* Library functions for temporal causal functionality
* shifting plotter function
* printing graph: best practices
* added docstrings
* moved datasets
* updated tutorial notebook
* sphinx documentation
* updated shifting columns with 0,1,..,max_lag
* support for dot format
* tigramite support
* updated filter to be a hidden function
* black and isort utils
* black and isort timeseries
* updated notebook text
Signed-off-by: Amit Sharma <amit_sharma@live.com>
* integer range fix
* correction in timestamp : notebook text
* time lagged causal estimation
* removed cell outputs
* find ancestors
* include ancestors in notebook
* formatting changes
* comments : notebook
* multiple time lags : csv graph'
* multiple time lags
* unrolled graph using bfs
* cleanup of functions
* removed find parents and ancestors
* tests for causal graph creation
* tests for adding lagged edges
* tests for shifting columns
* tigramite dependency added
---------
Signed-off-by: Amit Sharma <amit_sharma@live.com>
Co-authored-by: Amit Sharma <amit_sharma@live.com>
Add example notebook for counterfactual fairness using GCM
Following the paper Estimates Counterfactual Fairness following Kusner et al. (2018)
---------
Signed-off-by: Kriti Mahajan <kritimhj@amazon.com>
This should significantly speed up some calculations for larger graphs. In this regard, also increased the default samples for the anomaly attribution to balance the reduced number of Shapley run.
Signed-off-by: Patrick Bloebaum <bloebp@amazon.com>
Taking less samples to estimate ICC. This might lead to higher variance in the estimate, but speeds it up by multiple factors.
Signed-off-by: Patrick Bloebaum <bloebp@amazon.com>
Before, when a causal graph had causal mechanisms assigned, they were also used when creating a new GCM object based on it. Now, they are removed (from a copied version of the graph).
Signed-off-by: Patrick Bloebaum <bloebp@amazon.com>
The method merges p-values by taking the minimum of the adjusted p-values. This then represents a p-value for a global hypothesis.
Signed-off-by: Patrick Bloebaum <bloebp@amazon.com>
* [DOC] fix typo
Signed-off-by: yogabonito <yogabonito@users.noreply.github.com>
* [DOC] small fix
The text first informs that PC, FCI, and GES will be considered, but instead of FCI the notebook shows LiNGAM.
Signed-off-by: yogabonito <yogabonito@users.noreply.github.com>
---------
Signed-off-by: yogabonito <yogabonito@users.noreply.github.com>
This is very closely related to the RescaledMedianCDFQuantileScorer but a statistically more well-defined approach. However, it is also slightly more conservative for small sample sizes.
Furthermore, it changes the way equal samples are counted in MedianCDFQuantileScorer by also including the test sample itself. This prevents a p-value of 0.
Signed-off-by: Patrick Bloebaum <bloebp@amazon.com>
Before, we were loading it directly from the UCI dataset repository. However, this cause connection issues and build fails. The dataset is now added directly with its respective license.
Further, removed the output of the causal discovery example notebook to ensure that it gets compiled while building.
Signed-off-by: Patrick Bloebaum <bloebp@amazon.com>
Before, a single p-value still got adjusted using the 2 * ... formula, which led to an unnecessary adjustment.
Signed-off-by: Patrick Bloebaum <bloebp@amazon.com>
Add distribution_change_robust method
Adding a new distribution change method that more robust and converges faster. However, it does not support all kind of set functions. Currently, it supports mean and variance attributions.
Based on the paper:
Quintas-Martinez, V., Bahadori, M. T., Santiago, E., Mu, J., Janzing, D., and Heckerman, D. Multiply-Robust Causal Change Attribution, Proceedings of the 41st International Conference on Machine Learning, Vienna, Austria. PMLR 235, 2024.
Signed-off-by: Víctor Quintas-Martínez <vquintas@mit.edu>
This uses an improved version of the "twice the average" rule following recent results from M. Gasparini, R. Wang, and A. Ramdas, *Combining exchangeable p-values*, arXiv 2404.03484, 2024.
This new method is now used by default when merging p-values. Accordingly, the quantile based method was renamed to be more consistent with the naming pattern.
Signed-off-by: Patrick Bloebaum <bloebp@amazon.com>
* For consistency and avoidance of future issues, replace all occurrences of Pandas' get_dummies with skLearn's OneHotEncoder. Encoder lifespan: Reuses encoders for new estimate_effect() calls, and replaces existing encoders on CausalEstimator.fit(). Additional uses of get_dummies without side-effects or consistent encoding issues in do-Sampler Propensity Scores utilities also replaced for consistency.
Signed-off-by: DAVID RAWLINSON <dave@causalwizard.app>
* Add categorical encoding consistency tests for CausalEstimators. Fix bug in arg order for RegressionEstimator._do().
Signed-off-by: DAVID RAWLINSON <dave@causalwizard.app>
---------
Signed-off-by: DAVID RAWLINSON <dave@causalwizard.app>
Co-authored-by: DAVID RAWLINSON <dave@causalwizard.app>
Rahul Shrestha
Ashutosh Srivastava
Patrick Bloebaum
Chris Trevino
kmhj13
dependabot[bot]
yogabonito
allcontributors[bot]
Jon Cowie
Patrick Blöbaum
Christoph
Víctor Quintas-Martínez
Harsh Sinha
Lukas Heumos
Amit Sharma
drawlinson
Hongyu Lin