Calculator Options
Formula Used
PCA centers each feature before matrix decomposition begins. Standardization also divides values by their sample standard deviation. This process prevents measurement scale from dominating every component.
Standardized value: zᵢⱼ = (xᵢⱼ − μⱼ) / sⱼ
Covariance matrix: C = ZᵀZ / (n − 1)
Eigen equation: Cvₖ = λₖvₖ
Eigenvectors define principal directions across the prepared feature space. Eigenvalues measure variance captured by their corresponding directions. Scores equal prepared observations multiplied by selected eigenvectors.
Component scores: T = ZV
Explained variance: EVRₖ = λₖ / Σλ
Score distance: D² = Σ(tₖ² / λₖ)
How to Use This Calculator
Paste a rectangular dataset into the main input area. Mark headers and identify any optional label column. Confirm every remaining column contains numeric feature values.
Choose standardization when feature units or ranges differ greatly. Select components for each visible chart axis. Enable arrows, labels, ellipses, or grouping when useful.
Press the calculation button to generate results above. Review variance before interpreting distances or visual separation. Download score and loading tables for deeper analysis.
Understanding PCA Scatter Plots
Core Purpose
PCA converts many related features into fewer independent directions. Each direction captures a measurable share of total variance. The first components usually reveal the strongest structural patterns.
A scatter plot maps observations using selected principal component scores. Nearby points have similar combinations of original feature values. Separated groups may indicate classes, segments, or unusual behavior.
Reliable results begin with clean rectangular numeric data. Each row should represent one observation or measured sample. Each numeric column should describe one comparable input feature.
Preparation and Decomposition
Standardization gives every feature a comparable variance before decomposition. This choice helps when columns use different measurement units. Centering alone preserves original variance differences between features.
The covariance matrix summarizes relationships across every prepared feature. Eigenvectors identify directions producing uncorrelated component scores. Eigenvalues quantify variance captured along those discovered directions.
Explained variance percentages support informed component selection. Large early percentages suggest useful compression with limited information loss. Small percentages indicate weaker patterns or noisy dimensions.
Loadings and Biplots
Loadings connect principal components back to original variables. Large absolute loadings identify influential features for each direction. Opposite signs indicate contrasting movement along that component.
Biplot arrows display loading direction beside projected observations. Long arrows represent stronger contributions within the displayed plane. Similar arrow directions suggest positively related original features.
Groups can be colored using an optional label column. This view helps compare known classes against unsupervised structure. Unexpected overlap may expose weak separation or noisy labels.
Outliers and Dimensions
Score distance can flag observations far from the projection center. These points deserve review before removal or special treatment. Outliers may reflect errors, rare cases, or meaningful discoveries.
Confidence ellipses summarize expected spread across the first components. Points outside the boundary may have unusual score combinations. The selected probability changes how strict that boundary becomes.
Three dimensional plotting can reveal structure hidden in two dimensions. However, depth perception may complicate precise visual comparisons. Always compare three dimensional views with variance percentages.
Interpretation and Validation
PCA remains linear and may miss curved relationships. It also depends strongly on scaling and feature preparation choices. Interpret components alongside domain knowledge and validation evidence.
Component scores often support clustering, classification, and exploratory quality checks. They can greatly reduce harmful collinearity before downstream statistical modeling. Retain enough informative components for the intended analytical objective.
Repeated careful experiments strengthen conclusions about stable projection patterns. Compare several scaling methods, feature subsets, and selected component pairs. Document every important choice so others can reliably reproduce the complete analysis.
Frequently Asked Questions
1. What does each PCA point represent?
Each point represents one dataset observation. Its coordinates are principal component scores. Nearby points share similar multivariate feature patterns.
2. Should I standardize my features?
Standardize when features use different units or scales. Centering alone preserves original variance differences. Compare both methods when domain priorities remain uncertain.
3. How many components should I retain?
Review cumulative explained variance and downstream performance. Avoid relying on one universal percentage rule. Keep enough components for your analytical objective.
4. What do loading arrows show?
Arrows show feature direction within the displayed component plane. Longer arrows indicate stronger visible contributions. Similar directions suggest positive feature relationships.
5. Are flagged points definitely incorrect?
No, flags only identify unusual score distances. Review source records before changing anything. Rare observations may contain important information.
6. Can PCA use categorical variables directly?
Standard PCA requires numeric feature columns. Encode categories carefully before analysis. Avoid arbitrary numbers that imply false ordering.
7. Why can component signs change?
Eigenvector signs are mathematically interchangeable. Reversing signs preserves distances and variance. Compare absolute structure instead of sign alone.
8. When is a three dimensional plot useful?
Use three dimensions when the third component adds meaningful variance. Rotate the chart to inspect hidden separation. Confirm patterns with numeric summaries.
9. Does PCA prove clusters exist?
No, PCA only provides a linear projection. Apparent groups require additional validation. Use clustering metrics and domain evidence.