Quiz dotyczący nauki nadzorowanej

Odpowiedzi i wyjaśnienia pojawiają się w miarę rozwiązywania quizu.

1. Co w przypadku uczenia się nadzorowanego sprawia, że ​​zbiór danych szkoleniowych jest „oznaczony”?
   • Each example has been sorted into folders according to its file type and size
   • Each example is paired with a target output for the model to learn
   • Each example has been compressed so the model can read it in less time
   • Each example includes a timestamp recording exactly when it was first collected
2. Który typ problemu z nadzorowanym uczeniem się rozwiązuje model przewidujący, czy wiadomość e-mail jest spamem, czy nie?
   • Regression
   • Klasyfikacja
   • Dimensionality reduction
   • Clustering
3. Czego przykładem jest przewidywanie ceny sprzedaży domu na podstawie jego wielkości i lokalizacji?
   • Clustering, because similar houses are grouped together by area
   • Regression, because the output is a continuous number
   • Classification, because houses belong to distinct neighborhoods
   • Reinforcement learning, because prices reward sellers
4. Dlaczego praktycy ukrywają oddzielny zestaw testów, którego model nigdy nie widzi podczas szkolenia?
   • To give the model extra examples once its training accuracy peaks
   • To store a backup copy in case the training data is lost
   • To estimate how well the model performs on new, unseen data
   • To make training run faster by using fewer examples
5. When should preprocessing steps such as fitted scaling be learned for a supervised-learning evaluation?
   • From all rows before splitting so the test set helps estimate them
   • From test examples only
   • From the training partition, then applied to validation and test examples
   • From whichever partition gives the best final score
6. The guide’s classifier catches 30 of 40 urgent messages and flags 20 ordinary messages. Which recall and precision values follow?
   • Recall 75%; precision 60%
   • Recall 60%; precision 75%
   • Recall 100%; precision 50%
   • Recall 30%; precision 20%
7. A dataset has repeated records for the same person. What risk can a random row split create?
   • The model is guaranteed to have low recall
   • Related records can appear in both training and test sets, leaking information
   • The target labels automatically become numerical
   • The test set is no longer allowed to contain labels
8. What can changing a classification threshold do, according to the guide?
   • Change the tradeoff between false positives and false negatives without changing learned parameters
   • Change the training labels into features
   • Guarantee both error types become zero
   • Replace the need to evaluate model performance