Pre-model Interpretability and Explainability

Vocabulary flashcards covering key terms and definitions in pre-modeling explainability, exploratory data analysis, statistical measures, and feature engineering.

Pre-Modeling Explainability

The process of gaining insights into datasets before model building using techniques to understand data quality, distributions, and structure to support model explainability.

Exploratory Data Analysis (EDA)

A collection of statistical techniques and visualizations used to analyze data distribution, quality, relationships between features, and potential model bias.

Univariate Analysis

Statistical and graphical techniques focused on analyzing a single feature at a time to examine characteristics such as spread, range, and central tendency.

Multivariate Analysis

Analysis techniques used when multiple features interact, capturing local and global relationships across continuous and categorical data.

Mean

A central value around which data is distributed, calculated mathematically as $\mu = \frac{1}{N} \sum_{i=1}^N x_i$.

Variance

The average squared deviation of data points from the mean, represented as $\sigma^2 = \frac{1}{N} \sum_{i=1}^N (x_i - \mu)^2$.

Standard Deviation

The typical distance of data points from the mean, calculated as the square root of variance $\sigma = \sqrt{\sigma^2}$.

Skewness

A statistical measure that quantifies the asymmetry of a distribution relative to its mean, calculated as $\gamma_1 = \frac{\mu_3}{\sigma^3}$.

Kurtosis

A statistical measure representing the tailedness of a distribution, calculated as $\gamma_2 = \frac{\mu_4}{\sigma^4}$.

Excess Kurtosis

A statistic comparing a distribution's tailedness to a normal distribution, defined as $\gamma_2 - 3$.

Pivot Table

A summarization tool that reorganizes and aggregates data by grouping rows and columns and applying functions such as sum or mean.

Cross-Tabulation

A quantitative method (contingency table) used to analyze the relationship between two or more categorical variables using frequency counts or percentages.

Box Plot

A visual representation summarizing a dataset using five metrics: median (Q2), quartiles (Q1, Q3), whiskers ($1.5 \times IQR$), and individual outliers.

Distribution Plot

A chart visualizing data distribution using histograms for binned frequency counts and Kernel Density Estimation (KDE) for smooth density estimation.

Violin Plot

A visualization technique combining a boxplot with a Kernel Density Estimation (KDE) plot to showcase data density and shape.

Joint Distribution Plot

A bivariate visual method used to reveal local interactions between two continuous features using density estimates, histograms, or hexbin plots.

Heatmap

A graphical encoding of a pairwise correlation matrix where feature associations are colored based on correlation measures.

Spearman Rank Correlation

A non-parametric correlation metric that assesses monotonic relationships using the ranks of observations instead of raw values.

Parallel Coordinate Plot

A visualization technique that projects high-dimensional data onto two dimensions by drawing individual data points as polylines across parallel vertical axes.

Principal Component Analysis (PCA)

A linear dimensionality reduction technique that creates orthogonal principal components as linear combinations of original features to maximize captured variance.

Biplot

A composite visualization combining PCA scores and a loading plot to illustrate how raw features influence principal components in direction and magnitude.

t-SNE

t-Distributed Stochastic Neighbor Embedding; a non-linear dimensionality reduction technique that preserves local neighborhood structures using a perplexity parameter.

Isomap

Isometric Mapping; a non-linear dimensionality reduction method based on spectral theory that preserves geodesic distances across a neighborhood graph.

Autocorrelation

A measure of the linear relationship between a time-series variable and its lagged values.

Differencing

A time-series transformation computing the difference between consecutive observations to eliminate non-stationarity.

Topic Modeling

An unsupervised natural language processing approach (such as LDA) that models documents as probability distributions over topics and topics as distributions over terms.

Filter-based Feature Selection

A fast, model-independent feature selection approach that evaluates feature importance using statistical metrics like variance thresholds, Chi-squared, or ANOVA.

Fisher Score (F-score)

A feature selection metric measuring discriminatory power by comparing feature variance between classes to feature variance within classes.

Wrapper-based Feature Selection

A feature selection method that utilizes a machine learning algorithm directly to evaluate and optimize feature subsets based on predictive performance.

Embedded Feature Selection

A feature selection strategy integrated directly into model training, such as regularization methods (L1 Lasso, L2 Ridge) or tree-based splits.