Why Women Need to Be Involved in Data Science to Prevent Bias in Algorithms – Techstrong Con 2023

Hi everybody. I am Mikayla visani. I am the lead of machine learning team at frostrop.

And I almost also comana shindirector of kerosene Tech that is an on profit organization that seeks to get women into technology and girls. So today. I am going to talk about why we need more women in data science teams.

So first of all, why algorithms are biased nowadays more and more with tend to trust more into machines and algorithms. But and but we don't take into account that they might have errors and they might maybe biased. so to explain what biases in algorithms.

I like to take this example that is a in a tool that we usually do a mostly every day. At least me. I don't know you but if you put in Google Translate And she is an engineer and translate to running language that doesn't have a gender associated.

When we you try to translate this again, they are going trying to figure out what is the gender Associated and it is based on probabilistic and it's more probable that engineer will be a man that I get. So that's why they are going to choose to put heat. But what is happening here?

It means that A goal is wrong with this bias. It doesn't mean that it's wrong because the data says that it's not the algorithm. It might be wrong how the the design of the tool is made to to show this error.

And what is wrong is is the bias that is present in the society that is reflected in data as a consequence in algorithms. So this this talk will be about first of all what is data science? And what is Vias and then talk about what are the contributing factors that affect bias and what can we do to reduce it and finally a focus on diversity teams?

That is why to reduce wise. So first of all, what is data science? Data science is not only about algorithms.

We have a great focus on data. And also we need a really understand the business around that data to in order to get value then from the data. And we also need a lot of statistics and analytical skills to understand how the algorithms Works behind the scenes to choose the best algorithm for our problem and now how to tune this algorithms.

to get the best results what is machine learning inside data science machine learning is the use of algorithms to learn from data. How doing computations over the data by based on a statistics? And machine learning is used in many ways in all the industries.

We we can use it to find patterns in the data predict outcomes fine. That the disease or detects opportunities errors also in in the data or funding groups. In the data, so it applies on many Industries.

It can be Health Care as diseases or it can be e-commerce or otherwise So what is bias? bias transfer refers to an unfair outcome that in these cases the algorithm that the output of the algorithm that affect a certain group of people or a person or an idea. so the virus comes from first the society that has the bias and historical data available is with that bias.

so algorithms are trying on data and that is why the there are output might be biased also, so it is a fact that bias cannot be removed because the algorithms are trained in a set of information. They don't have the whole reality and but it's rude also that we can do something to reduce or mitigate this bias. So the biosity in data Sciences presented in all the pipeline from data insects in Houston.

So selecting. What is the data that we are going to use prepare the data? What are the Transformations that we do over the data to prepare it for the model how we we visualize that information and understand it.

What are the conclusions over the data we get and what data We Choose Or Not choose and that are modeling refers to the part of the training of the algorithm. Also, we humans can infer bias into the algorithms. And finally it also is important how we understand the results.

From from the algorithm. So in all the phases we can introduce bias. First of all, and now I I will go over through different examples that some tools the biases is present.

And in this example, especially refers to to the first example about the whole translate is not to say that it's not that Google is not doing anything about the problem. It's just it's a difficult problem to solve and Some years ago, they introduced this functionality. So when the algorithm doesn't it's not able to find what is the gender associated with the phrase?

They put their both options feminine and masculine. As as we can see here. But there are some cases that is still the algorithm doesn't work very well in this example.

I try it yesterday and we we can see if you know some Spanish that engineer engineero refers to to women. So in the translation something happens that I am not sure what what is happening here, but it says that I am in Canada instead of Henrietta. So there are still something to do we can help goal to resolve this problems.

Just sending free but here and and help. to to improve Another example, it's about cpt3 that has been a trending topic now with the new versions coming out. There was a study.

That they found that the adjectives related to women were more like Beauty easygoing or optimistic and the the adjectives related to men were more neutral. So this is a standard bias that is in the data and is reflected in the output of the algorithm. Of course now this kind of problems are being tackle and resolved but still.

The bias is present and it's difficult to be removed. Bias interest present only in text but also present in any matches so we can see here two examples of another big tech companies like Twitter and more photos. They they problem in Twitter was around the image crapping.

It had bias against certain kind of people. um goal photos producing resist outputs regarding a photo that was considered you search on goal and for gorilla and instead of gorillas, it came persons. So this kind of things can be really damaging for for some people in the society.

Another example is gender stress that this this girl made a study the when she found out that an algorithm didn't recognize her face. But if she uses a blank mask the algorithm did recognizer. So this kind of things you can see how the machine can be if this type of algorithms are starting uses us as a tool in in Industries like healthcare or other things.

Other example we start text images or in also invoices are present the bias is present since a for example, The Voice assistants all Syria, Alexa Cortana and will assistant all of these have feminine voices. So what is this? It's because the available information and regarding this type of work was a related to woman voices since the work was related more with women.

but we are in this way reinforcement the the gender stereotypes that are present in the society. So we can see that this tools are being used in an important decisions and the algorithms are not ready yet to to do that. So we cannot replace someone yet for for this type of retician using the algorin.

There was a tool that Amazon uses for recruiting people and it has gender Vias. So was the discriminated against women Another tool is compass that was used in. To to determine if a person how to to return to Shale or not.

And in healthcare also, they're found that Ascender bias in a tool that determine if they diagnosis of a patient. So as you can see, this is very dangerous because we will be discriminating group of societies using using the student not considering that it might be by us. So what are the contributive factors that may affect bias?

First of all, as we said that is why us it might be complete and probably not represent the whole population. and also not only data but we humans the ones that develop the the designs and the alchems we are biases and also we are the ones that choose the data and trying those algorithms. So we have to take this into consideration.

There are different types of bias of human bias these three years just as examples. Availability so we tend to overestimate. Answering an assessment refers to that.

We compare to our reference point. And that reference point doesn't have to be the whole truth and the other and the most common is stereotypes. And it's the one that maybe is most related with gender.

So If for example before I asked my mother how a programmer would like before I I became a programmer. She will think about something like this. So.

men with with with this kind of t-shirt and behind a computer. So this is a stereotype we have in our mind, but if you think about right now how many programs you know and look like this there are very few. So but it's a bias that is present in our society and the algorithm also.

imagine a program and like that based on the data it has So what is the problem with the stereotypes is it wrong? I don't think that it's wrong because it's our our brain grow. The our brain works we have to it is a shortcut that we do to not to better things faster.

But the problem is with the stereotypes is a that they became the only true available and this phrase is the this writer that I considered that is a Very good, right? So women are misrepresented in data in teams in algorithms and decisions. And in this way, we are missing half of the population View and why I say that it's very important that we get women into into Tech teams mostly in AI because nowadays AI is being used in every industry.

So it is important to create the solutions with our real. So here are some numbers right now. The AI professionals are 22% that researchers that wrote a paper in Earth X5 is 13 and the ones that participating conference are 18.

So it are very few women. that are in the industry but we are talking about bias and its presenting data why we just not remove the gender from the data and the problem is solved now. It's not that simple because the gender is In a way hidden in another variables, it's not that simple.

So what can we do to reduce the bias? We can do many things. First of all, really understand what is the data?

What is it coming from? Check the samples that we get from the data check if the data is balanced and if not get more data. Are also be aware of our own bias and the historical bias that might be present in the data.

and not take as an assumption anything instead of validate that assumption with data and with people Also, it's very important to analyze the final results into the categories that we have. For example, if we have a gender the both Sanders, but not only which genders with all the categories. And also I think that is very important go and takes this your algorithm with real people.

Be careful. Also how to you show the results. The interpretation of the results might be bias.

And also another recommendation is go and ask to another call you for for a review and they might find found out things that you didn't find out and finally, it's very important to have a diverse team to have different point of view. So why diversity teams not only in data science, but in every industry, there are several papers and investigations that confirm that core working in collaboration can bring many benefits different point of view and working from different disciplines. These come with high performance Innovation success and creating more roles Solutions.

So also having low diversity in increase. the having love diversity can narrow opportunities because for example if girls also a Think about it this stereotype for a programmer. They will consider that they can't not be a programming and that's not true.

So that's part of the work that we do at Garden Tech trying to get cars into technology and show them that they they can do it. Um in data science, as I said, it's very broad areas that you can work. So it doesn't mean that you are going to be behind a computer the day.

It there are many opportunities in all the businesses. on Industries so I encourage you to take an eye not only in AI but in all Tech teams to create more inclusive code and look at your teams team and see how it builds because the hook or call matters and bring more women into the team into teams and motivate more girls to get into the industry. So this is just a message that I want to give that technique needs you just as you are.

It doesn't matter. If you are a programmer or not this solutions. We need to create together as a whole the society needs also to be involved to reduce the bias.

Here, I just leave you some talks and that are very interesting and also books regarding this topic and then I can share the slide with you to have. those recommendations and now I just open up for questions and I hear my email just in case you want to to contact me.