Machine Learning: Unsupervised Learning, Neural Networks, and Reinforcement Learning

A variable that is directly measured and observed in the original dataset

A variable that combines two or more variables that are statistically strongly related to each other

A variable that has zero correlation with all other variables in the dataset

A variable that is always standardized to have a unit variance

The number of features in the original dataset

The direction of maximum variance in the data

The proportion of total variance in the initial data that is explained by each eigenvector

The number of principal components that must be retained in the model

Randomly, to ensure unbiased selection of components

From lowest to highest according to their eigenvalues

From highest to lowest according to their eigenvalues (i.e., in terms of usefulness in explaining total variance)

Alphabetically based on the names of the original features

To visualize the clusters formed in a dataset

To determine the optimal number of principal components to retain

To identify outliers in the original dataset

To compute the eigenvalues of the covariance matrix

The mean of each series is 0 and the standard deviation is 1

All features are converted to categorical variables

The total variance of the dataset is set to 0

The covariance matrix becomes a vector of zeros

It requires the data to be labeled in order to function

It always retains all original features, offering no dimensionality reduction

The resulting principal components typically cannot be easily labeled or directly interpreted by the analyst

It only works with datasets containing fewer than five features

High intra-cluster distance and low inter-cluster distance

Low intra-cluster distance (cohesion) and high inter-cluster distance (separation)

Equal variance within all clusters regardless of cluster size

All clusters having the same number of observations

Manhattan distance

Cosine similarity

Euclidean distance

Hamming distance

Determined automatically by the algorithm without any user input

A model hyperparameter that must be specified before running the algorithm

Always equal to the number of features in the dataset

Equal to the sample size of the dataset

All centroids coincide at the origin

The maximum number of iterations has been reached

No observation is reassigned to a new cluster (i.e., centroids no longer need to be recalculated)

The variance within each cluster is exactly zero

The observation that is farthest from all other observations in a cluster

The average value of the observations assigned to a cluster

The first observation included in the dataset

The mode of all categorical features within a cluster

The final cluster assignment can depend on the initial location of the centroids

It cannot handle datasets with more than 100 observations

It always produces clusters of equal size

It requires labeled data for training

Starts with all observations in a single cluster and progressively partitions them

Begins with each observation as its own cluster and iteratively merges the two closest clusters

Requires the analyst to pre-specify the exact number of clusters before any computation

Can only be applied to datasets with exactly two features

It starts with all observations in one cluster and progressively partitions them into smaller clusters

It merges the two closest clusters until only one cluster remains

It begins with k randomly chosen centroids and assigns each observation to the nearest centroid

It requires labeled training data for each observation

The number of observations contained in the cluster

The distance (dissimilarity) between the two clusters being combined

The number of features used to construct the dendrogram

The total variance explained by the clusters

Only an input layer and an output layer, with no intermediate processing

An input layer, one or more hidden layers, and an output layer

Multiple input layers and a single hidden layer, with no output layer

A single layer of nodes connected in a tree structure

Always produce a strictly linear output regardless of input

Transform the total net input into the final output of the node in a non-linear way

Compute the eigenvalues of the input feature matrix

Select the most important features for the model

Forward propagation

Backward propagation

Dimensionality reduction

Cluster assignment

A neural network with only one hidden layer

A neural network with many hidden layers (at least 2, often more than 20)

A neural network that uses only linear activation functions

A neural network that does not require any training

Requires direct labeled data for every observation in the training set

Learns by having the agent test new actions, observe its environment, and reuse previous experiences through trials and errors

Always produces identical results regardless of the sequence of actions taken

Can only be applied to numerical prediction problems, not classification