South Big Data Hub Education & Workforce Working Group, 15 Feb 2022

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Education & Workforce WG | Use of educational data to predict students’ performance

Description

February 2022
Use of educational data to predict students’ performance
Thereza Padilha; Georgia Tech

A

So, um first of all, thanks for having me and it's a little echo, no you're, okay, okay and for this opportunity to talk and share what I'm doing here in georgia, tech.

A

So today, uh I'll present a little bit the work that I me I'm doing here and I will talk about use of educational data to predict students performs, I am a visiting scholar in georgia, tech until may this year and I will come back to brazil and hopefully I will continue to meet everybody virtually so here um my advisor is professor richard cataprong.

A

uh We work in the problem-solve and educational technology lab at the school of psychology.

A

uh The agenda for this this talk, I divided basically an introduction just to see what I'm saying a contest contextualization and talk a little bit so the education data mining in that case of study following five steps: data collection, data, clean and transformation, feature selection, deep learning models and evaluation, and I will finally a brief conclusion.

A

uh The the main point of the research is great repetition.

A

It's an important topic for all educational institutions- great repetition here, I'm talking about students who keep in the same grades more than one year, so they did achieve at the minimum expected to go to next year night to next grade and for some psychologists, great repetition has not been seen as a fair strategy because it can cause some problems for these students like poor corona, pure interactions, aversion to school, behavior problems and probably you've dropped off a version through school.

A

In educational institutions, you know we have many academic and personal data that are stored and you can explore to find hidden, hidden or unknowing and knowledge yeah. I saw one mistake here, but okay, annoy or hidden knowledge for this is only uh to predict students, performance, user education. Data set is a topic that has been researched for several years, but there are some limitations.

A

uh And the basic goal is identify as soon as possible students who need assistance in the course because we can allow teachers to adopt pedagogical learning actions, avoid the students fail or drop out.

A

In this talk, I will share a case study, a case study involving data, deep learning models to predict students performance using different dimensionality reduction methods, and for these experiments we used uh one data set from portugal about high school.

A

Now that's available on the internet, I think until since 2014 yeah there is this kind of data set, so the education data mine, it is based on knowledge, discovery in database process. Very known. Very many people know this process, the fayad and others uh defining in 1996, but until now we can use to follow all the steps that basically, we have five steps.

A

We we are from raw data until the knowledge and we have selection, pre-process transformation data, mind and interpretation and evaluation. Now we did one case study base and all these five steps.

A

As I said, the case study is to predict the study to predict student performance in secondary education, using academic and personal data and based on this kdd process.

A

Now, basically, we let me take care of this.

A

We we collect data. We did some clean and transformation.

A

We choose four feature selection base. The idea is to minimize the amount of features that we have in the data uh pass present all this data for the deep learning models and see what we have and compare these models. Now. Let me explain all these steps that you we did the first one, the phase, one data collection, this data set uh it's possible to find in this website is, as I said, is from portugal.

A

We have data from portugal to portugal, portuguese schools. Basically, this data, we can't find academic records, democratic demographic information, technological resources, social attitudes, family background and social economic stats status. I brought just a simple.

A

To to everybody knows better how features that we have in this data set, that is school age, address family size, mother's education, father's education, free time? How? How how much time the students have to free no, no, no study, and if we I have, for example, if they drink during the week, drink just weekends.

A

So there is 33 uh fixtures in this data set the number of of the absences and for the records academic records. Basically, we have about the grades. We have the grade first period grade and the second grade for the phase 2. Basically, data and cleaning transformation.

A

I brought just some examples, but we transform nominal to numerical values like internet, yes or no. We trained for zero or one and also we reduced.

A

The number of possible values like g3 is a target value target feature that means pass or fail in the in the class 0 to 9 and 10 to 20, because we saw in understand how is the portuguese um system to pass one year to another, zero to nine means that failed and 10 to 20 or of average or above average, but anyway he can pass and move move to the next great, the phase three uh we had.

A

We have the selection features because we had 33, but we used some methods, some algorithms, to look all these features and try to see what's the most important and using information gain, we got 17 relief, 10 features, boruda, 14 and record fixture elimination just 13..

A

This is uh the features that it selected for each algorithm. Some I put in bold as bold because they are exactly the same. Like the two grades and the amount of failures.

A

The phase four basically was the deep learning models. Now then, we create five deep learning models: the google collab using tensorflow and keras libraries.

A

We had we present personal and data cadence use for neural network to see if pass or fail and the first layer.

A

We did some changes because sometimes we use 33, sometimes with 17 the same amount of each algorithm choose the phase. Four I have.

A

I brought just a valuation, but just because it's ten minutes- I I I don't want to spend too much time this, but I I brought just training and for tests accuracy.

A

If I use all the features like 33 the training that I use. 80 percent of the data we had 100 percent accuracy, but the most important you need to focus is on test test accuracy, just 20 percent and in this case, in this study, the boruto.

A

14 attributes got the best result, 0.939.

A

In the next slide is basically to see uh how the accuracy of the boruto algorithm and over the epochs as soon as we can see the examples he will improve over the time.

A

And for the conclusion, for this talk right to predict students performance, you have investigate the use of feature selection, algorithms, cho beta model with more accuracy than others. Would you like to see if it's possible to use legs features minimize the amount of data that you need to show for for the neural network, to try to classify if these students is possible to fight to fail? There is a high chance to fail in the future or not, and in this case the study de boruto helped us to find and to classify these students.

A

So um I'm looking for for other projects that I can collaborate in the future and I my idea just bring and share with you what I'm doing.

A