A cluster is a tight bundle; an outlier is a polyline that breaks from one. Two emphases of the same perceptual structure.
Similar points produce similar polylines. Concentrated data forms a visible bundle: bundle width reflects group spread, bundle path reflects group center.
A cluster is a tight bundle. An outlier is a polyline that peels away at specific axes. Both are readings of the same structure: bundles when looking for groups, deviations when looking for what doesn't fit.
The geometry follows precise rules. An elliptical cluster between two variables produces a hyperbolic signature between the corresponding parallel axes. Because each axis is one dimension, a deviation in specific dimensions maps to deviations at specific axis pairs.
The UCI Wine dataset contains 178 samples of wine from three different grape cultivars grown in the same region of Italy. Each sample is measured across 13 chemical properties. The three cultivars produce three distinct bundles that separate across several dimensions, most clearly on flavanoids, color intensity, proline, and od280/od315 dilution.
A scatterplot matrix would need 28 pairwise panels for the same 8 variables. Here all eight dimensions appear at once, and the three-class structure is visible as three bundles traveling in different directions.
Figure 1. The UCI Wine dataset (178 rows, 8 most discriminating features). Three grape cultivars form three distinct bundles. Cultivar 1 (blue) runs high on proline and flavanoids. Cultivar 3 (orange) is high on color intensity and low on od280/od315. Cultivar 2 (green) occupies the middle ground. Hover any polyline to isolate it.
Points clustered in an ellipse between two adjacent variables map to a family of line segments whose envelope is a hyperbolic "bowtie."
The boundary of the ellipse maps to a generalized hyperbola (an h-star) via the envelope construction: each boundary point produces a segment, and as the point traces the ellipse, the family of segments has the bowtie as its envelope.
Recall that for a line , the convergence point in parallel coordinates sits at . When data points cluster, their pairwise slopes cluster too, producing a concentration of convergence points. That density of convergence points is what forms the bowtie waist. The tighter the correlation, the narrower the waist.
Drag the correlation slider to change the ellipse eccentricity. Press Animate to watch the tangent point circle the ellipse.
Figure 2. An elliptical cluster (left) and its bowtie signature (right). The green dot traces the ellipse boundary; the corresponding tangent line sweeps through the parallel coordinates. The envelope of all tangent lines forms the bowtie. Narrow waist = strong correlation.
Bowtie width encodes correlation. A perfectly correlated cluster (r = 1) produces line segments that all pass through a single convergence point, collapsing the bowtie to a line. As correlation decreases, the convergence points spread, and the bowtie widens.
Bowtie orientation encodes sign. Positive correlation places the bowtie waist outside the axes (to the right of the right axis). Negative correlation places it between the axes, producing a crossing pattern.
Multiple clusters, multiple bowties. If the data contains several clusters, each contributes its own bowtie. Their separation in parallel coordinates reflects how well the clusters separate in data space.
An outlier typically has one or two anomalous values while the rest are ordinary. In parallel coordinates, this shows up as a polyline that follows the bundle for most axes but diverges sharply at the anomalous segments — the geometry tells you which dimensions are responsible.
Below, 80 points drawn from a multivariate normal in six dimensions form a tight cluster (every polyline hugs the bundle) with one injected outlier whose per-dimension deviations can be varied.
Drag the red dots on the axes or use the sliders to adjust the outlier's deviation in each dimension.
Figure 3. A cluster of 80 points (blue) and one injected outlier (red). The sliders control how many standard deviations the outlier deviates in each dimension. Watch the red polyline separate from the bundle only at the axes where you push the slider.
Move a single slider and the red line peels away from the bundle at that axis alone. Push two adjacent sliders and the deviation spans a wider segment. Push two non-adjacent sliders and you get two isolated breaks in an otherwise conforming polyline. The pattern of deviation is the outlier's signature: it tells you exactly which dimensions are responsible.
Intuition is useful, but we can make "deviation from the bundle" precise. The Mahalanobis distance measures how far a point lies from the center of a distribution, accounting for the shape and orientation of that distribution:
Here is the point, is the distribution mean, and is the covariance matrix. The inverse-covariance weighting means deviation along a tightly clustered dimension counts more than deviation along a spread-out one. In parallel coordinates, we can decompose this distance into per-axis contributions and see which axes drive the total.
Click any polyline to select it and see its distance breakdown.
Figure 4. Manufacturing measurements across six dimensions. Click a polyline to see its Mahalanobis distance decomposed by dimension. The dashed gray line is the population mean; connecting strokes on each axis show per-dimension deviation; bars below quantify the squared z-scores; the table gives the numerical breakdown. The largest contributors are highlighted.
Reading the distance breakdown. A high total Mahalanobis distance does not tell you where the problem is. The per-dimension decomposition does. An outlier at D = 8.3 might have all its distance concentrated in a single dimension (a spike defect) or spread across three dimensions (a drift defect). The treatment differs.
Contextual outliers. A point can be an outlier relative to its cluster but not relative to the global population. If your data has multiple clusters, a point at the edge of Cluster A might sit squarely inside Cluster B, its Mahalanobis distance is only meaningful with respect to the right reference distribution.
A bundle establishes an expectation; a deviation violates it at a specific axis pair. Bundle width sets the threshold for deviation; the axes where the break occurs identify which dimensions carry the anomaly.
One tool serves both: brushing. Drag a range on any axis to highlight polylines passing through it. The highlighted subset reveals cluster structure across the other axes. Confine every axis to its normal range and the polylines that fall outside become outliers.
The wine data without class colors. Brushing flavanoids high picks out cultivar 1; color intensity high picks out cultivar 3. The three-cluster structure can be rediscovered through interaction. The next explainer formalizes brushing geometrically.
Drag vertically on any axis to brush. Brushed polylines highlight, others dim. Click outside the brush on the same axis to clear it.
Figure 5. The wine dataset without class labels. Brush any axis to select a range. The highlighted polylines reveal latent cluster structure across all dimensions. Try brushing flavanoids high (top third) or color intensity high to isolate specific cultivars — or brush every axis to its "normal" middle range and watch the remaining polylines stand out as deviations.
Finding clusters and diagnosing outliers both reduce to reading the bundle. The bowtie explains the local geometry; the Mahalanobis decomposition quantifies and locates the deviation; brushing makes the reading interactive. The next explainer treats the cluster envelope — the h-star — as a geometric object whose interior defines membership.