Quickstart¶

Evaluate one projection¶

Prepare a finite two-dimensional original array and a projection with the same number of rows:

import numpy as np
from zadu import ZADU

rng = np.random.default_rng(0)
original = rng.normal(size=(200, 16))
projection = original[:, :2] + 0.05 * rng.normal(size=(200, 2))

specs = [
    {"id": "tnc", "params": {"k": 20}},
    {"id": "stress", "params": {}},
]

runner = ZADU(specs, original)
scores = runner.measure(projection)

print(scores[0]["trustworthiness"])
print(scores[0]["continuity"])
print(scores[1]["stress"])

Results follow specification order. Every result is a dictionary of finite Python scalar scores.

Use typed specifications¶

MEASURE and make_spec() provide autocomplete-friendly alternatives to raw dictionaries:

from zadu import MEASURE, ZADU, make_spec

specs = [
    make_spec(MEASURE.TNC, k=20),
    make_spec(MEASURE.STRESS),
]

scores = ZADU(specs, original).measure(projection)

The short aliases such as "tnc" and full IDs such as "trustworthiness_continuity" are both accepted.

Label-based measures¶

Pass a label vector to measure() when a specification contains nh, ca_tnc, dsc, ivm, c_evm, l_tnc, or cadi:

specs = [
    {"id": "nh", "params": {"k": 15}},
    {"id": "dsc", "params": {}},
]

scores = ZADU(specs, original).measure(projection, label=labels)

Labels may be strings or arbitrary numeric values.

Input checklist¶

original and projection must be finite numeric 2D arrays with the same number of rows.
For neighbor-based measures, use 1 <= k < n.
Standard T&C and class-aware T&C normalization additionally require k < n / 2.
Provide labels for every label-based specification.
Undefined inputs such as constant distances, a single class, or coincident neighborhoods raise an actionable ValueError instead of returning nan or inf.

Next, use Choose measures to match scientific questions to metrics, or read Evaluate projections for configuration and diagnostics.