Roadmap¶

This page is the high-level roadmap, summarising shipped milestones and planned work.

Released¶

ConceptGraph (tree, NetworkX-backed) with YAML / dict / NetworkX constructors.
Metrics: feature_counts, importance_sum, utilization, auc_drop (3 strategies).
Plots: sunburst, utilization_map, auc_drop_map.
Adapters: from_shap_explanation, from_permutation_importance, from_feature_importances_.
Tests, mypy strict, README quickstart, end-to-end notebook on Give Me Some Credit.

Fixed auc_drop_map rendering empty when skip_root=True (root feature_count was 0 with non-zero children, which Plotly silently dropped).

New metrics: feature_correlation, nullity_correlation, shap_correlation, joint_missing_rate, column_missing_rate, coherence_importance.
New plots: correlation_block, joint_missing_map, coherence_importance_scatter, regulatory_tag_overlay.
Cross-cutting decisions locked: switchable correlation method (Spearman default), joint_missing_rate is a standalone metric (no implicit fusion into auc_drop), shap stays an optional extra.

concept_violin — per-concept horizontal violin (KDE) of summed signed SHAP per sample, branch-tinted.
ConceptPredictionExplainer.waterfall(row=...) — single-prediction waterfall rolled up to a chosen tree depth.
Rendering defaults: root concept hidden by default; sunburst colours by branch (hierarchical shading) when no colorscale is given; utilization_map subsumes the standalone feature-count sunburst; correlation_block stacks nested concept labels in depth-stratified rows.

Uncertainty: bootstrap_importance + signed_concept_bar — bar chart with bootstrap confidence intervals.
Interactions: concept_interaction_matrix + concept_interaction_heatmap — concept × concept SHAP-interaction matrix; concept_sankey — multi-tier SHAP flow diagram with explicit ontological grouping.
Cohort: segment_importance + segment_concept_heatmap — concept × cohort heatmap; concept_pareto — faceted Lorenz curves per cohort.
Drift: attribution_drift + concept_drift_lines — multi-period attribution monitoring; concept_drift_delta + concept_drift_sunburst — period-to-period delta sunburst.

concept_disparity + concept_disparity_heatmap — concept × protected-group additive-gap matrix vs a reference group. protected accepts a pd.Series or column-name string (mirrors segment_importance).
Protected-attribute API locked at minimal (one attribute per call, explicit reference). Intersections, first-class ProtectedAttribute metadata on graph nodes, and multiple-reference baselines are filed for v0.6.x / v1.0.

No API changes. Shared helpers extracted (aligned_index_map, per_sample_per_concept, branch_colors, default qualitative palette). Drift-lines title now honestly reports "top K of N". All deprecation aliases removed (single-user package, no back-compat needed).
Test suite cleanup: drop the _v0X_ release-version prefix from test file names; hoist common fixtures into conftest.py.
Docs: full rewrite to story-first landing + guided Tour mirroring the notebook + question-shaped User Guide pages.

Intersection protected attributes (e.g. gender × age_band).
First-class ProtectedAttribute object on the graph with per-node sensitive-flag metadata.
Multiple-reference baselines (compare every group to the population mean instead of one reference group).

The four cross-cutting decisions locked during the v0.3 grooming session:

ID	Decision	Locked value
D1	Default correlation method	Switchable, default = Spearman
D2	`auc_drop` realism weight	Standalone `joint_missing_rate` metric, no implicit fusion
D3	SHAP dependency posture	Optional extra (`[shap]`)
D4	Single-prediction surface	`ConceptPredictionExplainer` class