AI-basic 2

Supervised Learning involves:

• Training models only on unlabeled data

• Using labeled examples to learn a function or classifier

• Guessing class labels without training

• Ignoring training data

A typical Supervised Learning dataset includes:

• Records with attributes and known class labels

• Only unlabeled raw data

• No attributes, just random numbers

• Class labels hidden from the learning algorithm

Decision Tree Induction:

• Creates a tree by splitting data based on attribute values

• Requires no attributes

• Cannot be pruned once grown

• Ignores impurity measures entirely

An overfitted decision tree:

• Generalizes well to new, unseen data

• Fits training data too closely, harming generalization

• Is always the best model

• Occurs only when trees have no leaves

Precision and Recall measure:

• The same quantity in classification performance

• Different aspects of predictive accuracy (true positives and coverage of actual positives)

• Only false negatives

• Irrelevant metrics in model evaluation

Cross-Validation:

• Uses multiple train/test splits to better estimate model performance

• Trains on the entire dataset without testing

• Is never used in practice

• Is identical to using a holdout set

Support Vector Machines (SVM):

• Find a hyperplane that separates classes with the maximum margin

• Never use kernels

• Do not allow for soft margins

• Are only applicable to linear problems

The k-Nearest Neighbor (kNN) algorithm:

• Builds a model by averaging the target function parameters ahead of time

• Is a lazy learner that classifies based on proximity to known examples

• Uses complex parametric training

• Cannot handle numeric attributes

Ensemble Methods like Bagging and Boosting:

• Combine multiple models to often improve predictive performance

• Always perform worse than single models

• Never use bootstrapped samples or sequential weighting

• Cannot handle classification tasks

F1-Score is:

• The ratio of correct predictions to total predictions

• The harmonic mean of precision and recall

• Ignored in model evaluation

• Always equal to accuracy