Principal component analysis draws every row of your table as one point on a map. Rows that behave alike across all of their numbers land close together. Rows that behave differently land far apart. The horizontal axis is the single strongest pattern in your data, and the vertical axis is the strongest remaining pattern that is independent of the first one.

Three things are worth spotting on any map: clumps of points, which are groups of similar rows; lonely points far from everything, which are unusual rows worth checking; and the overall direction of spread, which shows where most of the variation lives.

In the map plus drivers view, each arrow is one of your original columns. An arrow points toward the side of the map where that column has high values. Long arrows matter a lot for the visible plane, short arrows matter little. Two arrows pointing the same way are columns that rise and fall together, while arrows pointing in opposite directions are columns that trade off against each other.

The variance chart shows how much of the total variation each pattern captures, together with the running total. When the running total is high after two or three patterns, a flat map already tells most of the story. The feature compass places every column inside a circle: columns near the rim are represented well by the current plane, columns near the center are mostly living in other patterns. The heatmap shows every column against every computed pattern at once, so you can name each pattern by the columns that load on it most strongly.

Features are standardized to unit variance by default, robust scaled by mean centering plus MAD spread when outliers dominate, or centered only on request. The decomposition is an exact eigendecomposition of the covariance matrix by cyclic Jacobi rotations when the smaller data dimension allows it, and a matrix free orthogonal iteration with Rayleigh Ritz refinement for very large tables. Explained variance uses the full trace, so percentages are exact in both paths. Component signs are fixed so that the largest absolute loading of each component is positive.

The correlation circle shows the correlation between each variable and each component. Group ellipses are 95 percent normal confidence regions from each group covariance. Outlier scores use Hotelling T squared over the retained components against a chi square threshold. Cluster search runs k-means with a k-means++ start and a fixed seed on the leading component scores, and selects k by the mean silhouette, reporting the Calinski-Harabasz score as a second separation signal. Suggested component counts report the Kaiser rule and the broken stick model alongside cumulative variance. All computation happens locally and results are fully reproducible. The 3D scene is a local WebGL visualization with a canvas fallback; the quality mode changes only rendering density, never the PCA statistics.

Your data is processed entirely inside your browser tab. The page makes no network requests with your data, there is no upload and no server side processing, and closing the tab removes everything. Very wide or very long tables are handled with progressive computation, and the display thins extremely dense point clouds while statistics always use every row. PCA describes structure and correlation; it does not by itself prove causation.