Dr Genevieve Hayes
I think one of the interesting things in the United States is they don't just have the laws protecting people who have protected at.
00:33:36 Dr Genevieve Hayes
Tributes in insurance. They also have those laws preventing price optimization in some states, which is basically discrimination against people based on non protected attributes. For example this person in a wealthier area than this other person. So let's charge the richer.
00:33:56 Dr Genevieve Hayes
Person more.
00:33:57 Dr Fei Huang
Yes, yes, I think you raised a very interesting example here. So usually, generally speaking protected attributes are, you know, for every last business, they have to be very careful with you know either direct or indirect discriminating against those protected attributes. But, but.
00:34:19 Dr Fei Huang
For insurance it's a little bit different where there are other considerations, especially if you use non risk based pricing factor.
00:34:28 Dr Fei Huang
Price optimization is, is is something I think is commonly used in some markets including Australia I think. I think Europe is also using a press optimization a lot, but US is I think starting from 2015 up to now is around 20 states already.
00:34:48 Dr Fei Huang
And press optimization in the United States for the insurance pricing and underwriting, the main reason is that they believe the price, you know the price optimization or those non risk based actually.
00:35:00 Dr Fei Huang
It's it's not suitable to be used for insurance purposes, for example, people who are more willing to change.
00:35:10 Dr Fei Huang
Their insurers, or more waiting to shop around, get lower prices and more loyal customers get higher prices. And in the UK, historically, there's a price.
00:35:21 Dr Fei Huang
Working so the.
00:35:22 Dr Fei Huang
Longer you stay with, ensure the higher prices you get.
00:35:25 Dr Fei Huang
So it's it's like every year there's a price working. I'm starting from earlier this year, UK band price working.
00:35:32 Dr Fei Huang
So this is in similar I think. I think the idea is similar to the price optimization banned in the United States which is all bending the non risk based factors.
00:35:43 Dr Fei Huang
Yeah but it it's it's it's also something that it's it's not every not every jurisdiction are doing this but it's definitely these two discussions in in different markets including Australia here.
00:35:55 Dr Fei Huang
As well, yeah.
00:35:56 Dr Genevieve Hayes
Do you think that in Australia will one day have laws preventing price optimise?
00:36:00 Dr Genevieve Hayes
Session. I don't know.
00:36:02 Dr Fei Huang
Yeah, it's, it's very difficult to predict, but it's it's definitely something that people could think about.
00:36:09 Dr Fei Huang
Actually, in my current research I'm trying to looking at the the impact of this balance. So what's the real welfare cost?
00:36:21 Dr Fei Huang
Of these different regulations, uhm, I think I I I can provide you with the link of the paper after this interview.
00:36:29 Dr Fei Huang
So it's actually, I mean for lots of the regulations, it's fairness laws rather than fairness skin, I mean based on my simple model.
00:36:37 Dr Fei Huang
So it's I mean of course when regulators having those you know those the regulation into considerations they they are usually not just from the economic perspective from the welfare gain or welfare loss sometimes they think more of like the vulnerable vulnerable groups or people with.
00:36:57 Dr Fei Huang
Low social economic status. So that's another perspective that.
00:37:03 Dr Fei Huang
Why there are these type of regulations and also say accessibility to insurance. So I mean I think what I, I, I want to say is that more research is needed in this area to understanding the real impact of each of these regulations. They either press working band.
00:37:23 Dr Fei Huang
Or press optimization ban so that everybody is more informed what, what will happen if we have this band and idea? What are the options to protect the consumers where we're at the same time the firm?
00:37:38 Dr Fei Huang
I also you know gaining profit and also you know not highly impact on the on the sustainability, yeah.
00:37:48 Dr Genevieve Hayes
So you mentioned before fairness losses and fairness gains. So what you're saying is that in some circumstances by banning discriminatory actions?
00:37:59 Dr Genevieve Hayes
Either discriminatory based on protected attributes or.
00:38:02 Dr Genevieve Hayes
On non protected attributes that can actually make everyone worse off.
00:38:07 Dr Fei Huang
That scan, yes, it all of course it's, it's that's it depends on which perspective are you looking at, right.
00:38:15 Dr Fei Huang
But from at least from the economic welfare perspective it can list, everyone works. Well, it's possible under certain.
00:38:24 Dr Genevieve Hayes
Very interesting. What can you give me an example of one of the scenarios where does leave it leave everyone worse off?
00:38:31 Dr Fei Huang
Uhm, I think we're, I think it's probably easier if we, you know when, when, when the final paper is out where we have more rigorous studies that you also understand the assumptions.
00:38:42
OK.
00:38:43 Dr Fei Huang
Yeah, of all these circumstances. But I, I, I can. I can clearly see that for many of the combinations, for example.
00:38:52 Dr Fei Huang
If there are cost, there's a regulation.
00:38:55 Dr Fei Huang
Banning the cost, say when insurers do cost modelling the technical prices, if there are regulations there. As a group we mainly focus on the group level fairness there and or if there are regulations on price optimization on the final prices or regulations on achieving demographic.
00:39:15 Dr Fei Huang
Parity the group.
00:39:16 Dr Fei Huang
Level fairness on the final prices, or if there is a combination of both, banning the core, estimating the costs and estimating the prices. So we look.
00:39:27 Dr Fei Huang
At all sorts of.
00:39:28 Dr Fei Huang
Regulations and we we employed a simple economic model.
00:39:32 Dr Fei Huang
To to understand the welfare impact on both consumers and firms. Now we can see that lots of this combination leads to worse outcomes.
00:39:42 Dr Genevieve Hayes
It's very interesting if you if you send through the link, I'll include that in the show notes 'cause I'm sure a lot of people will.
00:39:48 Dr Fei Huang
Be interested in reading that. Yeah, but I have. And that again, I wanna, I wanna, you know, remind everyone.
00:39:54 Dr Fei Huang
And that that is from the economic welfare perspective. So because we do not have the data sets to understand the profiles linked to the consumers vulnerability or socioeconomic status, so we basically do not understand maybe, maybe there will be welfare again under vulnerable people from this perspective it's it's worth well.
00:40:15 Dr Fei Huang
It's better off, but from the purely the the data that we have access to and from purely the economic welfare perspective, it could lead these two words of scenarios, yeah.
00:40:26 Dr Genevieve Hayes
That's very interesting.
00:40:29 Dr Genevieve Hayes
So a lot of the points we've raised in this discussion, I can see potentially putting organisations off using machine learning models, or if they're already using them, making them think twice about using them, because basically what we've told people is that even if they're trying very hard not to, they can end up.
00:40:49 Dr Genevieve Hayes
Discriminating against people? Yeah. And even though this might not 'cause issues with anti discrimination laws now because it's sound.
00:40:59 Dr Genevieve Hayes
It's like they don't fully catered to AI and machine learning just yet, you know in the future this could lead lead to them breaching some future law.
00:41:11 Dr Genevieve Hayes
What would you say to an organisation that was thinking of going down the path to machine learning to try and reassure them that?
00:41:20 Dr Genevieve Hayes
You know, there are benefits that outweigh these risks of discriminatory activity.
00:41:27 Dr Fei Huang
Yeah, I think that's a very important question that I I believe every organisation should, should carefully think about. For example, if you are considering whether to use some additional data sets or if you want to implement a black box of very complex machine learning models, think about the risks versus the return.
00:41:47 Dr Fei Huang
Updates and also I think it's it's very important to understand the regulations in the jurisdiction that you have business.
00:41:55 Dr Fei Huang
For example, if it's, if it's an international firm, have business in different jurisdictions, then even just in Australia, if you have business in different states, I mean the product attributes you're facing can be slightly different.
00:42:08 Dr Fei Huang
So I mean.
00:42:09 Dr Fei Huang
I think that's something that the organisation have to be aware of. Another thing I want remind is that, as you said, mitigating direct discrimination does not necessarily mean mitigating indirect discrimination.
00:42:21 Dr Fei Huang
So also organisations should think about it there, you know, strong proxies there and how to mitigate indirect discrimination.
00:42:29 Dr Fei Huang
So although this is still a little bit vague area in terms of regulation guidance, but I think the regulators are making efforts in providing more guidance here so.
00:42:40 Dr Fei Huang
I think keep updates to the latest information in terms of what can be done and what cannot be done and then the final advice they wanna give is that actually.
00:42:53 Dr Fei Huang
Sometimes people tend to look, you know, too much on the say, the marginal gain in terms of prediction accuracy.
00:43:01 Dr Fei Huang
But tend to forget about what's the real return on your company. But for example, if I employ a machine learning black box model, I probably improve my prediction accuracy by say 12% and and then that means lots of efforts in this complex you lose the transparency and then we.
00:43:21 Dr Fei Huang
When you look at the real impact on the firm, say the profits or what's the 5% prediction accuracy really means for your, for, let's say for the profits of the firm's or the, the value of the firm's or the, the volume of the, the volume of the, say the.
00:43:39 Dr Fei Huang
The products that you could have then it could be minimum or even, you know, neglectable. So I think it's it's important to understand the real value of either getting a big data set or using a more complex algorithm and then also think about what you lost, you lost the transparency.
00:44:00 Dr Fei Huang
Or you lost the interpretability and and then something else. So I'm I'm not against black box models, but I just thought it had to be all taken into consideration.
00:44:10 Dr Genevieve Hayes
And if if you're using a black box model leads to either a massive fine or to a massive scandal that leads to a loss of business then.
00:44:20 Dr Genevieve Hayes
Any benefits you gain from that black box model are going to be eliminated.
00:44:24 Dr Fei Huang
Yeah, so I think that's something I think academics should work together with industries to make those so to to help them understand better the real benefits of the different, more fancy mode.
00:44:36 Dr Fei Huang
Those and also try to make the models more interpretable at the same time give you more prediction accuracy. So I think there's lots of, you know, space or gaps that could be filled in the future. I think lots of research needs to be done.
00:44:51 Dr Genevieve Hayes
So the key points I'm getting here that our businesses should take away our don't be scared of using machine learning models.
00:44:59 Dr Genevieve Hayes
However, if you're going to do it, go down that path and go in with your eyes.
00:45:03 Dr Genevieve Hayes
Open. Yeah. So yeah, consider different model.
00:45:08 Dr Genevieve Hayes
Rules not just about for what? Which model gives you the most accuracy, but which model gives you the best results? Yep, which is a tradeoff between accuracy and transparency.
00:45:21 Dr Fei Huang
Yeah. And also the with also the regulation risk, right? Uh, discrimination risk, fairness risk, all this into consideration.
00:45:28 Dr Genevieve Hayes
And generally try and be a good corporate citizen.
00:45:31 Dr Genevieve Hayes
Yeah. So, so if you're taking steps to demonstrate that you're actively trying to avoid being discriminatory and I would imagine that even if something went wrong, if you could say, well, we've deliberately not collected these protected attributes and we have done all these tests to ensure we're being fair.
00:45:51 Dr Genevieve Hayes
The individual or group level, then you could make a pretty fair case if you ever got.
00:45:59 Dr Genevieve Hayes
If you ever did have something, go.
00:46:01 Dr Fei Huang
Yes, actually in my class. So I teach data analytics at USW in my class. I in my first class, I always tell students have technology enhance humanity in mind.
00:46:12 Dr Fei Huang
So you have to.
00:46:13 Dr Genevieve Hayes
Have both, yeah. And. And that way you don't end up with a situation like Terminator where you end up with us versus the machines.
00:46:22 Dr Genevieve Hayes
Yeah, yeah.
00:46:25 Dr Genevieve Hayes
OK. So we're getting pretty close to the end. So, but I have a few final questions for.
00:46:30
Are you?
00:46:32 Dr Genevieve Hayes
Is there anything on your radar, in the AI, data and analytics space that you think is going to become important in the next three to five years?
00:46:41 Dr Fei Huang
Yeah, I I think still ethical and responsible use of AI and big data is still something that I I believe will attract more and more attention.
00:46:52 Dr Fei Huang
And then Miss Lawson means lots of work to be done in this area. So I think this definitely are trained in the next few years.
00:47:00 Dr Fei Huang
Another thing is related to this climate change towards net zero, right? So some big problems that the whole society or the whole world trying to resolve. It's also linked to the Echo and.
00:47:12 Dr Fei Huang
Responsible use of AI and and also.
00:47:16 Dr Fei Huang
Say that when when I say ethical is actually includes, but not limited to discrimination or fairness related thing. It's also about transparency, accountability, interpretability for example.
00:47:29 Dr Fei Huang
So it's it's a whole package of SD code, use of AI and big data, and in Australia there's.
00:47:36 Dr Fei Huang
Voluntary ethical framework that organisation could refer to come and they are. They also provide some examples case studies about some some other, you know, pioneers who how they how they adopted the ethical framework in their business.
00:47:52 Dr Genevieve Hayes
That framework sounds like it could be of great use to our listeners. Do you have a link that you?
00:47:57 Dr Genevieve Hayes
Could send me for that?
00:47:58 Dr Genevieve Hayes
Yeah, yeah, yeah. If you send me that, I'll include.
00:48:00 Dr Genevieve Hayes
That in the show notes, yeah.
00:48:02 Dr Genevieve Hayes
And what final advice would you give to organisations looking to maximise the value of their data through either fairness and anti discrimination or through anything else?
00:48:15 Dr Fei Huang
Yeah, my final devices probably I think it's it's still like have a good understanding of your data, your model and ask yourself whether you really need that, what's a real impact on your consumers and on your firms and also take all those regulation into consideration.
00:48:35 Dr Fei Huang
Potential regulation risk and also have a good understanding of the requirements of anti discrimination regulations and laws in the specific jurisdictions.
00:48:45 Dr Genevieve Hayes
So be a good person and act responsibly, basically, yeah, yeah.
00:48:50 Dr Genevieve Hayes
OK, this has been great having you here.
00:48:53 Dr Genevieve Hayes
For listeners who want to learn more about you or get in contact, is there some way in which they can do that?
00:48:59 Dr Fei Huang
Yeah absolutely. I'm very happy and open to be connected. So I I can send you my LinkedIn profile and also I have a public profile on USW website. I'm I'm happy to be connected via LinkedIn or my email address.
00:49:13 Dr Genevieve Hayes
Yes. And I'll put those in the show notes for anyone listening. Yeah. So thank you so much for joining me here today.
00:49:22 Dr Fei Huang
Thank you so much Genevieve. I really enjoyed our conversation today and I'm sure this is a great topic and more and more conversations and discussions is also going on in our society.
00:49:34 Dr Genevieve Hayes
Yeah, yeah. I can only say this. Getting bigger in the future, yeah.
00:49:39 Dr Genevieve Hayes
And for those in the audience, thank you for listening to me today. I'm doctor Genevieve Hayes. And this has been value driven data science brought to you by Genevieve Haste Consulting.