Possible to incrementally change the causal graph and learn the causal function only on the nodes affected in GCM?

[Yangliu-SY]

I use the functionality of attributing anomalies of a node to its parents and ancestors in GCM for root cause analysis. Sometimes, I need to add a new node to the graph and some edges between the new node and existing nodes. I would then like to learn the causal mechanism for the nodes affected by the change.
I have tried the following. I have an exiting causal graph and the associated causal model that has been learned. I modified the causal graph by adding the new node and associated new edges. I created a new causal model from the modified causal graph. I would like to copy the learned causal mechanisms in the existing causal model to the new causal model. Afterwards, I would learn the causal mechanisms on the nodes affected by the change in the causal graph.
But when I tried to assign the learned causal mechanism in the existing causal model to the new causal model, I received an error that said I couldn't do this assignment.

for node in causal_graph_exist.nodes:
    causal_model_new.causal_mechanism(node) = causal_model_exist.causal_mechanism(node).copy()

Is it possible to incrementally change the causal graph and learn the causal function only on the nodes affected? I looked up the documentation of DoWhy but I didn't find a way to achieve this. Let me know if you need more information.

• DoWhy version 0.11.1

[bloebp]

Hi,

yes that is possible, you can use the functions on single nodes directly. Here is an example:

import networkx as nx
import numpy as np
import pandas as pd

import dowhy.gcm as gcm

X = np.random.normal(loc=0, scale=1, size=1000)
Y = 2 * X + np.random.normal(loc=0, scale=1, size=1000)
Z = 3 * Y + np.random.normal(loc=0, scale=1, size=1000)
data = pd.DataFrame(data=dict(X=X, Y=Y, Z=Z))

causal_model = gcm.StructuralCausalModel(nx.DiGraph([('X', 'Y'), ('Y', 'Z')]))
gcm.auto.assign_causal_mechanisms(causal_model, data)
gcm.fit(causal_model, data)

# Generate data for a new node 'W'
W = 2 * Z + X + np.random.normal(loc=0, scale=0.1, size=1000)
data["W"] = W

# Add the new node to the existing graph
causal_model.graph.add_edges_from([("Z", "W"), ("X", "W")])

# Automatically assign a model to the new node only (or assign one manually)
gcm.auto.assign_causal_mechanism_node(causal_model, "W", data, gcm.auto.AssignmentQuality.GOOD)

# Alternatively, we can also call the assign_causal_mechanisms function, but need to make sure that override_models is false
# gcm.auto.assign_causal_mechanisms(causal_model, data, override_models=False)

# Fit only the new node
gcm.fitting_sampling.fit_causal_model_of_target(causal_model, "W", data)

print(gcm.draw_samples(causal_model, 5).head())

[Yangliu-SY]

Thanks a lot. I appreciate it very much.