Configuration

One ExperimentConfig (pydantic v2) describes a run completely. Unknown keys are errors, values are checked, instances are immutable, and the SHA-256 of the canonical JSON form identifies a configuration independently of its file format (ADR-011). YAML, TOML and JSON are read and written.

erdr config show                      # the default (literal) configuration
erdr config show --preset closest     # B = 2 + own-class z
erdr config init my.yaml              # start a file
erdr config check my.yaml             # validate; hash, preset, non-literal readings
erdr run -c my.yaml

The default (literal) configuration

schema_version: 1
name: erdr
description: ''
dataset:
  path: null              # null = datasets/default.dat; or a file, a dataset id, builtin:<name>
  format: auto            # auto | extended | erdr-workbook | csv | xlsx | parquet
  sheet: null
  class_column: null
  id_column: null
  feature_types: null     # {name: quantitative | nominal} overrides the file's types
  missing_values: error
preprocessing:
  normalizer: {name: minmax, params: {}}
distance:
  metric: {name: auto, params: {}}     # auto = the metric of the dataset type, or a metric
  type_metrics:                        # the metric auto takes for each dataset type (ADR-022)
    quantitative: {name: manhattan, params: {}}
    nominal: {name: hamming, params: {}}
    heterogeneous: {name: zhuravlev, params: {}}
  features: auto          # auto | all | quantitative | nominal | [feature names]
  decimals: 10
erdr:
  k: 3
  k_rule: enforce         # enforce | warn | off
  epsilon: 0.001
  d_decimals: 10
  readings:
    self_in_l: included   # A: included (1) | excluded (2)
    blocked_cover: kept   # B: kept (1) | dropped (2)
    step5_return: after-all   # C: after-all (1) | first-move (2)
    z_scope: all          # all | own-class
output:
  runs_dir: runs
  dataset_formats: [dat, csv]

A plug-in may be given by name alone: normalizer: rank.

Step 0

preprocessing.normalizer — minmax (default), none, z-score (ddof: 1 or 0), robust, max-abs, decimal-scaling, rank, unit-length. Every normalizer is fitted on the objects of E₀ and maps a feature without spread to 0; nominal codes are never changed.

distance.metric — auto (default: the metric of the dataset type, below), zhuravlev (aliases zhuravlyov, juravlev), weighted-zhuravlev (quantitative_weight, nominal_weight, feature_weights), heom, gower, manhattan, euclidean, chebyshev, minkowski (p ≥ 1), canberra, cosine, mahalanobis, hamming. distance.features: auto gives every metric its domain — all features for heterogeneous metrics, set I for Euclidean and the like, set J for Hamming; an explicit selection outside the domain is an error that names the features.

The metric follows the dataset type (Datasets, ADR-022). With metric: auto the metric comes from distance.type_metrics by the type of the dataset — quantitative → manhattan, nominal → hamming, heterogeneous → zhuravlev. These are the Zhuravlev metric and its two parts, so auto gives the Zhuravlev distances bit for bit on every dataset (the literal default). Each entry must suit its type: a quantitative or heterogeneous metric for quantitative datasets, Hamming or a heterogeneous metric for nominal ones, a heterogeneous metric for heterogeneous ones (erdr config check verifies it); an entry may carry parameters:

distance:
  metric: auto
  type_metrics:
    quantitative: {name: minkowski, params: {p: 3}}   # changes the results on quantitative data

A metric named in metric is used for every dataset it is defined on: on a dataset without its feature type (Euclidean on a nominal dataset) the run stops with an error that lists the metrics that suit; a single-type metric on a heterogeneous dataset runs on its part of the features and is flagged ⚠.

Steps 1–6

• erdr.k — k of z(Sᵢ, k).
• erdr.k_rule — enforce (default) rejects a k outside odd 3 … 2·min|Kᵢ| − 3; warn logs it and records it in the run; off only requires that z(Sᵢ, k) exists (ADR-005).
• erdr.epsilon — ε > 0 of the blocking; any ε gives the same result.
• erdr.d_decimals — rounding of d.
• erdr.readings — the open points of §1.1 (The algorithm).

Presets and shipped files

Preset	Readings	Use
literal (default)	A = B = C = 1, z of any class	the article's text; reproduces the workbook
closest	B = 2, own-class z	the reading closest to the article's Table 1

configs/default.yaml (the workbook), configs/heart-disease-270.yaml (literal, |D| = 198 for k = 3) and configs/heart-disease-270-closest.yaml (|D| = 75 for k = 3). Every YAML/TOML file written by the package starts with a comment header naming the preset and every non-literal reading.

Command-line overrides

erdr run accepts --preset, --k, --epsilon, --reading 121, --z-scope own-class, --metric, --normalizer and --k-rule on top of a configuration file; the result is validated like a file.