Learning and causality

Creative Commons License

aGrUM

interactive online version

In [1]:
import pyAgrum as gum
import pyAgrum.lib.notebook as gnb

Model

Let’s assume a process \(X_1\rightarrow Y_1\) with a control on \(X_1\) by \(C_a\) and a parameter \(P_b\) on \(Y_1\).

In [2]:
bn=gum.fastBN("Ca->X1->Y1<-Pb")
bn.cpt("Ca").fillWith([0.8,0.2])
bn.cpt("Pb").fillWith([0.3,0.7])

bn.cpt("X1")[:]=[[0.9,0.1],[0.1,0.9]]

bn.cpt("Y1")[{"X1":0,"Pb":0}]=[0.8,0.2]
bn.cpt("Y1")[{"X1":1,"Pb":0}]=[0.2,0.8]
bn.cpt("Y1")[{"X1":0,"Pb":1}]=[0.6,0.4]
bn.cpt("Y1")[{"X1":1,"Pb":1}]=[0.4,0.6]

gnb.flow.row(bn,*[bn.cpt(x) for x in bn.nodes()])
G Y1 Y1 X1 X1 X1->Y1 Pb Pb Pb->Y1 Ca Ca Ca->X1
Ca
0
1
0.80000.2000
X1
Ca
0
1
0
0.90000.1000
1
0.10000.9000
Y1
Pb
X1
0
1
0
0
0.80000.2000
1
0.20000.8000
1
0
0.60000.4000
1
0.40000.6000
Pb
0
1
0.30000.7000

Actually the process is duplicated in the system but the control \(C_a\) and the parameter \(P_b\) are shared.

In [3]:
bn.add("X2",2)
bn.add("Y2",2)
bn.addArc("X2","Y2")
bn.addArc("Ca","X2")
bn.addArc("Pb","Y2")

bn.cpt("X2").fillWith(bn.cpt("X1"),["X1","Ca"]) # copy cpt(X1) with the translation X2<-X1,Ca<-Ca
bn.cpt("Y2").fillWith(bn.cpt("Y1"),["Y1","X1","Pb"]) # copy cpt(Y1) with translation Y2<-Y1,X2<-X1,Pb<-Pb

gnb.flow.row(bn,bn.cpt("X2"),bn.cpt("Y2"))
G Ca Ca X2 X2 Ca->X2 X1 X1 Ca->X1 Y2 Y2 X2->Y2 Y1 Y1 Pb Pb Pb->Y1 Pb->Y2 X1->Y1
X2
Ca
0
1
0
0.90000.1000
1
0.10000.9000
Y2
Pb
X2
0
1
0
0
0.80000.2000
1
0.20000.8000
1
0
0.60000.4000
1
0.40000.6000

Simulation of the data

The process is partially observed : the control has been taken into account. However the parameter has not been identified and therefore is not collected.

In [4]:
#the base will be saved in completeData="out/complete_data.csv", observedData="out/observed_data.csv"
import os
completeData="out/complete_data.csv"
observedData="out/observed_data.csv"

# generating complete date with pyAgrum
size=35000
#gum.generateSample(bn,5000,"data.csv",random_order=True)
generator=gum.BNDatabaseGenerator(bn)
generator.setRandomVarOrder()
generator.drawSamples(size)
generator.toCSV(completeData)
In [5]:
# selecting some variables using pandas
import pandas as pd
f=pd.read_csv(completeData)
keep_col = ["X1","Y1","X2","Y2","Ca"] # Pb is removed
new_f = f[keep_col]
new_f.to_csv(observedData, index=False)

We will use now a database fixed_observed_data.csv. While both databases originate from the same process (the cell above), the use of fixed_observed_data.csv instead of observed_data.csv is made to guarantee a deterministic and stable behavior for the rest of the notebook.

In [6]:
fixedObsData="res/fixed_observed_data.csv"

statistical learning

Using a classical statistical learning method, one can approximate a model from the observed data.

In [7]:
learner=gum.BNLearner(fixedObsData)
learner.useGreedyHillClimbing()
bn2=learner.learnBN()
bn2
Out[7]:
G Ca Ca X2 X2 Ca->X2 X1 X1 Ca->X1 Y2 Y2 X2->Y2 Y1 Y1 X1->Y1 Y2->Y1

Evaluating the impact of \(X2\) on \(Y1\)

Using the database, a question for the user is to evaluate the impact of the value of \(X2\) on \(Y1\).

In [8]:
target="Y1"
evs="X2"
ie=gum.LazyPropagation(bn)
ie2=gum.LazyPropagation(bn2)
p1=ie.evidenceImpact(target,[evs])
p2=gum.Potential(p1).fillWith(ie2.evidenceImpact(target,[evs]),[target,evs])
errs=((p1-p2)/p1)
quaderr1=(errs*errs).sum()
gnb.flow.row(p1,p2,errs,rf"$${100*quaderr1:3.5f}\%$$",
              captions=['in original model','in learned model','relative errors','quadratic relative error'])

Y1
X2
0
1
0
0.62110.3789
1
0.45080.5492

in original model
Y1
X2
0
1
0
0.61830.3817
1
0.44150.5585

in learned model
Y1
X2
0
1
0
0.0044-0.0072
1
0.0205-0.0168

relative errors
$$0.07722\%$$
quadratic relative error

Evaluating the causal impact of \(X2\) on \(Y1\) with the learned model

The statistician notes that the change wanted by the user to apply on \(X_2\) is not an observation but rather an intervention.

In [9]:
import pyAgrum.causal as csl
import pyAgrum.causal.notebook as cslnb

model=csl.CausalModel(bn)
model2=csl.CausalModel(bn2)
In [10]:
cslnb.showCausalModel(model)
../_images/notebooks_14-Examples_CausalityAndLearning_20_0.svg
In [11]:
gum.config['notebook','graph_format']='svg'
cslnb.showCausalImpact(model,on=target, doing={evs})
cslnb.showCausalImpact(model2,on=target, doing={evs})
G Ca Ca X1 X1 Ca->X1 X2 X2 Ca->X2 Y1 Y1 X1->Y1 Pb Pb Pb->Y1 Y2 Y2 Pb->Y2 X2->Y2
Causal Model
$$\begin{equation*}P( Y1 \mid \text{do}(X2)) = \sum_{Ca}{P\left(Y1\mid Ca\right) \cdot P\left(Ca\right)}\end{equation*}$$
Explanation : backdoor ['Ca'] found.
Y1
0
1
0.57680.4232

Impact
G X1 X1 Y1 Y1 X1->Y1 X2 X2 Y2 Y2 X2->Y2 Y2->Y1 Ca Ca Ca->X1 Ca->X2
Causal Model
$$\begin{equation*}P( Y1 \mid \text{do}(X2)) = \sum_{X1}{P\left(Y1\mid X1,X2\right) \cdot P\left(X1\right)}\end{equation*}$$
Explanation : backdoor ['X1'] found.
Y1
X2
0
1
0
0.57430.4257
1
0.56280.4372

Impact

Unfortunately, due to the fact that \(P_a\) is not learned, the computation of the causal impact still is imprecise.

In [12]:
_, impact1, _ = csl.causalImpact(model, on=target, doing={evs})
_, impact2orig, _ = csl.causalImpact(model2, on=target, doing={evs})

impact2=gum.Potential(p2).fillWith(impact2orig,['Y1','X2'])
errs=((impact1-impact2)/impact1)
quaderr2=(errs*errs).sum()
gnb.flow.row(impact1,impact2,errs,rf"$${100*quaderr2:3.5f}\%$$",
              captions=[r'$P( Y_1 \mid \hookrightarrow X_2)$ <br/>in original model',
                        r'$P( Y_1 \mid \hookrightarrow X_2)$  <br/>in learned model',' <br/>relative errors',' <br/>quadratic relative error'])
Y1
0
1
0.57680.4232

$P( Y_1 \mid \hookrightarrow X_2)$
in original model
Y1
X2
0
1
0
0.57430.4257
1
0.56280.4372

$P( Y_1 \mid \hookrightarrow X_2)$
in learned model
Y1
X2
0
1
0
0.0044-0.0060
1
0.0243-0.0331


relative errors
$$0.17362\%$$

quadratic relative error

Just to be certain, we can verify that in the original model, \(P( Y_1 \mid \hookrightarrow X_2)=P(Y_1)\)

In [13]:
gnb.flow.row(impact1,ie.evidenceImpact(target,[]),
               captions=[r"$P( Y_1 \mid \hookrightarrow X_2)$ <br/>in the original model","$P(Y_1)$ <br/>in the original model"])
Y1
0
1
0.57680.4232

$P( Y_1 \mid \hookrightarrow X_2)$
in the original model
Y1
0
1
0.57680.4232

$P(Y_1)$
in the original model

Causal learning and causal impact

Some learning algorthims such as MIIC (Verny et al., 2017) aim to find the trace of latent variables in the data !

In [14]:
learner=gum.BNLearner(fixedObsData)
learner.useMIIC()
bn3=learner.learnBN()
In [15]:
gnb.flow.row(bn,bn3,f"$${[(bn3.variable(i).name(),bn3.variable(j).name()) for (i,j) in learner.latentVariables()]}$$",
              captions=['original model','learned model','Latent variables found'])
G Ca Ca X2 X2 Ca->X2 X1 X1 Ca->X1 Y2 Y2 X2->Y2 Y1 Y1 Pb Pb Pb->Y1 Pb->Y2 X1->Y1
original model
G Ca Ca X1 X1 Ca->X1 X2 X2 X2->Ca Y2 Y2 X2->Y2 Y1 Y1 X1->Y1 Y2->Y1
learned model
$$[('Y2', 'Y1')]$$
Latent variables found

A latent variable (common cause) has been found in the data betwenn \(Y1\) and \(Y2\) !

Therefore we can build a causal model taking into account this latent variable found by MIIC.

In [16]:
model3=csl.CausalModel(bn2,[("L1",("Y1","Y2"))])
cslnb.showCausalImpact(model3,target, {evs})
G L1 Y1 Y1 L1->Y1 Y2 Y2 L1->Y2 X1 X1 X1->Y1 X2 X2 X2->Y2 Ca Ca Ca->X1 Ca->X2
Causal Model
$$\begin{equation*}P( Y1 \mid \text{do}(X2)) = \sum_{X1}{P\left(Y1\mid X1\right) \cdot P\left(X1\right)}\end{equation*}$$
Explanation : backdoor ['X1'] found.
Y1
0
1
0.57250.4275

Impact

Then at least, the statistician can say that \(X_2\) has no impact on \(Y_1\) from the data. The error is just due to the approximation of the parameters in the database.

In [17]:
_, impact1, _ = csl.causalImpact(model, on=target, doing={evs})
_, impact3orig, _ = csl.causalImpact(model3, on=target, doing={evs})

impact3=gum.Potential(impact1).fillWith(impact3orig,['Y1'])
errs=((impact1-impact3)/impact1)
quaderr3=(errs*errs).sum()
gnb.flow.row(impact1,impact3,errs,rf"$${100*quaderr3:3.5f}\%$$",
              captions=['in original model','in learned model','relative errors','quadratic relative error'])
Y1
0
1
0.57680.4232

in original model
Y1
0
1
0.57250.4275

in learned model
Y1
0
1
0.0075-0.0102

relative errors
$$0.01588\%$$
quadratic relative error
In [18]:
print("In conclusion :")
print(rf"- Error with spurious structure and classical inference : {100*quaderr1:3.5f}%")
print(rf"- Error with spurious structure and do-calculus : {100*quaderr2:3.5f}%")
print(rf"- Error with correct causal structure and do-calculus : {100*quaderr3:3.5f}%")
In conclusion :
- Error with spurious structure and classical inference : 0.07722%
- Error with spurious structure and do-calculus : 0.17362%
- Error with correct causal structure and do-calculus : 0.01588%
In [ ]: