IFN580 Week 7 Unsupervised Learning Flashcards
(16 cards)
What is ‘curse of dimensionality’
when dimensionality increases, data becomes sparse and requires more data to learn properly
Both supervised and unsupervised require at least one:
input attribute
Supervised learning differs from unsupervised learning in that supervised learning
requires:
at least one attribute
When selecting features, are the attribute values changed in any way?
No, feature selection is simply selecting/excluding features without any change.
Both feature selection and Principal Component Analysis reduce the number of features for a given dataset. How does the process differ between these two techniques?
PCA creates new features
What is one approach for selecting the optimal number of components for PCA?
Graph the sum of total variance of all components and look for elbow point
When does PCA work optimally?
PCA works best when the data follows a normal distribution.
What is a scenario where PCA may not perform adequately?
It is sensitive to outliers and may not work optimally if data is sparser.
t-SNE and UMAP are ___________ methods for ___________.
machine learning, dimensionality reduction
t-SNE can only compute up to ? components.
3
What is perplexity?
a hyperparameter is t-sne that controls how many neighbours each point considers
Which hyperparameters does UMAP use?
“nearest neighbours” (n_neighbors): Controls how UMAP balances local
structure versus global structure in the data.
“minimum distance” (min_dist), which controls the distance between the
points in the low dimensional representation
What algorithm does UMAP use for optimisation?
Uses deterministic graph Laplacian-based optimisation,
What algorithm does t-SNE use for optimisation?
Uses stochastic gradient descent
Can you combine multiple dimensionality reduction approaches?
Yes, use PCA to extract greater number of components and then UMAP or t-SNE to reduce to 2 components
What is PCA?
a technique used to reduce the dimensionality of a dataset but preserve the variance
