Exploring and Modeling High-Dimensional Data: Glossary

data can be anything - as long as you can represent it in a computer

A dimension is a feature in a dataset - i.e. a column, but NOT an index.

an index is not a dimension

Linear regression models can be used to predict a target variable and/or to reveal relationships between variables

Linear models are most effective when applied to linear relationships. Data transformation techniques can be used to help ensure that only linear relationships are modelled.

Train/test splits are used to assess under/overfitting in a model

Different model evaluation metrics provide different perspectives of model error. Some error measurements, such as R-squared, are not as relevant for explanatory models.

All models are wrong, but some are useful.

Before reading into a model’s estimated coefficients, modelers must take care to account for essential predictor variables

Models that do not account for essential predictor variables can produce distorted pictures of reality due to omitted variable bias and confounding effects.

All models are wrong, but some are useful.

Before reading into a model’s estimated coefficients, modelers must take care to check for evidence of overfitting and assess the 5 assumptions of linear regression.

One-hot enoding, while necesssary, can often produce very sparse binary predictors which have little information. Predictors with very little variability should be removed prior to model fitting.

All models are wrong, but some are useful.

PCA transforms your original data by projecting it into new axes

primary components are orthogonal vectors through your data that explain the most variability

TODO

TODO