GCI World 2026 September Session8 Closing — Transcript
Full transcript
- 0:03Okay.
- 0:05Uh, you have two types of assignments.
- 0:07First one is homework, and the second
- 0:10one is the attendance survey.
- 0:13Here's how to submit your homework.
- 0:16First, click homework six for session
- 0:18eight, then submit your homework
- 0:20function from the blue button.
- 0:23A score of uh three three means your
- 0:26answer is correct.
- 0:30If you forgot to submit your homework by
- 0:33the deadline, please use the late
- 0:35submission portal. You can still get up
- 0:38to uh two points.
- 0:41Uh
- 0:44So the late submission portal is shown
- 0:47to everyone. So if you already
- 0:49submitted, you don't need to submit
- 0:51again. Uh you won't get extra points.
- 0:56Uh today's homework is due November 20th
- 1:02uh at 1:00 a.m. UTC.
- 1:04You can submit as many times as you
- 1:06want. Only your last submission will be
- 1:08graded.
- 1:10Uh you need at least uh 14 out of 24
- 1:15points in total.
- 1:17Also, please pay attention. The session
- 1:20six homework is due tomorrow.
- 1:23Uh and session seven had no homework.
- 1:28Today's attendance survey is also due
- 1:30November 12th at 1:00 a.m. UTC time. You
- 1:33can either join the lab lecture or watch
- 1:36the recording. You can only submit once,
- 1:38and there are no late submissions.
- 1:41You need at least seven out of uh 14
- 1:44surveys.
- 1:45And here's the reminder. The session six
- 1:48survey
- 1:49is due tomorrow, and the session seven
- 1:52survey is due November 13th, 1:00 a.m.
- 1:55UTC time.
- 1:57Uh one more reminder about the
- 2:00Please submit the reproducible code you
- 2:02used.
- 2:04It must run on Google Colab, otherwise
- 2:06you competition score may be
- 2:08invalidated.
- 2:10Now is the Q and A time.
- 2:13Questions about today's lecture. Let me
- 2:16share the Q and A website.
- 2:27Okay.
- 2:32First question. How should I choose
- 2:34among label encoding, one hot encoding,
- 2:37and cross features for different models
- 2:39and more
- 2:41complex tasks? Okay, I think it's a
- 2:46very good question and as we did in the
- 2:50exercise,
- 2:52the encoding method
- 2:53is truly depends on the model you use.
- 2:56For example, for tree models, label
- 2:58encoding is fine. Trees only
- 3:02trees only care about order, so the
- 3:04numbers don't confuse them.
- 3:08Label encoding is also good when the
- 3:10categories have a real order.
- 3:13For linear models, please use one hot
- 3:15encoding. With label encoding, the model
- 3:18thinks one category is bigger than
- 3:21another. For example, like the
- 3:23Titanic
- 3:25embarked we did before. Like it might
- 3:28think port S is bigger than port C. That
- 3:30makes no no sense.
- 3:33And
- 3:37and the cross features help when two
- 3:41features match together, like in the
- 3:44notebook, sex and sex combined with
- 3:48embarked show
- 3:50a much bigger survival gap than sex
- 3:52alone.
- 3:53So a linear model can't find this kind
- 3:56of pattern by itself, so you
- 3:58give it the combination, but don't make
- 4:02the
- 4:03cross features
- 4:05too much, because if you create a bunch
- 4:08of cross features, we will get overfit.
- 4:13Okay, and I will mark it as answered.
- 4:18And then the second question,
- 4:21how can we
- 4:25if
- 4:26actively choose between regularization
- 4:29and feature selection when
- 4:33dealing with severe overfitting between
- 4:36training and
- 4:38test
- 4:43Okay.
- 4:50I would recommend you to try
- 4:53regularization
- 4:54first
- 4:56because it's very easy.
- 5:00And
- 5:06and
- 5:07it's kind of like
- 5:09it wouldn't induce too much harms, and
- 5:13it can also
- 5:14stop the model from
- 5:17fitting too much noise like we
- 5:20[clears throat] saw like we saw the the
- 5:22the housing price prediction before.
- 5:26If we do not use the regularization,
- 5:29then
- 5:30then the
- 5:32predicted line, dashed line will
- 5:35pulled towards the very big value.
- 5:38That's
- 5:39that's will cause overfitting.
- 5:43So
- 5:47so
- 5:49I also recommend you to use feature
- 5:52selection when you
- 5:54have many useless features.
- 5:58And it's also better when you want to
- 6:01explain your model to someone.
- 6:07Okay, I hope this can answer your
- 6:11question.
- 6:15I will mark this
- 6:17down.
- 6:24Okay.
- 6:28Is gradient boosting better than XGBoost
- 6:32or CatBoost?
- 6:36Okay, [clears throat] I should think how
- 6:39to answer this question.
- 6:48Uh
- 7:06So actually the gradient
- 7:11boosting
- 7:17is the method and the XGBoost
- 7:20and CatBoost
- 7:22are faster.
- 7:25So
- 7:27they add extra protection against over
- 7:31fitting.
- 7:32So
- 7:34it's like
- 7:35no tool always wins. What I can suggest
- 7:38to you to do is to try two or three of
- 7:42them and compare them when you are doing
- 7:45the
- 7:46competition or the
- 7:48[clears throat] final assignment.
- 7:55Okay.
- 7:59And what's the difference between
- 8:01boosting
- 8:03and sampling and hand label picking?
- 8:07Okay, in
- 8:09in my understanding, the hand label
- 8:11picking means like you pick the
- 8:15proper
- 8:16feature by yourself.
- 8:21So
- 8:24the
- 8:26I would say there's no
- 8:32better or worse. The main difference is
- 8:35like who picks the
- 8:38examples. Like with hand picking, a
- 8:41person, you should look at the data and
- 8:44chooses the important or
- 8:46difficult examples.
- 8:50And people can be biased. And boosting
- 8:52does this automatically. First, it
- 8:55trains a simple model, then it checks
- 8:59then it checks which examples the model
- 9:03got wrong. Then next the model pays more
- 9:05attention to those mistakes. It repeats
- 9:09this many times and each new model fixes
- 9:12the errors of the one before at the end.
- 9:14All the models are
- 9:16com combined.
- 9:18So
- 9:23I would say
- 9:24you should try this method
- 9:27in your
- 9:29in your final assignment and
- 9:31competition.
- 9:34Okay.
- 9:38How to deal with overfitting because the
- 9:41train
- 9:43set accuracy 0. point
- 9:4753 and the test 0. 73.
- 9:53Okay, for the
- 9:56for the overfitting, like
- 9:59such
- 10:00big gap means the model
- 10:03learn a lot of noise from the training
- 10:06data
- 10:07training data.
- 10:08First, I would like to recommend you to
- 10:11check whether there's any data leakage.
- 10:14For example, if you wrongly pick the
- 10:17test
- 10:18test data into the
- 10:20training data.
- 10:24And
- 10:25and actually, when you use different
- 10:28model, there are different ways to
- 10:32prevent overfitting. If you are using
- 10:34the
- 10:35decision tree model, I would suggest you
- 10:38to use a smaller tree depth.
- 10:41And
- 10:43if you use the
- 10:46the
- 10:47linear model,
- 10:49today's feature engineering is a very
- 10:52good way for you to prevent overfitting.
- 10:55For example, you should
- 10:56preprocess the
- 10:58data
- 11:00using the method we learned today.
- 11:05Also, I would recommend you to use cross
- 11:08validation, not just
- 11:11same train and test split
- 11:15because one split can be very lucky or
- 11:17unlucky.
- 11:19Okay, I [clears throat] think
- 11:22[groaning]
- 11:27maybe I can ask
- 11:29I can answer one more question.
- 11:31Does standardization need to be applied
- 11:34to every numerical feature or can we
- 11:36leave some features
- 11:39and change?
- 11:41Okay, as we as we learned before, for
- 11:44tree models, you don't need to do
- 11:47standardization
- 11:49because scaling doesn't change the order
- 11:53the the order of the values.
- 11:55For linear models and KNN standard, I
- 11:59recommend you to standardize all number
- 12:01features
- 12:03because as we as I show you before
- 12:08in the housing prediction, if some
- 12:10values are very big, then the line will
- 12:13be
- 12:14pulled
- 12:15uh
- 12:16pulled towards such big values.
- 12:20So um
- 12:22yes, that's my suggestion.
- 12:28Okay, I think I will
- 12:30finish today's uh
- 12:32Q and A session.
- 12:47Okay, uh let's wrap up with a recap of
- 12:50today and an overview of uh
- 12:53session nine.
- 12:56Today we watched the lecture on feature
- 12:58engineering and ensemble learning,
- 13:01uh practiced them in the hands-on
- 13:03notebook, went over the homework and
- 13:05attendance survey, and had a Q and A
- 13:07session.
- 13:09Next week is session nine on November 12
- 13:13uh from uh 11:00 to 12 uh
- 13:1812:30 uh o'clock UTC time.
- 13:21Uh we will first uh get a quick kickoff
- 13:25and uh
- 13:2635 minutes lecture on marketing basics
- 13:30and terminate Q and A session, uh 25
- 13:33minutes final final assignment follow-up
- 13:36session, and a five minute recap.
- 13:40Uh
- 13:40so please bring your questions about the
- 13:43final assignment to the follow-up
- 13:46session.
- 13:48Okay, so it's sorry for it's a bit of
- 13:52early. That's all for today. Thank you
- 13:54for joining. Don't forget the homework,
- 13:58the attendance survey, and your
- 14:00competition code.
- 14:02Um see you next week. Bye-bye.
About this transcript
This page contains the full transcript of GCI World 2026 September Session8 Closing by GCI World (Matsuo-Iwasawa Lab), generated from the public captions YouTube serves with the video. The transcript has 1,300 words across 281 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.