v1.12.2

New Features

• Modified to enable the use of prior knowledge in the GPU-based DirectLiNGAM implementation. (#183)
• Updated the utils.evaluate_model_fit function to re-estimate parameters when computing fit indices. (#184, thanks @nishikawa-screen)
• Added support for causal explanation metrics based on LEWIS. (#186)

Examples

• Updated the example notebook to visualize three types of distributions for estimated causal coefficients during bootstrapping: X->Y, Y->X, and no directed edge between X and Y. (#185, thanks @nishikawa-screen)

Code Fixes

• Updated compatibility with changes to check_array arguments introduced in scikit-learn v1.8.

Installation

• Improved the CI pipeline with the following changes:
  • Added checks on push.
  • Migrated from setup.py to pyproject.toml.

v1.12.1

Installation Fix

• Added autograd to install_requires in setup.py. (#181, thanks @apaleyes)

v1.12.0

New Features

• Added Missingness-LiNGAM (m-LiNGAM) algorithm. This algorithm can handle datasets with MCAR, MAR, and MNAR missingness, accurately estimating LiNGAM model parameters and the mechanisms causing missing data. (#175, thanks @matteoceriscioli)
• Added MultiGroupRESIT algorithm. This performs causal discovery across multiple datasets under the assumption that they share the same causal ordering. It uses the RESIT algorithm to identify the causal structure.
• Added ABICLiNGAM algorithm. ABIC-LiNGAM is a score-based causal discovery method that extends the LiNGAM framework to accurately estimate causal structures in the presence of unmeasured confounders, by leveraging continuous optimization and assuming error terms follow a multivariate generalized normal distribution.
  • The implementation of ABIC-LiNGAM is based on the original code from https://github.com/Yoshimitsu-try/ABIC_LiNGAM.
• Added lingam.utils.MGGDEstimator. This estimates the parameters of a multivariate generalized Gaussian distribution (MGGD).

Examples

• Added an example of LiNGAM analysis for clinical data. (#172, thanks @F0guro)
  • Example_to_analyze_clinical_data.ipynb
• Added an example of causal discovery with imputation on discrete data using algorithms such as RCD, CAM-UV, and RESIT. (#179, thanks @nishikawa-screen )
  • Example_to_perform_causal_discovery_using_RCD_with_bootstrap_and_imputation_on_data_with_discrete_variables.ipynb
  • Example_to_perform_causal_discovery_using_CAMUV_with_bootstrap_and_imputation_on_data_with_discrete_variables.ipynb
  • Example_to_perform_causal_discovery_using_RESIT_with_bootstrap_and_imputation_on_data_with_discrete_variables.ipynb

Installation

• Added autograd, which is required for ABICLiNGAM, to requirements.txt.

v1.11.0

New Features

• Added GroupLiNGAM and GroupDirectLiNGAM algorithm.
• Add lingam.utils.get_common_edge_probabilities function that can analyze common edges in causal graphs.

Code Fixes

• Fixed an issue where edges are not removed properly in RESIT algorithm. (#169)

Installation

• Changed SciPy version to 1.13.1 or lower because semopy uses the deprecated scipy.stats.mvn. If semopy is fixed, this will be reverted.

v1.10.0

New Features

• Added MultiGroupCAMUV algorithm, CAM-UV for multiple datasets.
• Added option to set prior knowledge for CAMUV.
• Added option to set prior knowledge for RESIT.

Examples

• Added an example of causal discovery with imputation on discrete data.
  • examples/Example_to_perform_causal_discovery_using_bootstrap_with_imputation_on_data_with_discrete_variables.ipynb

Code Fixes

• Fixed referenced lag numbers when pruning edges in VARLiNGAM. (#161)
• Fixed VARLiNGAM bootstrap method to resample from original residuals. (#163)

Installation

• Changed supported Python version to 3.9 or higher.

v1.9.1

New Features

• Added option to set prior knowledge for LongitudinalLiNGAM.
• Modified lingam.CausalBasedSimulator for Causal Digital Twin. See the following Jupyter notebooks for usage:
  • examples/CausalBasedSimulator_introduction.ipynb
  • examples/CausalBasedSimulator_generating_simulated_data.ipynb
  • examples/CausalBasedSimulator_estimating_functional_relationships.ipynb
• Modified tools.bootstrap_with_imputation for https://github.com/cdt15/pruning_missing_data.

Examples

• Added an example of causal discovery for data with discrete variables.
  • examples/Example_to_perform_causal_discovery_on_data_with_discrete_variables.ipynb

Code Fixes

• Fixed wrong selection of predictors in LongitudinalLiNGAM._compute_residuals when n_lags>1. (#155)

v1.9.0

New Features

• Added option measure=pwling_fast to accelerate DirectLiNGAM by parallelizing causal order on GPU with CUDA.
• Added tools.bootstrap_with_imputation function to interpolate missing data by Multiple Imputation Method and perform causal discovery by bootstrap method for datasets containing missing value.
• Added HighDimDirectLiNGAM class, a DirectLiNGAM algorithm suitable for high-dimensional data.

Code Fixes

• Fixed the calculation of the total effect in LongitudinalLiNGAM. (#136)
• Changed from using AdaptiveLasso to LinerRegression to calculate the total effect. (#137)
• Fixed the calculation method of regression coefficients when calculating residuals in DirectLiNGAM. (#141)

v1.8.3

New Features

• Added utils.evaluate_model_fit function to calculate fit indices.
• Added ind_corr option to use F-correlation to determine independence in BottomUpParceLiNGAM, RCD, MultiGroupRCD, and CAMUV algorithms.
• Added utils.calculate_distance_from_root_nodes function to calculate the shortest distance from the root variable to other variables based on the structure of causal graphs.
• Added utils.calculate_total_effect function to calculate causal effects based on the structure of causal graphs.

Code Fixes

• Fixed the calculation of total effect in LongitudinalLiNGAM.estimate_total_effect. (#122)
• Changed the argument name of sklearn.OneHotEncoder used in experimental.CausalDataGenerator from sparse to sparse_output. This fix requires scikit-learn v1.2 or higher. (#123)

v1.8.2

New Features

• Added LiNA and MDLiNA algorithm.

Code Fixes

• Fixed standardization in lingam.utils.predict_adaptive_lasso function without normalize option in sklearn.linear_model.LassoLarsIC.
• Changed to use lingam.utils.predict_adaptive_lasso function in VARLiNGAM and VARMALiNGAM.
• Changed prune option to True by default in VARLiNGAM and VARMALiNGAM. Additionally, changed the description that this option is for the causal effect of lag.

v1.8.1

New Features

• Added utils.f_correlation function to calculate F-correlation.
• Added utils.visualize_nonlinear_causal_effect function to plot nonlinear causal effects.
• Added an option for LiM to perform the local search or not.
• Added resampled_indices_ property to BootstrapResult.

Code Fixes

• Fixed a bug that labels specified in the utlis.make_dot_highlight function were not reflected in graphs.
• Modified MultiGroupDirectLiNGAM algorithm to allow specifying datasets of different sizes.
• Removed LiNA and MDLiNA algorithm from this package. Independent as a new lina package. https://github.com/cdt15/lina

Installation

• Changed supported Python version to 3.8 or higher.
• Eliminated the fixed numpy version.