YT YouTube2Text

YouTube2Text → Transcripts → GCI World 2026 September Session7 Competition Part

GCI World 2026 September Session7 Competition Part — Transcript

by GCI World (Matsuo-Iwasawa Lab) · 4,752 words · 847 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Right, so let's get into the
0:02competition. So, um as you know, the
0:05competition is about creating your own
0:07predictive model.
0:08Um sorry that it is old, but so um this
0:12is basically through uh
0:14we we In the competition, you're
0:16required to do exploratory data
0:19analysis, meaning that you you would you
0:20need to analyze the data um
0:23and understand what the data set is
0:26about. You need to then create your own
0:28machine learning model, and then you
0:30need to write fully evaluate it and sort
0:32of improve your model
0:34um and aim for a higher score.
0:37Um just as a reminder, um
0:41the competition is sort of in the core
0:43um course completion requirements.
0:46So, if you are aiming for a standard
0:48completion, um the in-class competition
0:51is optional. So, it means that even if
0:54you don't participate at all, you will
0:56still get a standard completion.
1:00However, if you are aiming for either
1:03the honors or outstanding completion,
1:06you will need to participate in the
1:09competition
1:11and get in the higher rankings.
1:14Right? Um the the final assignment that
1:17we will that we will cover in the latter
1:19half of today's session, this is
1:22mandatory for any kind of completion,
1:25but the the in-class competition, this
1:28is op- optional for if you want if
1:30you're just aiming for standard
1:31completion.
1:33Right?
1:34Um
1:35so and then yeah, so the competition is
1:39So, you'll be working on the Home Credit
1:42Default Risk competition. This was
1:44originally held on Kaggle back in 20 18.
1:48Um and then your your task is to create
1:50a machine learning model that best
1:52predicts the probability of default for
1:55each applicant. So,
1:57imagine that you're um you
2:00you are
2:02you know, you you you you you you
2:04So, the data set is about your clients
2:06like their attributes such as their
2:08income or where they live etc. And then
2:12your task is to create a machine
2:14learning model that predicts whether
2:17the client is going to default on their
2:19loan.
2:22So, how to do this? Um so so how to get
2:24started? So, I'll first walk through
2:26using the slides and then I'll show you
2:29I'll actually show you um
2:32you know, how to do it as well. So, so
2:34first of all, you will need to go to the
2:37student drive and then you need to go
2:39into the competition folder
2:41and download the competition folder to
2:43your local PC.
2:45Um you know, and then
2:48you know, you can click either this more
2:50button here or click right click and
2:52then click download. Either way, since
2:55this is a folder, it's going to it
2:57should download as a zip file.
3:02All right.
3:03And then the next thing is that you will
3:05need to unzip
3:06the zip file that you just downloaded.
3:10By unzipping, it's going to show you it
3:12competition you know, it's
3:14it's going to the zip file is going to
3:16unzip into a folder
3:19that named competition.
3:21And you will need to then upload
3:24the competition folder to your Google
3:26Drive.
3:27This can be anywhere in your
3:30Google Drive but you you you will need
3:34to remember where you placed the zip
3:36file.
3:41Right.
3:42And then
3:43first thing is to
3:45um click the Jupiter notebook inside the
3:48competition folder
3:50um and
3:52the notebook inside it is basically it's
3:55your base line model meaning that
3:57if you run it for you know, if you run
3:59it entire from the beginning to the end
4:02you have your first machine learning
4:04model that's already working.
4:07Um you will need So, as I said earlier,
4:10there's one thing There's one place that
4:12you need to sort of change according to
4:16where you place your notebook.
4:19So, you need to change the path that
4:22points to the directory so that it So,
4:24it So, you need to change the path so
4:26that it points to whichever folder that
4:30you uploaded the competition folder.
4:35And then you run all the cells in the
4:37baseline to create and it's going to
4:39create your a submission.csv
4:41So, sub- submission.csv. So, it's going
4:44to create a CSV file.
4:48Once you have the CSV file, you will
4:51need to then Again, you need to download
4:54the submission.csv.
4:58And submit on on the campus.
5:03So, you go to
5:04You click competition, you click the
5:06submit CSV file button,
5:09and then click submit button,
5:11and you You're now officially in the
5:14competition.
5:17Right? So, that's all the steps that you
5:19need to do to join the competition.
5:23Right? So, let's actually
5:26um
5:27review them one by one. So, first thing
5:29you need to go to this the student
5:31folder, right?
5:32So, this is the student folder that we
5:34have. And then inside, you know, we have
5:37the lecture materials I think that
5:38you're
5:40you access every week, but you know, you
5:42know, you go to the 04 competition
5:45folder.
5:49And then inside it you see You can see
5:50that there's this advanced techniques
5:52folder, which I'll explain later, and
5:54then you have the competition folder,
5:56and also the comp- competition tutorial.
5:59So, this competition tutorial, this also
6:02has all the details you need to know
6:04about So, so what I'm explaining right
6:06now, you can find this in the
6:08competition tutorial PDF
6:12Right? So, what you need to do, you need
6:14to download this competition folder,
6:16right?
6:31A while this is downloading
6:41Okay.
6:42Yeah, so now I have my competition
6:44folder.
6:46So, now I'm going to unzip this.
6:52So, now I can you can see that I have my
6:55unzipped folder here.
6:58See?
7:01So, now I'm going to do is I'm going to
7:02my I go to my
7:05Google Drive
7:07and inside my Google Drive my My Drive,
7:10I have a folder named GCI.
7:13So, I am just going to let me just
7:17change this to competition. Okay.
7:21Let's change the Okay.
7:23Yeah, I'm just going to
7:25upload this entire folder
7:28to my Google Drive.
7:34It's going to take a bit of a while.
7:39Um
7:42Yeah. Yeah.
7:44Yeah, so I will answer to your questions
7:47a bit later, but in the meantime you can
7:50send the questions and then I
7:53and I will answer them.
7:55I'll I'll be looking through, but I'll
7:57take time later on to answer.
8:01It's going to take a while
8:03um
8:05for you to
8:11upload this.
8:20So,
8:29this should finish in maybe 5 more
8:31seconds.
8:37Right. So, now I have my competition
8:39folder. I'm just going to
8:43rename the competition. So, inside my
8:45Google Drive, my My Drive, I have this
8:47GCI folder folder. And then inside my
8:50GCI folder, I have the competition.
8:54And inside this, you can see that
8:56there's
8:57one one folder that says input, and then
9:00one Excel file, a readme notebook, and a
9:03tutorial.
9:06The readme has all the
9:09It says that all all the rules, so
9:11please take a look at this.
9:15Um yeah, and then but for today, I will
9:18go through the tutorial very
9:21quickly.
9:31So,
9:32you will need to go through the tutorial
9:34So,
9:35how to go through the tutorial? So, you
9:37can just run this to import the cells,
9:40import the libraries.
9:42And then the second cell, by running the
9:44second cell, you're going to connect
9:46your Google Drive with the
9:48Um you're you're going to connect your
9:50Google Collaboratory with um
9:52Google Drive.
9:55When you run the second cell, it's going
9:57to pop up this thing that asks you to
9:59connect it to Google Drive.
10:02So, you need to
10:05click approve, approve.
10:09And it's going to connect it.
10:12Now this third part is
10:15the third cell is where you need to
10:18change.
10:19So you need to change this so that it
10:22matches to where you uploaded
10:26um this competition folder to.
10:30So how you can do it? If you already
10:31know, you can just write it down here.
10:34If you are unsure, how you can do it is
10:36by clicking this folder icon.
10:39Go inside drive.
10:42My drive.
10:44Then this my drive folder is basically
10:46your Google Drive. So for my case, I
10:49have it inside GCI
10:51folder.
10:58Yeah, so and then here you can see that
10:59then this is the competition folder,
11:01right?
11:03So I can click this button here. Click
11:06copy path.
11:07And
11:09like this. Like this. Like so.
11:14And then run the fourth cell.
11:17And if you if it shows all files exist
11:19and path is correctly set, then that
11:21means that you're all set.
11:24So now all you need to do is just run
11:26everything.
11:33All right. You run everything.
11:38And it's going to
11:42create a submission CSV file.
11:53Right?
11:55So once you're done that,
11:57if you go back to competition, um you
12:00will see let's see. It's still running.
12:04Hurry. Hurry.
12:06Yep. Yep. There we go. So once you've
12:08run it, you'll be able to
12:12This is the input folder.
12:15So yeah, so now you should see an output
12:17folder. Inside you should see the
12:19submission.csv.
12:21So you could download this.
12:25You go to Omni Campus.
12:28Go to competition.
12:33Click submit CSV file.
12:35Um
12:36see where is it?
12:38Submission.csv.
12:41And then click submit.
12:49Right.
12:50Now you should be able to see your
12:53results.
12:55And
12:57yeah, there we go.
13:00Now your task is to improve your model
13:03um and then work yourself up the
13:05um leaderboard.
13:07Right?
13:14So
13:16that's all So you can you So if you want
13:18to um
13:20So you know, the the recording of this
13:22will be um released tomorrow. So you can
13:24follow along. You can watch it again.
13:27You can also So there's also the
13:29competition tutorial PDF file.
13:32So you can also take a look at that. It
13:34has the all you know, it has the same um
13:38materials.
13:39Right?
13:41But for the
13:42in the remaining time I have, um I want
13:44to also talk about the advanced
13:46techniques. So uh some of you already
13:49know, we've we we are releasing several
13:53tutorial notebooks um
13:56which you can use to um
13:59um and use it to um improve your um
14:04model.
14:05So the first one I recommend to So this
14:08is sort of built on top of each other.
14:10We have two more notebooks coming up.
14:13Um
14:14So I recommend you first take a look at
14:17the tutorial adding features and then
14:19using the tutorial feature engineering
14:22secondly.
14:24So today um, what I will do is use the
14:28tutorial feature engineering um,
14:30notebook.
14:32I actually have them downloaded
14:35um,
14:37here. So I will upload it
14:41and use it.
14:56Right. So again,
14:59I
15:00we will run
15:01all the cells.
15:05So when you're working on the
15:07competition, it's really important that
15:09you don't just
15:11um,
15:12you know, start exploring
15:15and like trying all different kinds of
15:18things at the same time. Um, it's really
15:21important to sort of
15:23set a strategy um, or or to to
15:26set a hypothesis
15:29um, and then you test it out. So you
15:30change all the you you you keep all the
15:34variables fixed and then you change one
15:36thing at a time.
15:38See if that improves your model.
15:41Um, and then if you if it does, you keep
15:43it. If it doesn't, you you disregard and
15:46think of a different way.
15:49So and then you need to repeat that
15:50process over and over again.
15:53So
15:55let me use some of the slides here.
16:00All right. So
16:02how do you test whether the
16:06hypotheses
16:08you build is a good one or not?
16:11You
16:12As you can see here um,
16:15so probably you will be using the
16:17cross-validation method that you that we
16:20covered during the class, right? So, you
16:23it means that you have a validation set.
16:26Um so, you or you you you have your
16:28training set, you divide it into
16:32some number of folds. Um
16:34here the the this example shows a
16:36five-fold um cross-validation.
16:40And then you use
16:42four of them as your training set and
16:43then use that the remaining one for your
16:45as your validation set.
16:47And then you repeat that process five
16:49times.
16:51And then you get so the overall sort of
16:53the um average um
16:56So, you have, you know, you you have
16:58your training set and the validation
16:59set. You run them five times, you get
17:01and you see, you know, you can see for
17:03which fold um
17:06what was your um gain regarding in terms
17:09of the evaluation metric.
17:12A good improvement, you um should have
17:15an improvement across all of the five
17:18folds.
17:19Um you know, even if the each of this
17:22fold, you know, improvement in each of
17:24the fold might be small, but if it's
17:26improving across all folds, then it
17:28means that whatever you added is very
17:31effective um across all kinds of the
17:34data that's present.
17:36On the other hand, for example, maybe
17:40one, you know, adding what a feature
17:42might lead to some improvements for some
17:46of the folds, but maybe no improvements
17:49for some and even like a, you know, it
17:51may
17:52um degrade the performance in some of
17:54the folds. So, this is a bad
17:56improvement, meaning that you should
17:58think again about what kind of, you
18:00know, whether adding this particular
18:03feature is really effective or not.
18:06So, it's not just about improving on
18:08average, it's really about whether it's
18:11improving across all folds. This is one
18:14thing to keep it, you know, keep it in
18:15mind.
18:17The other thing that I was um explaining
18:20earlier as well is to treat each task as
18:23a hypothesis. So,
18:25you just don't, you know, you don't um
18:28explore randomly
18:30um testing all different things at once.
18:33You test one by one. So, it's like an
18:35you know
18:36if you've done maybe research research
18:39is also similar. I think business is
18:42also very similar like to this.
18:45But you you need to first set up a
18:47hypothesis.
18:49Test it out.
18:50See if it leads to a an improvement on
18:53your model.
18:54Um if it's not, then you need to think
18:57about why it's not. Then you need to
18:59test, you know, rebuild your hypothesis
19:02and repeat it over and over again.
19:06So, for example, um you know, one
19:09hypothesis could be that, you know, um
19:11in one of the maybe you have
19:14a a a feature that's about um whether
19:18this person this client is employed or
19:20not. Then you find that
19:23you know, maybe this number
19:25um
19:27there's this strange number and you
19:30think you know, maybe this number
19:32represent, you know,
19:34um my
19:35it's not meaningful.
19:37And you replace this number with, you
19:40know, a nan, a missing value.
19:42But then you run the the model, see
19:44that, you know, only one of the folds
19:46get better.
19:47Um and then on average it's like um
19:50degrading your performance. So, you
19:53would then need to um
19:55think again, build another hypothesis.
19:58Um
20:00and you know, maybe add, you know, try
20:02something else.
20:03You rerun the model, see that your model
20:07gets really better across all the four
20:09five five folds.
20:11Um and then you, you know, you will keep
20:13that change.
20:15And you repeat that all um again and
20:17again.
20:21Um and then the other thing I want to
20:24say is
20:27your model performance is basically
20:31um
20:32and to improve across two axes.
20:36One is how much more how much data you
20:39give to the model.
20:41And the other axis is
20:44or the model's capacity.
20:47So, the tutorial notebook, that's your
20:51baseline. So, that's the first working
20:53model that you have.
20:55And what you should aim for is to sort
20:57of increase your model across both of
21:00these two axes.
21:02Um, the adding features, so if we go
21:06back to this folder, or this folder,
21:10you should you can see that there's this
21:12adding features notebook and this
21:13feature engineering notebook.
21:16So, what this adding features notebook,
21:18this is a notebook about adding more
21:20features, more more columns to
21:23um, to the model. So, this is more about
21:26this Y axis direction, right? So, you're
21:28going to increase the
21:30um amount of data you have.
21:33Um,
21:34that you give to the model.
21:37And then the next thing you should
21:39consider is
21:41maybe you want to change your model. May
21:43um, you know, the the baseline uses
21:46maybe like logistic regression. Um, and
21:49you might think, okay, I want to try
21:50something new, um, something better like
21:52a random forest or other um, good
21:55models.
21:56That is an improvement um, on the Y
22:01axis. So, um,
22:03oh, sorry, the X axis. So, you're
22:05increasing the
22:07the model capacity, basically.
22:10And then
22:11you can also try, you know, I think we
22:13covered model um, hyperparameter tuning
22:16in session five, I think. So, this is
22:20going to be about tuning your model. So,
22:22this is going to sort of map this is
22:24about maximizing
22:26the model performance. So, it that this
22:28is going to be increasing in the x-axis.
22:32And then, um the And then, the
22:35the fifth thing that you can do,
22:37um
22:38this is going to be covered next week,
22:40um feature engineering. So, this is
22:42about creating new features by yourself.
22:46So, this is good This is going to be
22:48increasing in this y-axis in this
22:50figure.
22:52And then, you create, you know, repeat
22:53this over and over again, trying to
22:55improve the model and also the data, you
22:57know, in in the direction of a model and
23:00also the data.
23:01And, you know, you're going to get a
23:04very good, um
23:06model, um model in the end.
23:10Right. And then,
23:12as we have covered, um as as the
23:16you know, we we released the, you know,
23:18adding features notebook especially in
23:19the first because,
23:21um you know, even if you have a good
23:23model, if you don't have enough data,
23:25then
23:27um the model isn't really going to
23:28capture all, um you know, you know, good
23:31features, um
23:33because the input is limited. So,
23:36my my advice My advice is to start by
23:40increasing, you know, using all the data
23:42available.
23:43And then, um trying to see if you can
23:46improve the model.
23:49Right.
23:52So, I I I will really um give you the
23:56slides um later on.
23:58So,
24:00yeah, so let's see. So, here, this is so
24:03I will cover the feature engineering
24:05notebook. So, this is about adding new
24:08features to your note um to your model.
24:11So, you know,
24:14of course, you can, you know, there's a
24:15lot of features that you can use
24:17already, but there are some features
24:19that you can also create by yourself.
24:22We will also, you know, we'll cover
24:23feature engineering
24:25in the next week's session, but I will
24:28or give you a preview of what this is
24:31going to look like.
24:33So, even when you're going to create new
24:34features, it's important that you just
24:36don't like
24:38um run off and create
24:41um new features uh
24:43you know, my randomly. It's really
24:45important to have a hypothesis and then,
24:49you know, um use that hypothesis to test
24:52it out um
24:53and convert that into feature
24:55engineering or model improvement or
24:59um etc.
25:01So, then for example, um
25:06one perspective you can have is like,
25:09"Okay, so
25:10our task is to predict whether a client
25:13is going to default or not." So,
25:16one question to think about is, "If you
25:19were the, you know, the person giving,
25:21you know, um deciding whether to um you
25:25know, um give the
25:27approval of a loan or not, you you would
25:31think you you can think, 'Okay, what
25:33would I check? Like, what kind of
25:35quality what kind of attributes of the
25:38clients should I check?'"
25:40Um you know, you want to think, "Okay,
25:42will this loan be paid back? Like, how
25:45am I going to analyze that, right?"
25:48So, you can for example think, you know,
25:50"The same payment
25:52the same amount of this loan is easier
25:55to, you know, pay
25:57um if you if the person has high
25:59income."
26:01Um but it's going to be more difficult
26:03if someone has a low income.
26:07So, the question is, "Can they afford
26:09the payment of their loans?"
26:11Um you know, and then of course we have
26:14the, you know, So, and then actually in
26:16the in you know, in the columns we have
26:19this annuity feature and also the income
26:22feature.
26:23So, what you know, one thing you can
26:24look at is, "Look, what if we take the
26:27ratio
26:29of the annuity and income?"
26:32So, this is going to sort of tell you um
26:36the sort of the
26:37um capacity of the client's ability to
26:41pay back the loan.
26:43So, what you can do
26:48is have this annuity income ratio,
26:53which is here. So, you know, you take
26:54the training that you know, training
26:56data set, you take the annuity and then
27:00you take the income total
27:02and you take the
27:04you you you you you you divide the
27:07annuity by the income total.
27:10So, that's this is how you can create an
27:12annuity income ratio feature to your
27:15notebook.
27:17You can also think about
27:19um you know,
27:23let's say you have two loans of the same
27:25size, but
27:27even with the same size, you know, this
27:30can be repaid over very different
27:32lengths of time. So,
27:34you can ask, "Okay, how many payments
27:37are um do we have until it's repaid?"
27:40So, this you can take um this you can
27:43calculate by taking
27:45the credit and then you divide it by the
27:48annuity and then you can create this
27:49credit term feature.
27:53So, you know, you can take the AMT
27:55credit feature and then you can divide
27:57it by annuity.
28:00So, so this is So, you know, um
28:03again, I will uh so, what I want to
28:05emphasize here is that it's very
28:07important to
28:09that a perspective, you know, you know,
28:12build a hypothesis or some kind of this
28:14question um in a question form
28:17and then try out on different kinds of
28:20um methods.
28:22It can be either, you know, you know,
28:24adding more features, adding more data,
28:26or improving the model. It can be either
28:28way, but
28:29it's important to have some kind of a
28:31hypothesis behind it.
28:34So, I know I'm running out of time. So,
28:37some other things I want to mention
28:39about.
28:40One is that
28:42the current leaderboard you're seeing,
28:44this is the public data set. So, the
28:47total is um
28:49we're not using the entire um test set.
28:52We're only using a subset of it to show
28:55the public leaderboard.
28:57The final ranking is going to be decided
28:59by using the entire um test set. So,
29:03you
29:04you shouldn't really be, you know, my
29:06advice is don't try, you know, don't aim
29:09for the highest score in the current
29:11leaderboard.
29:13Um
29:14aim for a higher um
29:17score based on your um cross-validation
29:20score.
29:23Um rules of thumb, you should definitely
29:26submit a simple baseline early
29:29um
29:30instead of, you know, waiting, you know,
29:31waiting to create the perfect model.
29:34You should repeat iterated the this
29:36iteration
29:38um as much as you can.
29:40You should look at what your model sees.
29:42Um analyze the features using explor-
29:46exploratory data analysis instead of
29:48just feeding everything to the data to
29:50the model and hope that it something
29:53happen.
29:55You should
29:56what you should do is to change one
29:58thing at a time. Write down what you
30:00find.
30:01Um and then repeat that process over and
30:03over again instead of,
30:05you know, changing five things or 10
30:07things at once and just hope, you know,
30:09hope for the best.
30:11You should iterate
30:13do the iteration fast instead of, you
30:16know, waiting for
30:18like an hour to test one and you should
30:19you you should change it keep the
30:21changes small.
30:22And you should trust the the
30:24cross-validation score instead of change
30:26chasing the public leaderboard. So,
30:28these are some of the rules of thumb and
30:30I'll
30:31um I'll I'll share this slide with you
30:33um
30:35after the lecture.
30:37Right. So,
30:38um I think I will take maybe
30:41two or three minutes to answer the
30:42question.
30:44So, yeah. So, I think I will
30:46start with this one. Is a perfect score
30:48of one possible in the competition and
30:50what score score should I aim for? So,
30:54score one, this is
30:55this is definitely either overfitting
30:59or you're using um external data. So,
31:02this is something that should not be
31:04achievable. So, um
31:07um no.
31:10This is not uh yeah, so this is some
31:12kind of a rule violation. Um
31:16It could be that you might created a a
31:19very very good model that no one had
31:21ever thought of.
31:22That's why we're asking if you're if
31:24you're aiming for an honors or
31:25outstanding student, you will need to
31:27submit the code and we will run it
31:31um internally and test if there is any
31:33kind of rule violation.
31:36So, um no, but in general, you don't
31:38need to sort of um
31:40do do get the score one and be worried.
31:42Um just aim for a higher score. Um
31:45somewhere around the top 20%.
31:48Um the competition will be evaluated um
31:52using the the the the final leader board
31:55that is going to be run after closing
31:58the competition um on the deadline date.
32:01Um
32:03yeah. Aim for this. So, so just aim for
32:05top 20%.
32:07All right?
32:09Um
32:11Yeah, so
32:15You Does using leaderboard feedback the
32:17tune risk data leakage or overfitting?
32:19Yes, so
32:21my recommendation is don't aim for like
32:230.9 or something. Maybe just first of
32:26all, I'm getting 0.7
32:28seven or seven like just going above
32:310.75-ish
32:33is a great achievement already. So, I
32:36would suggest starting from there.
32:38Um I think like the higher scores are
32:42probably due to data leakage or
32:44overfitting, which means that it's
32:46violating the rules.
32:51Um
32:53let's see. What are the rules for using
32:55AI in the competition? What are doing to
32:57share the AI chat URL? Um so instead of
33:01sharing the chat URL, we will we're
33:04going to look at your code um overall.
33:07So
33:08you you if you're aiming for honors or
33:11outstanding, you will need to attach the
33:14um the entire um folder that you have.
33:20Right?
33:21Um
33:22can I use a local IDE or and you
33:24repository for the project?
33:28Is this about
33:31Yeah, so I think this is both about the
33:33competition and final assignment. You
33:35say you say you can use
33:37a good Git repository or you know you
33:39can write it you can you you don't have
33:41to stick to Google Collaboratory.
33:44One thing I will note is that please
33:46don't that you know make the repository
33:48public. Please give it private private.
33:53Right?
33:55How can I tune my model by changing its
33:57code? So you can take a look. So um take
34:00a look at the advanced techniques. These
34:03show you so this is built on top of the
34:06tutorial notebook.
34:07Um and it also gives you actually also
34:10gives you some hints about what to think
34:13about
34:14in order to it's like you're like or to
34:17to do the next step. So take make sure
34:19to check that out.
34:22Um should we edit the baseline
34:24submission series before submitting? Um
34:27yes. So you start you can start by the
34:30using the baseline definitely and then
34:32work yourself upwards and I will share
34:35the competition part one that I'm
34:36sharing I'm sharing right now.
34:39All right. So um
34:42how are students outstanding selective
34:44one scores are tied. So,
34:47in general,
34:48we will be looking at the final
34:50assignment more than the competition.
34:52But, the top 20% will
34:55even if the scores are tied, then
34:58both of them are going to be in the top
35:0020 in the
35:01in in within the threshold.
35:06Right?
35:07So,
35:11Right. I think that I think my time is
35:14up. Can we use any machine learning
35:16models and techniques including
35:17ensembles and top PFN? Yes, you may use
35:20that for the competition
35:23and also the final assignment as well.
35:26The problem statement of the
35:27competition, you can find it in the
35:29readme.
35:31Right?
35:34How should we prioritize feature
35:36engineering error analysis and try
35:38different models to improve the results?
35:40I'll I'll share the slide later, but
35:43improve
35:44the data first and then improve the
35:46model.
35:49Or anything.
35:51Okay.
35:54Right.
35:56Can we use external data sets? No, you
35:58may not.
36:00Right?
36:01So, I think my time is up and
36:05I think I will hand over to Takashi
36:07first, right?

About this transcript

This page contains the full transcript of GCI World 2026 September Session7 Competition Part by GCI World (Matsuo-Iwasawa Lab), generated from the public captions YouTube serves with the video. The transcript has 4,752 words across 847 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.