Future Work & Research Directions

This ledger tracks potential research directions and enhancements to the calibrated explanations framework, organized by theme and linked to relevant literature.

Active Research Areas

1. Adaptive Binning Strategies

Current state: Fixed or Mondrian-based discretization.

Research question: Can we develop adaptive binning that optimizes for both explanation fidelity and computational efficiency?

Literature connections:

  • Conformal prediction literature on adaptive binning

  • Information-theoretic approaches to discretization

Potential implementation: v0.10.x or later

2. Multi-Calibration Fairness

Current state: Basic Mondrian categorizer support for conditional fairness.

Research question: How can we extend the framework to guarantee multi-calibration across intersectional protected attributes?

Literature connections:

  • Multi-calibration (Hébert-Johnson et al., 2018)

  • Intersectional fairness frameworks

Potential implementation: v1.1.x (requires ADR for fairness primitives)

3. Conformal Guarantees for Explanations

Current state: Probabilistic intervals for predictions.

Research question: Can we provide distribution-free coverage guarantees for feature importance rankings?

Literature connections:

  • Conformal prediction theory

  • Venn-Abers calibration

Potential implementation: Research prototype first

4. Temporal Calibration Drift

Current state: Static calibration data.

Research question: How should the framework handle concept drift and when should recalibration be triggered?

Literature connections:

  • Online conformal prediction

  • Adaptive learning systems

Potential implementation: v1.2.x (monitoring tooling)

5. High-Dimensional Feature Interactions

Current state: Pairwise conjunctions via add_conjunctions().

Research question: What are computationally feasible approaches for explaining higher-order interactions while maintaining calibration guarantees?

Literature connections:

  • Shapley interaction indices

  • Functional ANOVA decompositions

Potential implementation: Plugin system extension (v1.x)

Long-Term Vision

Theoretical Foundations

  • Formal analysis of calibration error bounds for explanations

  • Convergence guarantees for iterative calibration methods

  • Connections to information theory and rate-distortion tradeoffs

Computational Efficiency

  • Parallel calibration strategies (continuation of ADR-004)

  • Incremental calibration for streaming data

  • GPU acceleration for large-scale deployments

Ecosystem Integration

  • Standardized interchange formats for calibrated explanations

  • Integration with MLOps platforms (MLflow, Weights & Biases)

  • AutoML-friendly presets and hyperparameter guidance

Contributing Research

We welcome research contributions! If you’re working on any of these areas or have related ideas:

  1. Open a discussion in the GitHub Discussions forum

  2. Reference relevant ADRs (see docs/improvement/adrs/)

  3. Consider prototyping as a plugin (see ADR-006, ADR-013, ADR-015)

  4. For published work, cite the calibrated explanations framework and we’ll add your paper to our literature tracker

References

This ledger complements the citing guide which lists foundational papers and recommended citation practices.