Machine Learning Flashcards
Method for computer-aided, automated learning of models that
represent relationships between the features π1, β¦ , ππ and the
associated target variable π .
Machine Learning
The three types of learning that Machine Learning combines.
Unsupervised learning, supervised learning and reinforcement learning.
The two types of supervised learning tasks.
Classification and regression
A qualitative / categorical supervised learning task is a ___ task.
Classification
Which type of task should we use to predict the outcome of a COVID test (positive or negative), a spam filter, the credit worthiness of a customer, the result of a match (win, tie, lose)?
A Classification task
A quantitative / numerical supervised learning task is a ___ task.
Regression
Which type of task should we use to predict mileage of electric cars, traffic volume at a street segment, expected waiting time (e.g., until e-scooter is fully charged), climate change (incl. global warming, or emission of greenhouse gases), amount of energy that is produced by wind parks?
A Regression task
The two machine learning intentions.
Prediction and explanation
It defines what should be modeled within the ML pipeline, it comprises feature data and target variable.
A task
An algorithm that trains the model based on training data, it encapsulates the actual supervised learning model.
A learner
It is an instantiation of the learner to the data (with model parameters that are optimal for the given data).
A model
(True or false) Hyperparameters are defined within the learner.
True. A learner is a (data independent) general construct defining which ML algorithm (and
hyperparameters) will be used.
Performance measure that quantifies deviation between true and predicted values of the target variable.
Loss function
The aggregation of all losses across (training) data.
Empirical risk
The best parametrization of the learner when training a model.
Risk minimization
MLR3 learner in which for each (test) data point x = (π₯1, β¦ , π₯π), i.e., each observation for which we want to make a prediction, compute the neighborhood ππ (π₯1, β¦ , π₯π) of its π nearest
neighbors.
K Nearest Neighbor Classifier
(True or false) There is no training step for k-NN models, just storing the training data to process it during the predict step.
True
If a model has learned too many details from the training data, which donβt necessarily generalize to unseen data sets, it is called an ____ model.
Overfitted
If a model is too general to capture all the relevant patterns from the data, then it is called an ____ model.
Underfitted
What kind of fit does a model have if it has low error in the training data as well as in the test data.
Underfit
One of the main disadvantages of this learner is that it can be computationally expensive for large datasets as it needs to store a lot of data per observation.
KNN Learner
Trick that supports classifiers as knn and SVMs to map the problem into a higher-dimensional space, then classifies the transformed problem and then use the outcome of the aforementioned classification for the original problem.
The kernel trick
The name given to data in which classes are distributed unevenly.
Imbalanced data.
Two typical countermeasures to handling unbalanced data.
- Over-/undersampling,
- Tuning class thresholds
