Skip to content

Plotting Results

The plot capability draws charts using matplotlib and Seaborn from scientific results. The input is either a .csv file produced by the summarise action or a metrics workflow cache (.db) produced by the evaluate_graph action. The action writes no workflow cache, only an image file such as a .png or .jpg; a step whose output is not a workflow cache runs once per workflow regardless of any matrix definition.

Keeping the .csv input is useful for plotting simple ad-hoc data without building a workflow cache.

The plotting code is migrated from the legacy experiments/plot.py module in the causaliq-discovery repository, so figures such as the variable-ordering CPDAG F1 chart can be replicated exactly.

Parameters

Parameter CLI Required Description
input -i/--input Yes Input .csv file (summarise) or metrics cache .db (evaluate_graph)
output -o/--output Yes Output chart file, e.g. a .png or .jpg image
type --type No Type of plot required: line, regression, histogram, box, violin, bar or scatter (default line)
subplot --subplot Yes Column name which defines the subplot, e.g. network
group --group Yes Column name which defines the legend groups, e.g. series
x -x/--x Yes Column name which provides the x-axis values, e.g. sample_size
y -y/--y Yes Column name which provides the y-axis values, e.g. f1.mean
property -p/--property No Chart property in name=value format with a Python literal value, repeatable

Input

Summarise CSV

A .csv produced by the summarise action has one row per subplot, x-value and group combination. For example:

network series sample_size f1.mean
asia BNLEARN/HC_STD 100 0.14
asia BNLEARN/HC_STD 1000 0.27
asia BNLEARN/HC_OPT 100 0.25
sports BNLEARN/HC_OPT 1000 0.48

Metrics Cache

A metrics workflow cache (.db) produced by evaluate_graph holds one metrics object per entry. The plot columns are the entry's matrix values (the workflow matrix keys, e.g. network, series, sample_size) plus the metric names recorded by evaluate_graph (e.g. f1, shd, equiv.f1, equiv.shd). Entries whose run failed hold provenance only — no object — and contribute no row.

This is how the variable-ordering figure is produced: plot reads the metrics cache written by the evaluate_graph step.

CLI Example

This command replicates the ord_hc_f1 figure with one subplot per network, a line per series, a log-scale x-axis and the publication fonts:

causaliq-analysis plot -i results/summary.csv -o figures/ord_hc_f1.png \
    --type line --subplot network --group series \
    --x sample_size --y f1.mean \
    -p "figure.title=" \
    -p "subplot.aspect=1.05" \
    -p "figure.subplots_left=0.04" \
    -p "figure.subplots_right=0.86" \
    -p "figure.subplots_top=0.98" \
    -p "figure.subplots_hspace=0.22" \
    -p "subplot.grid=True" \
    -p "subplot.grid_colour=lightgray" \
    -p "subplot.background=white" \
    -p "xaxis.scale=log" \
    -p "xaxis.label=Sample size" \
    -p "yaxis.label=F1 (CPDAG)" \
    -p "legend.title=variable ordering" \
    -p "palette=['#66bd63','#d73027','#000000']"

Workflow Example

steps:
  - name: "Plot CPDAG F1"
    uses: "causaliq-analysis"
    with:
      action: "plot"
      input: "results/graphs-metrics.db"
      output: "figures/ord_hc_f1.png"
      type: "line"
      subplot: "network"
      group: "series"
      x: "sample_size"
      y: "equiv.f1"
      properties:
        - "xaxis.label=Sample size"
        - "yaxis.label=F1 (CPDAG)"

Properties

Properties use the names from the legacy plotting module. Each property is a string of the form <name>=<value> where <value> is written in Python literal syntax, so the full range of Python types can be specified:

  • int.property=22 sets the integer 22
  • float.property=0.23 sets the float 0.23
  • string.property='string value' sets the string string value
  • tuple.property=(2, 'dad') sets the tuple (2, 'dad')
  • list.property=['a', 1, 2.3] sets the list ['a', 1, 2.3]
  • dict.property={'key1': 1, 'key2': 'me'} sets a dict
  • set.property={1, 'two'} sets the set {1, 'two'}
  • bool.property=True sets True, and none.property=None sets None

Values which are not valid Python literals (e.g. lightgray or Sample size) are treated as plain strings, and a blank value (e.g. figure.title=) sets an empty string. The ¬ character sets None. Strings may also be quoted: figure.title='Test Figure' is equivalent to figure.title=Test Figure.

YAML note: a dict value contains : (colon-space), which ends a YAML plain scalar, so dict properties must be double-quoted in the workflow file, e.g. - "dict.property={'key1': 1, 'key2': 'me'}".

Supported property names include:

  • Figure: figure.title, figure.title_fontsize, figure.dpi, figure.per_row, figure.background
  • Subplot: subplot.kind, subplot.aspect, subplot.title, subplot.title_fontsize, subplot.axes_fontsize, subplot.background, subplot.grid, subplot.grid_colour
  • Axes: xaxis.label, xaxis.range, xaxis.scale, xaxis.ticks, xaxis.ticks_fontsize, xaxis.shared, yaxis.label, yaxis.range, yaxis.scale, yaxis.ticks, yaxis.ticks_fontsize, yaxis.shared
  • Figure layout: figure.subplots_top, figure.subplots_left, figure.subplots_right, figure.subplots_bottom, figure.subplots_hspace, figure.subplots_wspace
  • Legend: legend.title, legend.title_fontsize, legend.fontsize, legend.labels, legend.outside, legend.key, legend.loc, legend.ncol
  • Lines: line.sizes, line.dashes
  • Other: palette, violin.scale, violin.width