YouTube2Text

GCI World 2026 September Session8 Closing — Transcript

by GCI World (Matsuo-Iwasawa Lab) · 1,300 words · 281 segments · language en · Watch on YouTube

Full transcript

  1. 0:03Okay.
  2. 0:05Uh, you have two types of assignments.
  3. 0:07First one is homework, and the second
  4. 0:10one is the attendance survey.
  5. 0:13Here's how to submit your homework.
  6. 0:16First, click homework six for session
  7. 0:18eight, then submit your homework
  8. 0:20function from the blue button.
  9. 0:23A score of uh three three means your
  10. 0:26answer is correct.
  11. 0:30If you forgot to submit your homework by
  12. 0:33the deadline, please use the late
  13. 0:35submission portal. You can still get up
  14. 0:38to uh two points.
  15. 0:41Uh
  16. 0:44So the late submission portal is shown
  17. 0:47to everyone. So if you already
  18. 0:49submitted, you don't need to submit
  19. 0:51again. Uh you won't get extra points.
  20. 0:56Uh today's homework is due November 20th
  21. 1:02uh at 1:00 a.m. UTC.
  22. 1:04You can submit as many times as you
  23. 1:06want. Only your last submission will be
  24. 1:08graded.
  25. 1:10Uh you need at least uh 14 out of 24
  26. 1:15points in total.
  27. 1:17Also, please pay attention. The session
  28. 1:20six homework is due tomorrow.
  29. 1:23Uh and session seven had no homework.
  30. 1:28Today's attendance survey is also due
  31. 1:30November 12th at 1:00 a.m. UTC time. You
  32. 1:33can either join the lab lecture or watch
  33. 1:36the recording. You can only submit once,
  34. 1:38and there are no late submissions.
  35. 1:41You need at least seven out of uh 14
  36. 1:44surveys.
  37. 1:45And here's the reminder. The session six
  38. 1:48survey
  39. 1:49is due tomorrow, and the session seven
  40. 1:52survey is due November 13th, 1:00 a.m.
  41. 1:55UTC time.
  42. 1:57Uh one more reminder about the
  43. 2:00Please submit the reproducible code you
  44. 2:02used.
  45. 2:04It must run on Google Colab, otherwise
  46. 2:06you competition score may be
  47. 2:08invalidated.
  48. 2:10Now is the Q and A time.
  49. 2:13Questions about today's lecture. Let me
  50. 2:16share the Q and A website.
  51. 2:27Okay.
  52. 2:32First question. How should I choose
  53. 2:34among label encoding, one hot encoding,
  54. 2:37and cross features for different models
  55. 2:39and more
  56. 2:41complex tasks? Okay, I think it's a
  57. 2:46very good question and as we did in the
  58. 2:50exercise,
  59. 2:52the encoding method
  60. 2:53is truly depends on the model you use.
  61. 2:56For example, for tree models, label
  62. 2:58encoding is fine. Trees only
  63. 3:02trees only care about order, so the
  64. 3:04numbers don't confuse them.
  65. 3:08Label encoding is also good when the
  66. 3:10categories have a real order.
  67. 3:13For linear models, please use one hot
  68. 3:15encoding. With label encoding, the model
  69. 3:18thinks one category is bigger than
  70. 3:21another. For example, like the
  71. 3:23Titanic
  72. 3:25embarked we did before. Like it might
  73. 3:28think port S is bigger than port C. That
  74. 3:30makes no no sense.
  75. 3:33And
  76. 3:37and the cross features help when two
  77. 3:41features match together, like in the
  78. 3:44notebook, sex and sex combined with
  79. 3:48embarked show
  80. 3:50a much bigger survival gap than sex
  81. 3:52alone.
  82. 3:53So a linear model can't find this kind
  83. 3:56of pattern by itself, so you
  84. 3:58give it the combination, but don't make
  85. 4:02the
  86. 4:03cross features
  87. 4:05too much, because if you create a bunch
  88. 4:08of cross features, we will get overfit.
  89. 4:13Okay, and I will mark it as answered.
  90. 4:18And then the second question,
  91. 4:21how can we
  92. 4:25if
  93. 4:26actively choose between regularization
  94. 4:29and feature selection when
  95. 4:33dealing with severe overfitting between
  96. 4:36training and
  97. 4:38test
  98. 4:43Okay.
  99. 4:50I would recommend you to try
  100. 4:53regularization
  101. 4:54first
  102. 4:56because it's very easy.
  103. 5:00And
  104. 5:06and
  105. 5:07it's kind of like
  106. 5:09it wouldn't induce too much harms, and
  107. 5:13it can also
  108. 5:14stop the model from
  109. 5:17fitting too much noise like we
  110. 5:20[clears throat] saw like we saw the the
  111. 5:22the housing price prediction before.
  112. 5:26If we do not use the regularization,
  113. 5:29then
  114. 5:30then the
  115. 5:32predicted line, dashed line will
  116. 5:35pulled towards the very big value.
  117. 5:38That's
  118. 5:39that's will cause overfitting.
  119. 5:43So
  120. 5:47so
  121. 5:49I also recommend you to use feature
  122. 5:52selection when you
  123. 5:54have many useless features.
  124. 5:58And it's also better when you want to
  125. 6:01explain your model to someone.
  126. 6:07Okay, I hope this can answer your
  127. 6:11question.
  128. 6:15I will mark this
  129. 6:17down.
  130. 6:24Okay.
  131. 6:28Is gradient boosting better than XGBoost
  132. 6:32or CatBoost?
  133. 6:36Okay, [clears throat] I should think how
  134. 6:39to answer this question.
  135. 6:48Uh
  136. 7:06So actually the gradient
  137. 7:11boosting
  138. 7:17is the method and the XGBoost
  139. 7:20and CatBoost
  140. 7:22are faster.
  141. 7:25So
  142. 7:27they add extra protection against over
  143. 7:31fitting.
  144. 7:32So
  145. 7:34it's like
  146. 7:35no tool always wins. What I can suggest
  147. 7:38to you to do is to try two or three of
  148. 7:42them and compare them when you are doing
  149. 7:45the
  150. 7:46competition or the
  151. 7:48[clears throat] final assignment.
  152. 7:55Okay.
  153. 7:59And what's the difference between
  154. 8:01boosting
  155. 8:03and sampling and hand label picking?
  156. 8:07Okay, in
  157. 8:09in my understanding, the hand label
  158. 8:11picking means like you pick the
  159. 8:15proper
  160. 8:16feature by yourself.
  161. 8:21So
  162. 8:24the
  163. 8:26I would say there's no
  164. 8:32better or worse. The main difference is
  165. 8:35like who picks the
  166. 8:38examples. Like with hand picking, a
  167. 8:41person, you should look at the data and
  168. 8:44chooses the important or
  169. 8:46difficult examples.
  170. 8:50And people can be biased. And boosting
  171. 8:52does this automatically. First, it
  172. 8:55trains a simple model, then it checks
  173. 8:59then it checks which examples the model
  174. 9:03got wrong. Then next the model pays more
  175. 9:05attention to those mistakes. It repeats
  176. 9:09this many times and each new model fixes
  177. 9:12the errors of the one before at the end.
  178. 9:14All the models are
  179. 9:16com combined.
  180. 9:18So
  181. 9:23I would say
  182. 9:24you should try this method
  183. 9:27in your
  184. 9:29in your final assignment and
  185. 9:31competition.
  186. 9:34Okay.
  187. 9:38How to deal with overfitting because the
  188. 9:41train
  189. 9:43set accuracy 0. point
  190. 9:4753 and the test 0. 73.
  191. 9:53Okay, for the
  192. 9:56for the overfitting, like
  193. 9:59such
  194. 10:00big gap means the model
  195. 10:03learn a lot of noise from the training
  196. 10:06data
  197. 10:07training data.
  198. 10:08First, I would like to recommend you to
  199. 10:11check whether there's any data leakage.
  200. 10:14For example, if you wrongly pick the
  201. 10:17test
  202. 10:18test data into the
  203. 10:20training data.
  204. 10:24And
  205. 10:25and actually, when you use different
  206. 10:28model, there are different ways to
  207. 10:32prevent overfitting. If you are using
  208. 10:34the
  209. 10:35decision tree model, I would suggest you
  210. 10:38to use a smaller tree depth.
  211. 10:41And
  212. 10:43if you use the
  213. 10:46the
  214. 10:47linear model,
  215. 10:49today's feature engineering is a very
  216. 10:52good way for you to prevent overfitting.
  217. 10:55For example, you should
  218. 10:56preprocess the
  219. 10:58data
  220. 11:00using the method we learned today.
  221. 11:05Also, I would recommend you to use cross
  222. 11:08validation, not just
  223. 11:11same train and test split
  224. 11:15because one split can be very lucky or
  225. 11:17unlucky.
  226. 11:19Okay, I [clears throat] think
  227. 11:22[groaning]
  228. 11:27maybe I can ask
  229. 11:29I can answer one more question.
  230. 11:31Does standardization need to be applied
  231. 11:34to every numerical feature or can we
  232. 11:36leave some features
  233. 11:39and change?
  234. 11:41Okay, as we as we learned before, for
  235. 11:44tree models, you don't need to do
  236. 11:47standardization
  237. 11:49because scaling doesn't change the order
  238. 11:53the the order of the values.
  239. 11:55For linear models and KNN standard, I
  240. 11:59recommend you to standardize all number
  241. 12:01features
  242. 12:03because as we as I show you before
  243. 12:08in the housing prediction, if some
  244. 12:10values are very big, then the line will
  245. 12:13be
  246. 12:14pulled
  247. 12:15uh
  248. 12:16pulled towards such big values.
  249. 12:20So um
  250. 12:22yes, that's my suggestion.
  251. 12:28Okay, I think I will
  252. 12:30finish today's uh
  253. 12:32Q and A session.
  254. 12:47Okay, uh let's wrap up with a recap of
  255. 12:50today and an overview of uh
  256. 12:53session nine.
  257. 12:56Today we watched the lecture on feature
  258. 12:58engineering and ensemble learning,
  259. 13:01uh practiced them in the hands-on
  260. 13:03notebook, went over the homework and
  261. 13:05attendance survey, and had a Q and A
  262. 13:07session.
  263. 13:09Next week is session nine on November 12
  264. 13:13uh from uh 11:00 to 12 uh
  265. 13:1812:30 uh o'clock UTC time.
  266. 13:21Uh we will first uh get a quick kickoff
  267. 13:25and uh
  268. 13:2635 minutes lecture on marketing basics
  269. 13:30and terminate Q and A session, uh 25
  270. 13:33minutes final final assignment follow-up
  271. 13:36session, and a five minute recap.
  272. 13:40Uh
  273. 13:40so please bring your questions about the
  274. 13:43final assignment to the follow-up
  275. 13:46session.
  276. 13:48Okay, so it's sorry for it's a bit of
  277. 13:52early. That's all for today. Thank you
  278. 13:54for joining. Don't forget the homework,
  279. 13:58the attendance survey, and your
  280. 14:00competition code.
  281. 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.