figure-design-patterns

Figure & Table Design Patterns

This skill provides a vocabulary of established figure types for ML / computational biology papers. When suggesting visualizations, default to these patterns. Deviations should be justified.

When to Use Which Figure Type

Situation	Default figure	Alternative	Anti-pattern
Compare 3-10 models on 1 metric	Bar chart with error bars	Dot plot	Pie chart
Compare N models on M tasks (N,M > 4)	Heatmap	Multi-panel bar	Stacked bar
Show distribution of a quantity	Box plot or violin	Histogram	Mean ± SD bar (hides shape)
Compare 2 quantities across items	Scatter plot	Slopegraph	Two parallel bars
Show ranking stability across conditions	Critical Difference (CD) diagram	Bumpchart	Multiple bar charts
Show learning over training	Line plot with shaded CI	Step plot	Bar per epoch
Show a process / pipeline	Schematic diagram	Flowchart	Wall of text
Compare full-distribution forecasts	Forest plot	Caterpillar plot	Mean-only bar
Show 2D embedding structure	Scatter colored by category	Hexbin if dense	3D rotation
Show time × condition × value	Faceted line plots	Heatmap with time on x-axis	Animated GIF

Table Type Patterns

Situation	Default table	Notes
Main results, N methods × M metrics	Bold best per column, underline second	Add std as ±x.xx or in parens
Ablation, removing components	Each row removes one component, last row = full	Mark significance vs. full
Dataset statistics	Group by attribute (size, splits, source)	Header rows with categories
Hyperparameters / configs	Method on rows, hyperparam on columns	OK to put in appendix
Qualitative comparison of methods	Methods on rows, properties (✓/✗/partial) on columns	Use for related work surveys

References

For deeper guidance on specific figure categories, see:

• references/ablation-table-templates.md — patterns for ablation tables
• references/statistical-figures.md — significance, CIs, effect sizes
• references/benchmark-paper-figures.md — patterns specific to benchmark papers (CD diagrams, rank stability, etc.)

General Principles

1. One message per figure. If you can't state the message in one sentence, split the figure.
2. Show distributions, not just means. Box plots beat bar-of-mean plots for revealing what's actually happening.
3. Color sparingly and meaningfully. Categorical color for categorical variables; sequential for ordered; diverging for signed deltas. Never decorative.
4. Direct labels beat legends. When possible, label lines/bars directly on the plot.
5. Don't sacrifice precision for prettiness. A clean Matplotlib figure beats a pretty Seaborn figure that hides outliers.
6. Captions tell the story, axes report the data. Caption answers "what should I take away?", axes answer "what am I looking at?".

When Suggesting Figures, Brainstormer Should

1. State the message in one sentence
2. Pick the default figure type from the table above
3. Mention 1-2 alternatives if the default has caveats
4. Flag if the data isn't ready (e.g., "needs CIs computed", "needs N>3 seeds")
5. Suggest where in the paper structure the figure should live