YouTube2Text

【G検定 聞き流し 練習問題】③機械学習の概要|練習問題52問・解答解説付き ~2026年試験対応~ #ワイズデータ — Transcript

by ワイズデータ [聞き流しで受かるIT・データ資格チャンネル] · 6,665 words · 1,289 segments · language en · Watch on YouTube

Full transcript

  1. 0:02In this video, I will present 52
  2. 0:04practice questions for Chapter 3:
  3. 0:06Overview of Machine Learning for the
  4. 0:08G-Certification.
  5. 0:10The questions are original, based on the
  6. 0:13G-Certification Syllabus 2024 Version
  7. 0:161.4, revised on May 11, 2026, by the
  8. 0:19Japan Deep Learning Association.
  9. 0:22We will cover four areas: supervised
  10. 0:25learning, unsupervised learning,
  11. 0:26reinforcement learning, and model
  12. 0:28selection and evaluation. After reading
  13. 0:31the question and the four options, I
  14. 0:32will provide 8 seconds of thinking time
  15. 0:34before presenting the correct answer and
  16. 0:36an explanation.
  17. 0:38You can learn by just listening or by
  18. 0:39watching the screen.
  19. 0:41Let's begin with Question 1.
  20. 0:45Question 1: Which of the following is
  21. 0:47the most appropriate description of
  22. 0:49training data in supervised learning?
  23. 0:52A: Provide only input data, with no
  24. 0:54correct labels given at all.
  25. 0:58B: Provide input data paired with the
  26. 1:00corresponding correct answer information
  27. 1:02.
  28. 1:04C: Provide only rewards, without
  29. 1:06providing any input data.
  30. 1:08D: Do not let the model memorize answers
  31. 1:10, but require human correction at every
  32. 1:12step.
  33. 1:22The correct answer to Question 1 is B:
  34. 1:24Provide input data paired with the
  35. 1:25corresponding correct answer information
  36. 1:27.
  37. 1:29In supervised learning, the model is
  38. 1:31trained using pairs of input features
  39. 1:33and the correct labels or target
  40. 1:35variables to be predicted.
  41. 1:37This is the fundamental concept for both
  42. 1:39classification and regression. A is
  43. 1:40similar to unsupervised learning, and C
  44. 1:42brings to mind the rewards used in
  45. 1:44reinforcement learning. The sequential
  46. 1:46human correction mentioned in D is not a
  47. 1:48required condition.
  48. 1:50Question 2: Which of the following
  49. 1:52supervised learning tasks generally
  50. 1:54qualifies as a regression problem?
  51. 1:58A: Predicting the price of a house based
  52. 2:01on its area, etc. B: Categorizing emails
  53. 2:04as spam or not spam.
  54. 2:06C: Sorting images of animals into dogs,
  55. 2:09cats, or birds. D: Determining whether
  56. 2:11or not a customer will cancel their
  57. 2:13subscription.
  58. 2:22The correct answer to Question 2 is A:
  59. 2:24Predicting the sale price based on
  60. 2:26housing area, etc. Regression is a task
  61. 2:28that predicts continuous numerical
  62. 2:30values. Typical examples include
  63. 2:32predicting housing prices, sales revenue
  64. 2:35, or temperature. B, C, and D are all
  65. 2:37classification problems that predict
  66. 2:39discrete categories.
  67. 2:42Question 3: Which is the most
  68. 2:44appropriate difference between simple
  69. 2:45linear regression and multiple linear
  70. 2:47regression?
  71. 2:49A: Simple regression performs
  72. 2:51classification, while multiple
  73. 2:53regression performs clustering. B: In
  74. 2:55simple regression, there is no target
  75. 2:57variable.
  76. 2:59C: Simple regression uses one
  77. 3:01explanatory variable, while multiple
  78. 3:03regression uses multiple.
  79. 3:05D: Simple regression is a linear model,
  80. 3:07while multiple regression is always a
  81. 3:09non-linear model.
  82. 3:19The correct answer for question 3 is C:
  83. 3:21simple regression uses one explanatory
  84. 3:23variable, while multiple regression uses
  85. 3:26several.
  86. 3:27Simple regression analysis uses one
  87. 3:28explanatory variable, while multiple
  88. 3:30regression analysis uses multiple
  89. 3:32variables to analyze the relationship
  90. 3:34with an objective variable. Both can be
  91. 3:37treated as linear regression.
  92. 3:39A misunderstands the analysis task.
  93. 3:41B denies the need for an objective
  94. 3:43variable in regression.
  95. 3:45D is incorrect in stating that multiple
  96. 3:47regression must be non-linear.
  97. 3:50Question 4: Which is the most
  98. 3:51appropriate application for logistic
  99. 3:53regression?
  100. 3:56A: Divide videos into clusters without
  101. 3:58training data.
  102. 4:00B: Maximize the discounted reward for
  103. 4:02the next action.
  104. 4:04C: Apply minimum information to
  105. 4:06continuous objective variables.
  106. 4:09D: Estimate the probability of a
  107. 4:11customer churning and classify whether
  108. 4:13they will churn or not.
  109. 4:23The correct answer for question 4 is D:
  110. 4:25Estimate the probability of a customer
  111. 4:27churning and classify whether they will
  112. 4:29churn or not. Logistic regression is a
  113. 4:31method that models the probability of an
  114. 4:33event occurring, especially for binary
  115. 4:35classification. Although the name
  116. 4:36contains "regression," it is widely used
  117. 4:38for classification. A is unsupervised
  118. 4:40learning, B is reinforcement learning,
  119. 4:42and C is a description close to standard
  120. 4:45linear regression.
  121. 4:47Question 5: Which is the most
  122. 4:48appropriate explanation for the basic
  123. 4:50mechanism of decision trees?
  124. 4:53A: Always project data into a random
  125. 4:55low-dimensional space.
  126. 4:58B: Branch data based on conditions for
  127. 5:00features to perform classification or
  128. 5:02prediction.
  129. 5:03C: Refers only to neural networks that
  130. 5:06propagate errors to subsequent layers. D
  131. 5:08: Select actions based solely on
  132. 5:10state-value functions.
  133. 5:20The correct answer for question 5 is B:
  134. 5:22Branch data based on conditions for
  135. 5:24features to perform classification or
  136. 5:26prediction.
  137. 5:28Decision trees are models that
  138. 5:29repeatedly branch data based on feature
  139. 5:31conditions to predict classes or
  140. 5:33numerical values.
  141. 5:34A is close to the explanation for
  142. 5:36dimensionality reduction, C is about
  143. 5:38neural network learning, and D is a
  144. 5:40topic of reinforcement learning.
  145. 5:44Question 6: Which is the most
  146. 5:46appropriate explanation of bootstrap
  147. 5:48sampling used in bagging?
  148. 5:52A: Create multiple training sets by
  149. 5:54sampling with replacement from the
  150. 5:56original data.
  151. 5:58B: Sort the original data only once and
  152. 6:01provide it to all models in the same
  153. 6:03order.
  154. 6:04C: Always remove teacher labels and
  155. 6:06change it to unsupervised learning.
  156. 6:10D: Exclude all samples one by one and
  157. 6:12learn only one model.
  158. 6:23The correct answer for Question 6 is A:
  159. 6:25Create multiple training sets by
  160. 6:27sampling with replacement from the
  161. 6:29original data.
  162. 6:31Bootstrap sampling performs sampling
  163. 6:33with replacement from the original
  164. 6:35dataset.
  165. 6:37In bagging, multiple models are trained
  166. 6:39on different samples and their
  167. 6:40predictions are aggregated.
  168. 6:42B does not mention sampling with
  169. 6:44replacement. C does not require the
  170. 6:46deletion of teacher labels. D differs in
  171. 6:48that it is limited to only one model.
  172. 6:51Question 7: Which is the most
  173. 6:53appropriate characteristic of random
  174. 6:55forests?
  175. 6:57A: Always train multiple linear
  176. 6:59regression models in sequence.
  177. 7:03B: Create only one decision tree and fix
  178. 7:05the features.
  179. 7:08C: Train multiple decision trees and
  180. 7:10aggregate their predictions.
  181. 7:13D: Always train by stacking deep neural
  182. 7:15networks.
  183. 7:25The correct answer for Question 7 is C:
  184. 7:27Train multiple decision trees and
  185. 7:29aggregate their predictions.
  186. 7:32Random forest is an ensemble method that
  187. 7:34uses multiple decision trees.
  188. 7:36It randomizes the data sampling and
  189. 7:38features used for branching to aggregate
  190. 7:40predictions from diverse trees.
  191. 7:43A does not necessarily use linear
  192. 7:44regression.
  193. 7:46B differs in that it uses only one tree
  194. 7:48, and D is not deep learning.
  195. 7:52Question 8: Which is the most
  196. 7:53appropriate comparison between bagging
  197. 7:55and boosting?
  198. 7:58A: Both cannot use teacher labels.
  199. 8:02B: Bagging is always one model, boosting
  200. 8:04always has an infinite number of models
  201. 8:07; C: Bagging is always a neural network
  202. 8:09.
  203. 8:11Boosting always uses decision trees.
  204. 8:14D: Bagging trains models in parallel for
  205. 8:16easy aggregation, while boosting often
  206. 8:18trains sequentially to compensate for
  207. 8:20previous weaknesses.
  208. 8:31The correct answer for Question 8 is D:
  209. 8:33Bagging trains models in parallel for
  210. 8:35easy aggregation, while boosting often
  211. 8:37trains sequentially to compensate for
  212. 8:39previous weaknesses.
  213. 8:41Bagging is a representative ensemble
  214. 8:43method that trains multiple models
  215. 8:45independently and aggregates their
  216. 8:46predictions.
  217. 8:48Boosting sequentially adds models to
  218. 8:49compensate for the errors of the
  219. 8:51previous models.
  220. 8:53A can be used in supervised learning.
  221. 8:55B misunderstands the number of models. C
  222. 8:58unnecessarily limits the models to be
  223. 9:00used.
  224. 9:02Question 9: What is the most appropriate
  225. 9:04basic concept of gradient boosting?
  226. 9:09A: Sequentially add weak learners to
  227. 9:11improve the loss function.
  228. 9:14B: Use the initially selected single
  229. 9:16predictor without modification.
  230. 9:19C: Delete all training data and perform
  231. 9:21clustering.
  232. 9:23D: Convert multiple features into
  233. 9:25principal components.
  234. 9:35The correct answer to Question 9 is A:
  235. 9:37Sequentially add weak learners to
  236. 9:39improve the loss function.
  237. 9:42Gradient boosting is an ensemble method
  238. 9:44that sequentially adds weak learners to
  239. 9:46improve predictions, using clues like
  240. 9:48the negative gradient of the loss
  241. 9:50function.
  242. 9:51B does not involve sequential model
  243. 9:53addition. C is unsupervised learning. D
  244. 9:55is a description close to principal
  245. 9:57component analysis.
  246. 10:00Question 10: Which is the most
  247. 10:01appropriate description of the
  248. 10:03fundamental concept of Support Vector
  249. 10:05Machines (SVM)?
  250. 10:09A: The sole purpose is to randomize the
  251. 10:11order of training data.
  252. 10:13B: Emphasize the distance, or margin,
  253. 10:15between the boundary separating classes
  254. 10:17and the nearest data points.
  255. 10:20C: Learn the discount rate for rewards
  256. 10:22from training data.
  257. 10:24D: Always classify all data into the
  258. 10:26same class.
  259. 10:35The correct answer to Question 10 is B:
  260. 10:37Emphasize the distance, or margin,
  261. 10:39between the boundary separating classes
  262. 10:41and the nearest data points. SVM is a
  263. 10:43classification method that uses the
  264. 10:45concept of maximizing the margin between
  265. 10:47the classification boundary and the data
  266. 10:49. Using soft margins allows for a
  267. 10:51certain degree of misclassification.
  268. 10:53A is not the objective of learning.
  269. 10:56C is a term from reinforcement learning
  270. 10:58, and D is inappropriate as a classifier
  271. 11:00.
  272. 11:01Question 11: Which is the most
  273. 11:03appropriate description of the kernel
  274. 11:05trick used in SVMs and other methods?
  275. 11:09A: Calculate by deleting all features.
  276. 11:13B: Guarantee learning without using
  277. 11:15training labels.
  278. 11:17C: Calculate inner products in a
  279. 11:19high-dimensional space without explicit
  280. 11:22mapping.
  281. 11:23D: Ensure that all data can always be
  282. 11:25separated by a straight line.
  283. 11:36The correct answer to Question 11 is C:
  284. 11:38Calculate inner products in a
  285. 11:39high-dimensional space without explicit
  286. 11:42mapping.
  287. 11:44Kernel functions allow for the direct
  288. 11:46calculation of inner products in a
  289. 11:47high-dimensional feature space.
  290. 11:50While useful for handling non-linear
  291. 11:52boundaries, it does not guarantee that
  292. 11:53all data can be perfectly separated.
  293. 11:56A does not delete features.
  294. 11:59B does not prohibit the use of labels.
  295. 12:01E does not guarantee separability.
  296. 12:05Question 12: Which of the following
  297. 12:07tasks best corresponds to multi-class
  298. 12:09classification?
  299. 12:11A Predict a customer's purchase amount
  300. 12:13for the next month.
  301. 12:15B Group customers by similarity using
  302. 12:17only their purchase history.
  303. 12:20C Learn a delivery route based on
  304. 12:22rewards.
  305. 12:23D Classify a handwritten digit image
  306. 12:26into one of ten classes, 0 through 9.
  307. 12:36The correct answer to question 12 is D,
  308. 12:39classify a handwritten digit image into
  309. 12:41one of ten classes, 0 through 9.
  310. 12:44Multi-class classification is a task of
  311. 12:46choosing one from three or more discrete
  312. 12:48classes. Handwritten digit 10-class
  313. 12:50classification is a typical example. A
  314. 12:52is regression, B is unsupervised
  315. 12:54clustering, and C is an application of
  316. 12:56reinforcement learning.
  317. 12:59Question 13: Which is the most
  318. 13:01appropriate description of the
  319. 13:03autoregressive (AR) model and vector
  320. 13:06autoregressive (VAR) model in
  321. 13:07time-series analysis?
  322. 13:11A. AR handles past values of one series
  323. 13:14, while VAR handles past values of
  324. 13:16multiple series.
  325. 13:19B AR is exclusively for image
  326. 13:21recognition and VAR is for speech
  327. 13:22synthesis.
  328. 13:25C AR is limited to unsupervised learning
  329. 13:28and VAR is limited to reinforcement
  330. 13:30learning.
  331. 13:32D Neither uses past values of time
  332. 13:34series.
  333. 13:44The correct answer to question 13 is A.
  334. 13:46AR handles past values of one series,
  335. 13:48while VAR handles past values of
  336. 13:50multiple series.
  337. 13:52The AR model explains the current value
  338. 13:54using past values of a single time
  339. 13:56series.
  340. 13:58The VAR model aggregates multiple time
  341. 14:00series and models them using the past
  342. 14:02values of each.
  343. 14:04B incorrectly limits the usage.
  344. 14:06C makes an inappropriate limitation
  345. 14:08regarding the learning method.
  346. 14:10D contradicts the concept of
  347. 14:11autoregression.
  348. 14:14Question 14: Which is the most
  349. 14:16appropriate benefit of ensemble learning
  350. 14:18, which combines multiple predictive
  351. 14:20models?
  352. 14:22A. It can guarantee a 100%accuracy rate
  353. 14:25for all training data.
  354. 14:28B. It can sometimes suppress the
  355. 14:30weaknesses or variance of individual
  356. 14:31models by combining their predictions.
  357. 14:35C. It always converts all models into
  358. 14:37random predictions.
  359. 14:39D. It can guarantee optimal predictions
  360. 14:41even without any training data.
  361. 14:52The correct answer to question 14 is B.
  362. 14:54It can sometimes suppress the weaknesses
  363. 14:56or variance of individual models by
  364. 14:58combining their predictions.
  365. 15:00Ensemble learning is a method of
  366. 15:01combining the predictions of multiple
  367. 15:03models.
  368. 15:04Bagging and boosting are representative
  369. 15:06examples, which can improve prediction
  370. 15:08accuracy and robustness depending on
  371. 15:10conditions. A and D are incorrect as
  372. 15:12they guarantee performance
  373. 15:13unconditionally. C is not a method that
  374. 15:16randomizes model predictions.
  375. 15:19Question 15.
  376. 15:20Which is the most appropriate
  377. 15:22description of unsupervised learning?
  378. 15:25A. It always minimizes the error against
  379. 15:27correct labels.
  380. 15:29B. It aims only to maximize rewards.
  381. 15:32C. It finds structures or regularities
  382. 15:35from data without correct labels.
  383. 15:37D. It predicts solely based on pre-set
  384. 15:40business rules.
  385. 15:49The correct answer to Question 15 is C.
  386. 15:51It finds structures or regularities from
  387. 15:53data without correct labels.
  388. 15:56Unsupervised learning typically learns
  389. 15:58cluster structures or low-dimensional
  390. 16:00representations from unlabeled data.
  391. 16:02Clustering and dimensionality reduction
  392. 16:04are typical examples.
  393. 16:06A describes supervised learning, B
  394. 16:08describes reinforcement learning, and D
  395. 16:10describes a rule-based system.
  396. 16:13Question 16.
  397. 16:15Which is the most appropriate
  398. 16:16description of the basic process of
  399. 16:18K-means?
  400. 16:20It alternately updates cluster centers
  401. 16:22and the assignment of each data point.
  402. 16:25B. It learns decision tree branching
  403. 16:27conditions based on correct labels.
  404. 16:30C. It selects actions by calculating
  405. 16:32future discounted rewards.
  406. 16:34D. It performs linear regression using
  407. 16:36only past values from a time series.
  408. 16:46The correct answer to Question 16 is
  409. 16:50alternately updating cluster centers and
  410. 16:52the assignment of each data point.
  411. 16:54In K-means, you set the number of
  412. 16:56clusters K, and repeat the process of
  413. 16:58assigning each data point to the nearest
  414. 17:00center and updating the centers. It is
  415. 17:02influenced by the choice of initial
  416. 17:03values and the number of clusters. B is
  417. 17:05about supervised learning, C is about
  418. 17:07reinforcement learning, and D is about
  419. 17:09autoregressive analysis.
  420. 17:12Question 17. Which of the following
  421. 17:14represents the results obtained from
  422. 17:15hierarchical clustering as a tree
  423. 17:17diagram along with the order of merging
  424. 17:19and distances?
  425. 17:23A. ROC curve, B. Confusion matrix, C.
  426. 17:26Decision boundary, D. Dendrogram.
  427. 17:38The correct answer to Question 17 is D.
  428. 17:40Dendrogram.
  429. 17:42A dendrogram is a diagram that
  430. 17:44visualizes the order and distance in
  431. 17:46which clusters are merged in
  432. 17:47hierarchical clustering.
  433. 17:49A represents evaluation of
  434. 17:51classification models, B represents
  435. 17:53aggregation of classification results,
  436. 17:54and C represents classifier boundaries;
  437. 17:56none are dendrograms for hierarchical
  438. 17:58clustering.
  439. 18:00Question 18. Which is the most
  440. 18:02appropriate description of Ward's method
  441. 18:04in hierarchical clustering?
  442. 18:08A. It always merges based solely on the
  443. 18:10maximum distance between clusters. B. It
  444. 18:12chooses the pair that results in the
  445. 18:14smallest increase in within-cluster sum
  446. 18:16of squares when merged. C. It maximizes
  447. 18:19the classification accuracy against
  448. 18:21external labels.
  449. 18:23D. It can only be used if principal
  450. 18:24components have been determined first.
  451. 18:35The correct answer to question 18 is B.
  452. 18:37Ward's method, which selects pairs that
  453. 18:39result in the smallest increase in
  454. 18:41within-cluster sum of squares, is a
  455. 18:43representative hierarchical clustering
  456. 18:45technique that proceeds based on the
  457. 18:47increase in within-cluster variance when
  458. 18:49two clusters are merged.
  459. 18:51A is an explanation close to the
  460. 18:53complete-linkage method. C is a
  461. 18:54supervised evaluation metric. D is
  462. 18:57incorrect because it requires Principal
  463. 18:59Component Analysis.
  464. 19:01Question 19: Which is the most
  465. 19:03appropriate main purpose of Principal
  466. 19:06Component Analysis (PCA)?
  467. 19:09A. Determine the action that maximizes
  468. 19:11rewards.
  469. 19:13B. Automatically add correct labels to
  470. 19:15observed data.
  471. 19:17C. Find directions that preserve as much
  472. 19:19data variance as possible to represent
  473. 19:21it in low dimensions.
  474. 19:23D. Always assign a specific theme name
  475. 19:26to unknown text.
  476. 19:35The correct answer to question 19 is C,
  477. 19:37finding directions that preserve as much
  478. 19:39data variance as possible to represent
  479. 19:41it in low dimensions.
  480. 19:43PCA is a linear dimensionality reduction
  481. 19:45method that finds directions with large
  482. 19:47data variance as principal components
  483. 19:49and aggregates information into fewer
  484. 19:51axes. A refers to reinforcement learning
  485. 19:53. B refers to labeling. D is related to
  486. 19:56text classification and is not the
  487. 19:58primary purpose of PCA.
  488. 20:00Question 20: Which is the most
  489. 20:02appropriate explanation for Singular
  490. 20:05Value Decomposition (SVD)?
  491. 20:07A. A method that decomposes a matrix
  492. 20:09into singular values and corresponding
  493. 20:11orthogonal matrices, also used for
  494. 20:13dimensionality reduction.
  495. 20:16B. A method to calculate rewards for
  496. 20:18multiple agents.
  497. 20:20C. An evaluation metric that separates
  498. 20:22classification results into true
  499. 20:23positives and false positives. D. A
  500. 20:25decision tree that always converts
  501. 20:27categories into a single value.
  502. 20:37The correct answer to question 20 is A.
  503. 20:39A method that decomposes a matrix into
  504. 20:42singular values and corresponding
  505. 20:43orthogonal matrices, also used for
  506. 20:45dimensionality reduction. SVD is a
  507. 20:47method that decomposes a matrix into a
  508. 20:49product of orthogonal matrices and a
  509. 20:51matrix containing singular values. It is
  510. 20:53used for dimensionality reduction and
  511. 20:55recommendations through low-rank
  512. 20:56approximation and other techniques. B
  513. 20:58relates to reinforcement learning, C to
  514. 21:00a confusion matrix, and D to a decision
  515. 21:02tree.
  516. 21:04Question 21: Which is the most
  517. 21:06appropriate purpose for using t-SNE?
  518. 21:10A. Directly calculate the accuracy of a
  519. 21:12classification model.
  520. 21:14B. Update action values in reinforcement
  521. 21:17learning.
  522. 21:18C. Predict sales based solely on past
  523. 21:21time-series values. D. Visualize
  524. 21:23high-dimensional data in 2D or 3D by
  525. 21:25emphasizing neighborhood relationships.
  526. 21:35The correct answer for Question 21 is D
  527. 21:37: Visualize in 2D or 3D while
  528. 21:39prioritizing the neighborhood
  529. 21:41relationships of high-dimensional data.
  530. 21:43t-SNE is a non-linear dimensionality
  531. 21:45reduction method used for visualization
  532. 21:47by representing local neighborhood
  533. 21:49relationships of high-dimensional data
  534. 21:50in a low-dimensional space. Care should
  535. 21:52be taken when interpreting the distance
  536. 21:54between distant clusters. A is about
  537. 21:56classification evaluation, B is
  538. 21:58reinforcement learning, and C is about
  539. 22:00prediction by autoregression.
  540. 22:04Question 22: Which is the most
  541. 22:06appropriate description of
  542. 22:07Multidimensional Scaling (MDS)?
  543. 22:11A: Create decision trees equal to the
  544. 22:13number of data types. B: Arrange in a
  545. 22:15low-dimensional space to reproduce the
  546. 22:17relationships of dissimilarity or
  547. 22:19distance between objects as much as
  548. 22:21possible. C: Always assign correct
  549. 22:23labels to data that lacks teacher labels
  550. 22:25.
  551. 22:26B: Determine the next action using an
  552. 22:28epsilon-greedy policy.
  553. 22:38The correct answer for Question 22 is B
  554. 22:40: Arrange in a low-dimensional space to
  555. 22:42reproduce the relationships of
  556. 22:43dissimilarity or distance between
  557. 22:45objects as much as possible. MDS is a
  558. 22:47method for finding a low-dimensional
  559. 22:48arrangement that preserves the
  560. 22:50relationships between objects based on
  561. 22:51their distances or dissimilarities as
  562. 22:53much as possible. It differs from PCA in
  563. 22:55terms of input and the criteria to be
  564. 22:57optimized.
  565. 22:58A does not describe an ensemble either.
  566. 23:00C is labeling, and D is action selection
  567. 23:03in reinforcement learning.
  568. 23:06Question 23: Which is the most
  569. 23:08appropriate description for using Latent
  570. 23:11Dirichlet Allocation (LDA) in document
  571. 23:13analysis?
  572. 23:16A: Assign every document to only one
  573. 23:18identical topic.
  574. 23:20B: It is a classifier that only
  575. 23:22determines the correctness of each
  576. 23:24document.
  577. 23:25C: Model documents as a mixture of
  578. 23:27multiple topics, where words appear for
  579. 23:30each topic.
  580. 23:32D: It is a method for resizing images.
  581. 23:44The correct answer for Question 23 is C
  582. 23:46: Model documents as a mixture of
  583. 23:48multiple topics, where words appear for
  584. 23:50each topic.
  585. 23:52The topic model LDA is a generative
  586. 23:54model that assumes each document has a
  587. 23:56mixture ratio of multiple topics, and
  588. 23:58each topic has a distribution of word
  589. 24:00occurrences.
  590. 24:01A denies the mixture of multiple topics
  591. 24:03. B is binary classification. D is a
  592. 24:06description of image processing.
  593. 24:09Question 24: Which is the most
  594. 24:11appropriate description of collaborative
  595. 24:13filtering used for recommendations on EC
  596. 24:15sites?
  597. 24:17A: Recommend products by utilizing user
  598. 24:19behavior history and similarity in
  599. 24:21preferences.
  600. 24:23B: Recommend by analyzing only the words
  601. 24:26in product descriptions.
  602. 24:27C: Display product names by sorting them
  603. 24:30in alphabetical order.
  604. 24:32D Change all products to the same price
  605. 24:34without using teacher labels.
  606. 24:45The answer to Question 24 is A:
  607. 24:47Recommend products using user behavior
  608. 24:49history and similarity in preferences.
  609. 24:52Collaborative filtering provides
  610. 24:54recommendations using interactions
  611. 24:56between users and products, such as
  612. 24:58ratings, views, and purchases. You can
  613. 25:00reference the behavior of similar users
  614. 25:02or items.
  615. 25:03B is close to content-based filtering, C
  616. 25:06is sorting, and D is price changing;
  617. 25:08none are recommendation methods.
  618. 25:12Question 25: Which is the most
  619. 25:14appropriate description of the
  620. 25:15cold-start problem in recommendation
  621. 25:17systems?
  622. 25:20A: The problem where all recommended
  623. 25:22products are physically cold.
  624. 25:24B: The problem where all user ratings
  625. 25:26are too high to calculate errors. C: The
  626. 25:28problem where a model's temperature
  627. 25:30parameter cannot be initialized.
  628. 25:33D: The problem where it is difficult to
  629. 25:35ensure recommendation accuracy for new
  630. 25:37users or products due to a lack of
  631. 25:39history.
  632. 25:48The answer to Question 25 is D: It is
  633. 25:50difficult to ensure recommendation
  634. 25:52accuracy for new users or products due
  635. 25:55to a lack of history.
  636. 25:57The cold-start problem is the difficulty
  637. 25:59in making recommendations because there
  638. 26:01is not enough behavior history or
  639. 26:03ratings for new users or items.
  640. 26:06Content-based methods that use product
  641. 26:07attribute information can be a
  642. 26:09supplement.
  643. 26:10A is a literal translation of the term,
  644. 26:12B is about evaluation metrics, and C is
  645. 26:14about different parameters; all differ
  646. 26:16from the recommendation cold-start
  647. 26:18problem.
  648. 26:19Question 26: Which is the most
  649. 26:21appropriate description of the basic
  650. 26:23structure of reinforcement learning?
  651. 26:27A: Learn using only pairs of correct
  652. 26:29labels and input images.
  653. 26:31B: An agent acts in an environment and
  654. 26:33learns a policy based on the rewards it
  655. 26:35receives.
  656. 26:37C: Unsupervised clustering of known text
  657. 26:40.
  658. 26:42D: Human-fixed agent actions without
  659. 26:44using rewards.
  660. 26:54The answer to Question 26 is B: An agent
  661. 26:56acts in an environment and learns a
  662. 26:58policy based on the rewards it receives
  663. 27:00.
  664. 27:01In reinforcement learning, the agent
  665. 27:03observes the state of the environment,
  666. 27:05takes actions, and receives rewards. It
  667. 27:07learns a policy for actions that
  668. 27:09maximizes the cumulative future reward.
  669. 27:12A is supervised learning, C is
  670. 27:13unsupervised learning, and D differs
  671. 27:15from the reinforcement learning concept
  672. 27:17that uses rewards.
  673. 27:20Question 27: What is the most
  674. 27:22appropriate combination of the basic
  675. 27:24elements of a Markov Decision Process (
  676. 27:27MDP)?
  677. 27:31A: State, action, state transition,
  678. 27:33reward, discount rate B: Teacher label,
  679. 27:36regression coefficient, mean squared
  680. 27:39error, decision tree C: Principal
  681. 27:41component, singular value, cluster
  682. 27:44center, dendrogram D: Precision, recall
  683. 27:47, confusion matrix,
  684. 27:51AUC
  685. 28:00The correct answer for Question 27 is A.
  686. 28:02State, action, state transition, reward
  687. 28:05, and discount rate.
  688. 28:07An MDP represents an environment using
  689. 28:10states, actions, transition
  690. 28:11probabilities, and rewards, and uses a
  691. 28:14discount rate to evaluate future rewards
  692. 28:16.
  693. 28:17This is the fundamental framework of
  694. 28:18reinforcement learning. B is mainly
  695. 28:20related to supervised learning, C to
  696. 28:22unsupervised learning, and D relates to
  697. 28:24evaluation metrics for classification
  698. 28:26models.
  699. 28:27Question 28: Which is the most
  700. 28:29appropriate description of the
  701. 28:31difference between the state-value
  702. 28:33function V (s) and the action-value
  703. 28:35function Q (s, a) in reinforcement
  704. 28:36learning?
  705. 28:38A: V does not consider rewards, and Q
  706. 28:41only uses today's rewards.
  707. 28:43B: V evaluates only actions, and Q
  708. 28:46evaluates only states.
  709. 28:48C: V represents the value of a state,
  710. 28:50and Q represents the value of a
  711. 28:52state-action pair.
  712. 28:54D: V and Q are always the same value, so
  713. 28:56there is no point in distinguishing them
  714. 28:58.
  715. 29:07The correct answer for Question 28 is C.
  716. 29:09V represents the value of a state, and Q
  717. 29:11represents the value of a state-action
  718. 29:13pair. The state-value V (s) represents
  719. 29:16the expected future reward from a state
  720. 29:18, while the action-value Q (s, a)
  721. 29:20represents the expected future reward
  722. 29:22when choosing a specific action in that
  723. 29:24state.
  724. 29:25A's explanation of learning information
  725. 29:27is incorrect. B's objects of evaluation
  726. 29:29are reversed.
  727. 29:32D denies the distinction between V and
  728. 29:34Q. Question 29: In reinforcement
  729. 29:36learning, where the discount rate gamma
  730. 29:38is set between 0 and 1, which is the
  731. 29:40most appropriate general effect of
  732. 29:42making gamma smaller?
  733. 29:45A: It becomes easier to prioritize
  734. 29:47future rewards over current rewards.
  735. 29:50B: Rewards are no longer used at all
  736. 29:52during learning.
  737. 29:53C: The number of states automatically
  738. 29:55becomes zero. D: The contribution of
  739. 29:58rewards in the distant future becomes
  740. 30:00relatively smaller.
  741. 30:09The correct answer for Question 29 is D
  742. 30:11: The contribution of rewards in the
  743. 30:13distant future becomes relatively
  744. 30:14smaller. With a smaller discount rate,
  745. 30:17the further away in time a reward is,
  746. 30:18the smaller its contribution to the
  747. 30:20present value, meaning rewards in the
  748. 30:22near future are relatively prioritized.
  749. 30:24A represents a line of thinking closer
  750. 30:26to when the discount rate is large. B
  751. 30:28does not mean that rewards are not used
  752. 30:30, and C is unrelated to the number of
  753. 30:32states.
  754. 30:35Question 30: Which is the most
  755. 30:36appropriate characteristic of Q-learning
  756. 30:39?
  757. 30:41A: Q-values are updated using the
  758. 30:43maximum value of the action-value in the
  759. 30:45next state.
  760. 30:47B: Each piece of data is always assigned
  761. 30:49to a specific class.
  762. 30:51C: Only the cross-entropy for the
  763. 30:53correct label is calculated.
  764. 30:55D: The variance in the direction of the
  765. 30:57principal component is maximized.
  766. 31:08The correct answer for Question 30 is A
  767. 31:10: Q-values are updated using the maximum
  768. 31:12value of the action-value in the next
  769. 31:14state.
  770. 31:15Q-learning is a representative
  771. 31:17value-based reinforcement learning
  772. 31:18method that updates Q-values using the
  773. 31:20maximum action value obtained in the
  774. 31:22next state. It is an off-policy method
  775. 31:24where the behavior policy and the target
  776. 31:26policy can differ. B describes K-means
  777. 31:29clustering, C describes loss in
  778. 31:31supervised learning, and D describes PCA
  779. 31:33.
  780. 31:35Question 31: Which is the most
  781. 31:37appropriate explanation for Q-value
  782. 31:39updating in SARSA?
  783. 31:41A: Updates by assuming only the best
  784. 31:43action in the next state.
  785. 31:46B: Updates using the Q-value of the next
  786. 31:48action actually selected.
  787. 31:51C: Updates using only teacher labels
  788. 31:53without using Q-values for all actions.
  789. 31:57D: Selects the same action every time
  790. 31:58without observing the state.
  791. 32:09The correct answer to Question 31 is B:
  792. 32:11Updates using the Q-value of the next
  793. 32:13action actually selected.
  794. 32:16SARSA is a representative on-policy
  795. 32:18method that updates using information on
  796. 32:20the current state, action, reward, next
  797. 32:23state, and the next action actually
  798. 32:25selected.
  799. 32:26A is an explanation close to the update
  800. 32:28rule of Q-learning.
  801. 32:30C confuses reinforcement learning with
  802. 32:32supervised learning, and D does not
  803. 32:34describe policy-based learning.
  804. 32:39Question 32: Which is the most
  805. 32:41appropriate explanation of the
  806. 32:43epsilon-greedy policy?
  807. 32:48A: Always chooses the action with the
  808. 32:50least reward.
  809. 32:51B: Always executes all observed actions
  810. 32:54simultaneously.
  811. 32:56C: Usually chooses the action with high
  812. 32:58estimated value and tries a random
  813. 33:00action with a certain probability.
  814. 33:03D: Does not change the action once
  815. 33:05chosen thereafter.
  816. 33:15The correct answer to Question 32 is C:
  817. 33:17Usually chooses the action with high
  818. 33:19estimated value and tries a random
  819. 33:21action with a certain probability.
  820. 33:24The epsilon-greedy policy usually
  821. 33:26selects the action estimated to be best
  822. 33:28currently, while choosing a random
  823. 33:30action with probability epsilon to
  824. 33:32balance exploration and exploitation.
  825. 33:35A is not a strategy to increase value.
  826. 33:38B does not assume simultaneous execution
  827. 33:40. D is an explanation that eliminates
  828. 33:42exploration.
  829. 33:45Question 33: Which is the most
  830. 33:46appropriate feature of the UCB policy
  831. 33:49used in the bandit problem?
  832. 33:52A: Never chooses untried actions.
  833. 33:56B: Uses only the average of estimated
  834. 33:58rewards for each action, ignoring the
  835. 34:00number of trials.
  836. 34:02C: Uses only the classification accuracy
  837. 34:04of correct labels as a criterion. D:
  838. 34:07Selects actions by adding an upper bound
  839. 34:09that considers uncertainty to the
  840. 34:11estimated reward.
  841. 34:20The correct answer for Question 33 is B
  842. 34:22: Select actions by adding an upper
  843. 34:24bound that accounts for uncertainty to
  844. 34:26the estimated reward.
  845. 34:29UCB stands for Upper Confidence Bound.
  846. 34:32By adding a bonus for uncertainty to the
  847. 34:35estimated reward, it allows for the
  848. 34:36exploration of less-understood actions.
  849. 34:39A hinders exploration.
  850. 34:41B ignores uncertainty derived from the
  851. 34:43number of trials, and C is a
  852. 34:45classification metric, not a bandit
  853. 34:47action selection method.
  854. 34:49Question 34: Which of the following is
  855. 34:51the most appropriate description of the
  856. 34:53fundamental characteristics of the
  857. 34:54multi-armed bandit problem?
  858. 34:57A: Balance the exploration of uncertain
  859. 34:59rewards with the exploitation of
  860. 35:01promising choices.
  861. 35:04B: Always predict image labels.
  862. 35:07C: Perform regression on multiple time
  863. 35:10series simultaneously.
  864. 35:12D: Update only one cluster center.
  865. 35:24The correct answer for Question 34 is A
  866. 35:26: Balance the exploration of uncertain
  867. 35:28rewards with the exploitation of
  868. 35:30promising choices.
  869. 35:33In multi-armed bandit problems, because
  870. 35:35it is unclear which option yields a
  871. 35:37higher reward, balancing exploration and
  872. 35:39exploitation is crucial. An example is
  873. 35:41selecting which advertisements to
  874. 35:43display. B describes image
  875. 35:45classification, C describes the VAR
  876. 35:47model, and D is close to the K-means
  877. 35:49method.
  878. 35:50Question 35: Which of the following is
  879. 35:52the most appropriate description
  880. 35:54regarding policy gradient methods?
  881. 35:57A: Always combine decision trees for
  882. 35:59classification.
  883. 36:01B: Parameterize the policy and directly
  884. 36:03update it in the direction that
  885. 36:05increases expected reward. C: Find only
  886. 36:08the axes where the variance of the
  887. 36:09acquired quantities is maximized.
  888. 36:12D: Learn multiple decision trees using
  889. 36:14bootstrap sampling of training data.
  890. 36:25The correct answer for Question 35 is B
  891. 36:27: Parameterize the policy and directly
  892. 36:29update it in the direction that
  893. 36:31increases expected reward. Policy
  894. 36:33gradient methods are techniques for
  895. 36:34optimizing by parameterizing the policy
  896. 36:36used to select actions itself.
  897. 36:38REINFORCE is a representative policy
  898. 36:40gradient method that uses episode
  899. 36:42rewards.
  900. 36:43A and D describe decision trees and
  901. 36:45bagging, while C describes PCA.
  902. 36:49Question 36: Which of the following is
  903. 36:51the most appropriate description of the
  904. 36:53actor-critic method?
  905. 36:56A: The actor defines the reward, and the
  906. 36:58critic deletes the dataset.
  907. 37:01B: The actor and the critic always play
  908. 37:03the same role.
  909. 37:05C: The actor learns the action policy,
  910. 37:07and the critic supports learning by
  911. 37:09evaluating values.
  912. 37:11D: The actor performs clustering, and
  913. 37:13the critic performs dimensionality
  914. 37:15reduction.
  915. 37:24The correct answer for Question 36 is C
  916. 37:26: The actor learns a policy for actions
  917. 37:28, and the critic supports learning by
  918. 37:30evaluating values.
  919. 37:33In the actor-critic method, the actor
  920. 37:35learns the policy for choosing actions,
  921. 37:37while the critic estimates state values
  922. 37:39to assist in updating the actor.
  923. 37:42It is a method that combines value-based
  924. 37:44and policy-based approaches.
  925. 37:46A incorrectly describes the roles. B
  926. 37:48contradicts the point that the roles are
  927. 37:50divided.
  928. 37:50D is confused with unsupervised learning
  929. 37:53.
  930. 37:54Question 37: Which is the most
  931. 37:56appropriate task for applying
  932. 37:58reinforcement learning?
  933. 38:00A: A process that simply follows a fixed
  934. 38:02rule, such as always turning on the heat
  935. 38:04when the temperature is low.
  936. 38:07B: Classifying images using only
  937. 38:09provided labels.
  938. 38:11C: Compressing data into two dimensions
  939. 38:14while preserving variance.
  940. 38:16D: A robot improving its movement
  941. 38:18strategy through trial and error and
  942. 38:20rewards.
  943. 38:29The correct answer for Question 37 is D
  944. 38:31: A robot improving its movement
  945. 38:33strategy through trial and error and
  946. 38:35rewards.
  947. 38:36Reinforcement learning is suitable for
  948. 38:38tasks where actions affect the
  949. 38:39environment, and rewards obtained as a
  950. 38:41result are used to improve continuous
  951. 38:43decision-making.
  952. 38:45A is a simple control rule, B is
  953. 38:47supervised classification, and C is
  954. 38:49dimensionality reduction such as PCA.
  955. 38:53Question 38: What is the most
  956. 38:54appropriate description of the basics of
  957. 38:56holdout validation?
  958. 38:58A: Performing training and final
  959. 39:00evaluation using only the same data.
  960. 39:03B: Dividing data into training and
  961. 39:05evaluation sets, then checking
  962. 39:07performance with unused data.
  963. 39:10C: Making random predictions without
  964. 39:12performing model training.
  965. 39:14D: Adding the correct answers from the
  966. 39:16evaluation data to the training features
  967. 39:18.
  968. 39:27The correct answer for Question 38 is B
  969. 39:29: Dividing data into training and
  970. 39:31evaluation sets, then checking
  971. 39:33performance with unused data.
  972. 39:36Holdout validation is a method of
  973. 39:37splitting data into training and
  974. 39:39evaluation sets to estimate
  975. 39:40generalization performance using data
  976. 39:42not used for training.
  977. 39:45A easily overlooks overfitting to
  978. 39:46training data, and C does not learn the
  979. 39:48model.
  980. 39:50D causes data leakage.
  981. 39:54Question 39: What is the most
  982. 39:55appropriate description of K-fold
  983. 39:57cross-validation?
  984. 39:59A: Dividing data into K partitions, and
  985. 40:02for each iteration, using one partition
  986. 40:04for validation and the rest for training
  987. 40:06to repeat the evaluation.
  988. 40:08B: Training K times without using any
  989. 40:10validation data at all.
  990. 40:12C: Using the same partition for both
  991. 40:14training and validation every time.
  992. 40:17Always ensure the sum of the K
  993. 40:19evaluation metrics equals 100 points.
  994. 40:30The correct answer to Question 39 is A:
  995. 40:32Divide the data into K parts, and repeat
  996. 40:34the evaluation using one part for
  997. 40:36validation and the rest for training
  998. 40:38each time.
  999. 40:40In K-fold cross-validation, data is
  1000. 40:42split into K parts, and we train and
  1001. 40:44evaluate K times while rotating the
  1002. 40:46validation partition, then aggregate the
  1003. 40:49results.
  1004. 40:50B does not satisfy the purpose of
  1005. 40:52cross-validation. C does not separate
  1006. 40:54training and validation within the same
  1007. 40:56iteration. D is an incorrect rule
  1008. 40:58regarding evaluation metrics.
  1009. 41:00Question 40: Which statement best
  1010. 41:02describes the relationship between
  1011. 41:04training error and generalization error
  1012. 41:06?
  1013. 41:07A: If the training error is low, the
  1014. 41:09error on unknown data will also be low.
  1015. 41:12B: Both are always identical by
  1016. 41:14definition.
  1017. 41:15C: Even if training error is small,
  1018. 41:18generalization error on unknown data may
  1019. 41:20be large.
  1020. 41:21D: Generalization error does not need to
  1021. 41:23be considered for trained models.
  1022. 41:34The correct answer to Question 40 is C:
  1023. 41:36Even if training error is small,
  1024. 41:38generalization error on unknown data may
  1025. 41:40be large.
  1026. 41:42If a model overfits to the training data
  1027. 41:44, it may have low training error but
  1028. 41:45fail to perform well on unknown data.
  1029. 41:48This is a classic case of overfitting.
  1030. 41:51A and B equate training performance with
  1031. 41:53generalization performance. D denies the
  1032. 41:55importance of generalization performance
  1033. 41:57, which is vital in real-world
  1034. 41:58operations.
  1035. 41:59Question 41: Which of the following
  1036. 42:02corresponds to a false positive (FP) in
  1037. 42:04binary classification?
  1038. 42:07A: Actually positive and predicted as
  1039. 42:09positive.
  1040. 42:11B: Actually negative and predicted as
  1041. 42:13negative.
  1042. 42:15C: Actually positive but predicted as
  1043. 42:17negative.
  1044. 42:19D: Actually negative but predicted as
  1045. 42:21positive.
  1046. 42:31The correct answer to Question 41 is D:
  1047. 42:33Actually negative but predicted as
  1048. 42:35positive.
  1049. 42:38A false positive is a case where the
  1050. 42:40reality is negative, but the model
  1051. 42:42predicts it as positive.
  1052. 42:44FP affects calculations such as
  1053. 42:46precision.
  1054. 42:48A is a true positive (TP), B is a true
  1055. 42:51negative (TN), and C is a false negative
  1056. 42:54(FN).
  1057. 42:56Question 42.
  1058. 42:58Out of 20 cases that a classification
  1059. 43:00model predicted as positive, 15 were
  1060. 43:03truly positive.
  1061. 43:05What is the precision in this case?
  1062. 43:10A: 60%, B: 75%, C: 80%, D: 93.75%
  1063. 43:27The correct answer to Question 42 is B:
  1064. 43:3175%.
  1065. 43:32Precision is the ratio of actual
  1066. 43:34positives among those predicted as
  1067. 43:35positive.
  1068. 43:3715 divided by 20 equals 0.75. Therefore
  1069. 43:41, it is 75%. A, C, and D all differ from
  1070. 43:44the result of calculating 15 divided by
  1071. 43:4620.
  1072. 43:47Question 43.
  1073. 43:49There are 40 cases that are actually
  1074. 43:51positive, and of those, the model
  1075. 43:53correctly predicted 30 as positive.
  1076. 43:56What is the recall in this case?
  1077. 44:01A: 75%B: 60%C: 80%D: 90%
  1078. 44:17The correct answer to Question 43 is A,
  1079. 44:2075%.
  1080. 44:23Recall is the proportion of actual
  1081. 44:24positives that were correctly predicted
  1082. 44:26as positive.
  1083. 44:2930 divided by 40 equals 0.75, so it is
  1084. 44:3375%. B, C, and D are all different from
  1085. 44:35the result of 30 divided by 40.
  1086. 44:38Question 44. A binary classification
  1087. 44:42model has a precision of 0.8 and a
  1088. 44:44recall of 0.5.
  1089. 44:46Which value is closest to the F1-score,
  1090. 44:48the harmonic mean?
  1091. 44:52A: 0.4 B: 0.5 C: 0.615 D: 0.8
  1092. 45:08The correct answer to Question 44 is C,
  1093. 45:120.615.
  1094. 45:13The F1-score is 2 times precision times
  1095. 45:15recall divided by precision plus recall
  1096. 45:18.
  1097. 45:192 times 0.8 times 0.5 divided by 1.3
  1098. 45:22equals approximately 0.615. A, B, and D
  1099. 45:25do not match the result of the harmonic
  1100. 45:27mean calculation.
  1101. 45:30Question 45. In a binary classification
  1102. 45:33task using data where 99%of the total is
  1103. 45:35negative, which is the most appropriate
  1104. 45:37evaluation if the model always predicts
  1105. 45:40negative?
  1106. 45:42A: Since the accuracy is high, the
  1107. 45:44performance in finding positives must
  1108. 45:46also be high.
  1109. 45:48B: Since the accuracy is high, there are
  1110. 45:51no false negatives.
  1111. 45:53C: The accuracy will always be 50%.
  1112. 45:57D: Even if the accuracy is 99%, the
  1113. 46:00recall for positives will be 0%.
  1114. 46:12The correct answer to Question 45 is D:
  1115. 46:15Even if the accuracy is 99%, the recall
  1116. 46:18for positives will be 0%.
  1117. 46:21If you always predict negative, you get
  1118. 46:2499%of the total correct, but since you
  1119. 46:26find zero actual positives, the recall
  1120. 46:29is 0%.
  1121. 46:30In class imbalance, the selection of
  1122. 46:32metrics is important. A and B evaluate
  1123. 46:35the positive detection ability
  1124. 46:36incorrectly.
  1125. 46:38C ignores the condition that 99%are
  1126. 46:40negative.
  1127. 46:43Question 46. Which is the most
  1128. 46:45appropriate explanation for the ROC
  1129. 46:47curve and AUC?
  1130. 46:50A: The ROC curve shows the relationship
  1131. 46:52between false positive rate and true
  1132. 46:54positive rate, and AUC is the area under
  1133. 46:57that curve.
  1134. 46:58B: The ROC curve only shows the
  1135. 47:00residuals of regression problems.
  1136. 47:03C: AUC represents the training time of
  1137. 47:05the classification model itself.
  1138. 47:08D: The ROC curve does not change when
  1139. 47:10the prediction probability threshold is
  1140. 47:12altered.
  1141. 47:21The correct answer to Question 46 is A:
  1142. 47:24The ROC curve shows the relationship
  1143. 47:26between false positive rate and true
  1144. 47:28positive rate, and AUC is the area under
  1145. 47:30that curve.
  1146. 47:32The ROC curve shows the relationship
  1147. 47:34between the false positive rate and the
  1148. 47:36true positive rate as the classification
  1149. 47:38threshold changes, and the AUC is the
  1150. 47:40area under that curve.
  1151. 47:42B is a confusion with regression
  1152. 47:43residuals. C refers to a different
  1153. 47:45evaluation target. D rejects the ROC
  1154. 47:48concept of plotting points by varying
  1155. 47:50the threshold.
  1156. 47:53Question 47: If a regression model's
  1157. 47:56predicted value is 24 and the actual
  1158. 47:58value is 17, what is the Mean Squared
  1159. 48:01Error (MSE)?
  1160. 48:05A: 2, B: 4, C: 5, D: 10
  1161. 48:20The correct answer to Question 47 is C:
  1162. 48:225. The errors are 1 and -3, and the
  1163. 48:25squared errors are 1 and 9.
  1164. 48:27Averaging them gives (1 + 9)/2 = 5, so
  1165. 48:31the MSE is 5.
  1166. 48:33A, B, and D do not match the average of
  1167. 48:35the squared errors.
  1168. 48:38Question 48: Which is the most
  1169. 48:40appropriate description of RMSE and MAE
  1170. 48:42in regression models?
  1171. 48:46A: RMSE is the mean of errors, and MAE
  1172. 48:48is the square root of errors.
  1173. 48:53B: RMSE is the root mean squared error,
  1174. 48:55and MAE is the mean absolute error.
  1175. 48:59C: RMSE and MAE are both precision
  1176. 49:01metrics for classification models.
  1177. 49:05D: RMSE is always smaller than MAE.
  1178. 49:18The correct answer to Question 48 is B:
  1179. 49:20RMSE is the root mean squared error, and
  1180. 49:23MAE is the mean absolute error.
  1181. 49:27RMSE is the square root of MSE, while
  1182. 49:29MAE is the average of absolute errors.
  1183. 49:33For the same data and error units, RMSE
  1184. 49:35is greater than or equal to MAE and
  1185. 49:36gives relatively more weight to large
  1186. 49:38errors.
  1187. 49:40A is incorrect in its definition. C
  1188. 49:41confuses regression evaluation with
  1189. 49:43classification evaluation, and D has the
  1190. 49:46size relationship between RMSE and MAE
  1191. 49:48reversed.
  1192. 49:51Question 49: Which is the most
  1193. 49:52appropriate common concept for the
  1194. 49:55Akaike Information Criterion (AIC) and
  1195. 49:57the Bayesian Information Criterion (BIC
  1196. 49:59)?
  1197. 50:02A: More complex models are
  1198. 50:03unconditionally rated higher.
  1199. 50:06B: Unlabeled data must be classified
  1200. 50:09into clusters.
  1201. 50:10C: Models are evaluated solely based on
  1202. 50:13prediction accuracy.
  1203. 50:15D: Models are compared by considering
  1204. 50:17both goodness-of-fit and model
  1205. 50:19complexity.
  1206. 50:28The correct answer to Question 49 is D:
  1207. 50:30Models are compared by considering both
  1208. 50:32goodness-of-fit and model complexity.
  1209. 50:35AIC and BIC compare models by
  1210. 50:37considering not only the fit to the data
  1211. 50:39but also penalties for model complexity
  1212. 50:41, such as the number of parameters.
  1213. 50:44Generally, smaller values are preferred
  1214. 50:46. A is the opposite of the penalty for
  1215. 50:48complexity. B differs in that it
  1216. 50:50concerns clustering, and C is limited to
  1217. 50:52classification accuracy.
  1218. 50:55Question 50: Which of the following is
  1219. 50:57the most appropriate description of
  1220. 50:59Occam's razor in machine learning model
  1221. 51:01selection?
  1222. 51:04A: If explanatory power is similar,
  1223. 51:06prioritize considering a simpler model.
  1224. 51:10B: The more parameters there are, the
  1225. 51:12higher the generalization performance
  1226. 51:14always becomes. C: If the model is made
  1227. 51:16more complex, validation data becomes
  1228. 51:19unnecessary. D: Outliers included in the
  1229. 51:21training data must always be deleted.
  1230. 51:32The correct answer to Question 50 is A:
  1231. 51:35If explanatory power is similar,
  1232. 51:37prioritize considering a simpler model.
  1233. 51:40Occam's razor is the concept that if
  1234. 51:42explanatory power is similar, one should
  1235. 51:44not unnecessarily increase complex
  1236. 51:46processes. It serves as a guideline that
  1237. 51:48emphasizes simplicity in model selection
  1238. 51:49as well. B is incorrect because
  1239. 51:51complexity does not guarantee
  1240. 51:52generalization performance. C does not
  1241. 51:54make validation unnecessary. D is not a
  1242. 51:57mandatory rule for outlier processing.
  1243. 52:00Question 51: The accuracy on training
  1244. 52:03data was 99%, while the accuracy on
  1245. 52:06validation data was 65%.
  1246. 52:09Which of the following is the most
  1247. 52:10appropriate phenomenon to suspect first
  1248. 52:12?
  1249. 52:14A: Automatic conversion to unsupervised
  1250. 52:17learning, B: Convergence of K-means, C:
  1251. 52:19Drop in generalization performance due
  1252. 52:22to overfitting, D: Change in the
  1253. 52:24definition of the ROC curve.
  1254. 52:34The correct answer to Question 51 is C:
  1255. 52:37Drop in generalization performance due
  1256. 52:39to overfitting. Since it shows good
  1257. 52:41performance on training data but worsens
  1258. 52:43significantly on validation data,
  1259. 52:45overfitting to the training data is
  1260. 52:46suspected.
  1261. 52:47Additional investigation into
  1262. 52:49differences in data distribution or
  1263. 52:51leakage is also necessary. A, B, and D
  1264. 52:53are not the primary phenomena directly
  1265. 52:55suspected from this performance gap.
  1266. 52:59Question 52: Which evaluation metric
  1267. 53:01should be particularly prioritized if
  1268. 53:03you want to minimize misses (false
  1269. 53:05negatives) in primary cancer screening?
  1270. 53:09A: Inference time only.
  1271. 53:12B: Recall, which equals sensitivity.
  1272. 53:16C: Mean Absolute Error (MAE).
  1273. 53:20D: Akaike Information Criterion (AIC)
  1274. 53:23only.
  1275. 53:32The correct answer to Question 52 is B:
  1276. 53:35Recall, which equals sensitivity.
  1277. 53:38Prioritizing recall, which is the rate
  1278. 53:40of correctly detecting actually positive
  1279. 53:42individuals as positive, leads to
  1280. 53:43evaluation that focuses on reducing
  1281. 53:45false negatives. However, false
  1282. 53:47positives and operational constraints
  1283. 53:49should also be considered together.
  1284. 53:51A does not directly represent the miss
  1285. 53:52rate.
  1286. 53:54C is a regression metric, and D is a
  1287. 53:56model selection metric; neither is
  1288. 53:58direct for evaluating misses in positive
  1289. 54:00detection.

About this transcript

This page contains the full transcript of 【G検定 聞き流し 練習問題】③機械学習の概要|練習問題52問・解答解説付き ~2026年試験対応~ #ワイズデータ by ワイズデータ [聞き流しで受かるIT・データ資格チャンネル], generated from the public captions YouTube serves with the video. The transcript has 6,665 words across 1,289 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.