Guides

Quickstart

Read these in order:

• What are global and regional effects: the concepts
• effector's API: the whole API in 3 minutes
  • (a) The input layer: numpy, models, adapters, the schema
  • (b) effector's report: the one-liner, .show(), and the HTML page
  • (c) The interactive API: the five engines, global and regional effects

Page by page

Every component of the report and every verb of the interactive API has its own page.

Inside (b) effector's report:

• The data & model header: what was explained, and how good the model is
• The explained variance ledger: how much of the model the explanation captures
• The rejected splits: every split considered and refused, with the reason
• The ranked features: the ranking, its units, and the coverage cut
• The triage plane: importance versus heterogeneity, where to look first
• The regional analysis: partition trees and per-leaf plots
• The global baseline: what you would have believed without regions
• Configuring the report: every knob of explain(...)

Inside (c) the interactive API:

• Construct and fit: the five engines and .fit()
• Customize .fit(): binning, centering, search depth
• plot: effects and their heterogeneity, drawn
• eval: effects as numbers, model free
• importance and heter_score: the twin scalars, in output units
• find_regions: subregions and the Partition value
• select_regions: which splits earn their keep
• compare and plot_triage: the cross-engine views

Going deeper

• The mental model: the thinking behind the API; one engine, values not state, two entrances
• Methods: the math reference: how each method defines the effect, its heterogeneity, and the two scalars, per feature type
• The design contract: the rules the API is built on, R1 to R14, one breath each
• Efficiency of global methods: count the model calls; everything after the fit is free
• Efficiency of regional methods: the regional search costs zero model calls