Targets for v0 release (targeting April 1, 2023)

v0 release is targeting the following:

• Formalizing the "context" abstraction.
• Algorithms
  • Fully observed
    • Constraint-based
      • PC
    • Score-based
      • GES algorithm with BIC score for discrete data
    • Common baseline algorithms
      • GOLEM
      • NOTEARS
    • Deep learning-based algorithms
      • DECI
    • SCM-based
      • DirectLINGAM
  • Latent variable
    • Constraint-based
      • GIN
      • FCI
      • Psi-FCI
• Support for include lists and exclude lists in constraint and score-based algorithms
• Suite of conditional independence and discrepancy tests
  • Traditional Chi-squared, G-squared, and z-score
  • Classifier conditional independence
  • Conditional mutual information
  • Kernel conditional independence
  • Kernel conditional discrepancy
  • Bregman divergence conditional discrepancy
• Tutorials and case studies demonstrating above tools

Out of scope for v0 but on the roadmap

• Graph conversion algorithms
  • DAG2PDAG: Convert a DAG to a PDAG
  • DAG2ADMG: Convert a DAG to an ADMG (for when the DAG has latent variables)
  • ADMG2PAG: Convert an ADMG to a PAG
• Discretization of continuous to discrete data
• Mixed (discrete, continuous) score-based algorithms
• Consistent extension algorithms (generate a instance of an equivalence class): PAG2ADMG, ADMG2DAG, PDAG2DAG
• Atomic interventions
• Local causal discovery
• Score-based latent variable algorithms
• An assumptions abstraction: Graph learning output should return the set of assumptions used by the model for downstream analysis.
• Abstractions for graph priors and ensembles