From dfbbbca369d28f0382ba88f732c8166e86b78795 Mon Sep 17 00:00:00 2001 From: Hongyu Lin <67950264+hon-gyu@users.noreply.github.com> Date: Fri, 9 Feb 2024 09:43:21 +0000 Subject: [PATCH] Fix small typo in user guide (#1129) Signed-off-by: Hongyu Lin <67950264+hon-gyu@users.noreply.github.com> --- .../user_guide/modeling_gcm/graphical_causal_model_types.rst | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/docs/source/user_guide/modeling_gcm/graphical_causal_model_types.rst b/docs/source/user_guide/modeling_gcm/graphical_causal_model_types.rst index 4432072586..96b7952594 100644 --- a/docs/source/user_guide/modeling_gcm/graphical_causal_model_types.rst +++ b/docs/source/user_guide/modeling_gcm/graphical_causal_model_types.rst @@ -55,11 +55,11 @@ use cases. However, DoWhy also has different types of mechanisms, specifically f natively supporting different types of data: - `Additive Noise Models `_ (continuous data) of the form :math:`X_i = f_i(PA_i) + N_i`, where :math:`f_i` can be any kind of regression function (e.g., from scikit-learn) and the noise :math:`N_i` is unobserved noise. When fitting an ANM, this then boils down to fitting the :math:`f_i` model (e.g., via least squares) and fitting :math:`N_i` based on the residuals :math:`N_i = X_i - f_i(PA_i)`. As mentioned throughout the user guide, ANMs are the most commonly used types of causal models due to their simplicity and ability to answer counterfactual questions. -- `Post-nonlinear models `_ (continuous data) of the form :math:`X_i = g_i(f_i(PA_i) + N_i)`, where :math:`g_i` is assumed to be invertible. These are a generalization of ANMs, allowing more complex relationships between :math:`N_i` and :math:`PA_i`. +- `Post-nonlinear Models `_ (continuous data) of the form :math:`X_i = g_i(f_i(PA_i) + N_i)`, where :math:`g_i` is assumed to be invertible. These are a generalization of ANMs, allowing more complex relationships between :math:`N_i` and :math:`PA_i`. - `Discrete Additive Noise Models `_ (discrete data), which have a similar definition as ANMs but are restricted to discrete values. - `Classifier-based Functional Causal Models `_ (categorical data) of the form :math:`X_i = f_i(PA_i, N_i)`, which cannot be used for rung 3 queries, since :math:`f_i` is typically not invertible here with respect to :math:`N_i`, but can be used for algorithms relying only on interventional queries (rung 2). Here, :math:`f_i` can be based on any classification model (e.g., from scikit-learn) and :math:`N_i` is by definition a uniform distribution on [0, 1] used to sample from the conditional class probabilities. In all mechanisms, causal sufficiency is assumed, i.e., :math:`N_i` is assumed to be independent of :math:`PA_i`. More details and justification of these types of causal mechanisms can be found in the correspondingly linked papers. Note that when using the auto assignment function, DoWhy tries to use invertible FCMs, such as ANMs, due to their flexibility -in addressing rung 3 queries. For categorical data, make sure to represent them as strings. \ No newline at end of file +in addressing rung 3 queries. For categorical data, make sure to represent them as strings.