Skip to content

Probablistic Inference with pyAgrum

Creative Commons LicenseaGrUMinteractive online version

In this notebook, we will show different basic features for probabilistic inference on Bayesian networks using pyagrum.

First we need some external modules:

Then we import pyagrum and the pyAgrum’s notebook module, that offers very usefull methods when writting a notebook.

This first example shows how you can load a BayesNet and show it as graph. Note that pyAgrum handles serveral BayesNet file format such as DSL, BIF and UAI.

import pyagrum as gum
import pyagrum.lib.notebook as gnb
bn = gum.loadBN("res/alarm.bgum")
gnb.showBN(bn, size="9")

svg

print(bn)
BN{nodes: 37, arcs: 46, domainSize: 10^16.2389, dim: 509, mem: 5Ko 896o}

From there, it is easy to get a posterior using an inference engine :

ie = gum.LazyPropagation(bn)
ie.makeInference()
print(ie.posterior(bn.idFromName("CATECHOL")))
CATECHOL │
NORMAL │HIGH │
─────────│─────────│
0.0512 │ 0.9488 │

But since we are in notebook, why not use pyAgrum notebook’s methods ?

gnb.showPosterior(bn, evs={}, target="CATECHOL")
PyAgrum inline image

You may also want to see the graph with some posteriors

## due to matplotlib, format is forced to png.
gnb.showInference(bn, evs={}, targets={"VENTALV", "CATECHOL", "HR", "MINVOLSET"}, size="11")

svg

.. and then observe the impact of evidence :

gnb.showInference(
bn,
evs={"CO": 1, "VENTLUNG": 1},
targets={
"VENTALV",
"CATECHOL",
"HR",
"MINVOLSET",
"ANAPHYLAXIS",
"STROKEVOLUME",
"ERRLOWOUTPUT",
"HBR",
"PULMEMBOLUS",
"HISTORY",
"BP",
"PRESS",
"CO",
},
size="10",
)

svg

You can even compute all posteriors by leaving the targets parameter empty (which is its default value).

gnb.showInference(bn, evs={"CO": 1, "VENTLUNG": 1}, size="14")

svg

To have a global view of the knowledge brought by the inference, you can also draw the entropy of all nodes

from pyagrum import explain
explain.showInformation(bn, {}, size="14")
Unnamed KINKEDTUBE KINKEDTUBE VENTLUNG VENTLUNG KINKEDTUBE->VENTLUNG PRESS PRESS KINKEDTUBE->PRESS HYPOVOLEMIA HYPOVOLEMIA STROKEVOLUME STROKEVOLUME HYPOVOLEMIA->STROKEVOLUME LVEDVOLUME LVEDVOLUME HYPOVOLEMIA->LVEDVOLUME INTUBATION INTUBATION SHUNT SHUNT INTUBATION->SHUNT INTUBATION->VENTLUNG MINVOL MINVOL INTUBATION->MINVOL INTUBATION->PRESS VENTALV VENTALV INTUBATION->VENTALV MINVOLSET MINVOLSET VENTMACH VENTMACH MINVOLSET->VENTMACH PULMEMBOLUS PULMEMBOLUS PAP PAP PULMEMBOLUS->PAP PULMEMBOLUS->SHUNT INSUFFANESTH INSUFFANESTH CATECHOL CATECHOL INSUFFANESTH->CATECHOL ERRLOWOUTPUT ERRLOWOUTPUT HRBP HRBP ERRLOWOUTPUT->HRBP ERRCAUTER ERRCAUTER HRSAT HRSAT ERRCAUTER->HRSAT HREKG HREKG ERRCAUTER->HREKG FIO2 FIO2 PVSAT PVSAT FIO2->PVSAT LVFAILURE LVFAILURE LVFAILURE->STROKEVOLUME LVFAILURE->LVEDVOLUME HISTORY HISTORY LVFAILURE->HISTORY DISCONNECT DISCONNECT VENTTUBE VENTTUBE DISCONNECT->VENTTUBE ANAPHYLAXIS ANAPHYLAXIS TPR TPR ANAPHYLAXIS->TPR CO CO STROKEVOLUME->CO TPR->CATECHOL BP BP TPR->BP PCWP PCWP LVEDVOLUME->PCWP CVP CVP LVEDVOLUME->CVP VENTMACH->VENTTUBE SAO2 SAO2 SHUNT->SAO2 VENTTUBE->VENTLUNG VENTTUBE->PRESS VENTLUNG->MINVOL VENTLUNG->VENTALV EXPCO2 EXPCO2 VENTLUNG->EXPCO2 ARTCO2 ARTCO2 VENTALV->ARTCO2 VENTALV->PVSAT ARTCO2->EXPCO2 ARTCO2->CATECHOL PVSAT->SAO2 SAO2->CATECHOL HR HR CATECHOL->HR HR->HRBP HR->HRSAT HR->CO HR->HREKG CO->BP
PyAgrum inline image

… and then observe the impact of an evidence on the whole bayes network :

explain.showInformation(bn, {"CO": 0}, size="9")
Unnamed KINKEDTUBE KINKEDTUBE VENTLUNG VENTLUNG KINKEDTUBE->VENTLUNG PRESS PRESS KINKEDTUBE->PRESS HYPOVOLEMIA HYPOVOLEMIA STROKEVOLUME STROKEVOLUME HYPOVOLEMIA->STROKEVOLUME LVEDVOLUME LVEDVOLUME HYPOVOLEMIA->LVEDVOLUME INTUBATION INTUBATION SHUNT SHUNT INTUBATION->SHUNT INTUBATION->VENTLUNG MINVOL MINVOL INTUBATION->MINVOL INTUBATION->PRESS VENTALV VENTALV INTUBATION->VENTALV MINVOLSET MINVOLSET VENTMACH VENTMACH MINVOLSET->VENTMACH PULMEMBOLUS PULMEMBOLUS PAP PAP PULMEMBOLUS->PAP PULMEMBOLUS->SHUNT INSUFFANESTH INSUFFANESTH CATECHOL CATECHOL INSUFFANESTH->CATECHOL ERRLOWOUTPUT ERRLOWOUTPUT HRBP HRBP ERRLOWOUTPUT->HRBP ERRCAUTER ERRCAUTER HRSAT HRSAT ERRCAUTER->HRSAT HREKG HREKG ERRCAUTER->HREKG FIO2 FIO2 PVSAT PVSAT FIO2->PVSAT LVFAILURE LVFAILURE LVFAILURE->STROKEVOLUME LVFAILURE->LVEDVOLUME HISTORY HISTORY LVFAILURE->HISTORY DISCONNECT DISCONNECT VENTTUBE VENTTUBE DISCONNECT->VENTTUBE ANAPHYLAXIS ANAPHYLAXIS TPR TPR ANAPHYLAXIS->TPR CO CO STROKEVOLUME->CO TPR->CATECHOL BP BP TPR->BP PCWP PCWP LVEDVOLUME->PCWP CVP CVP LVEDVOLUME->CVP VENTMACH->VENTTUBE SAO2 SAO2 SHUNT->SAO2 VENTTUBE->VENTLUNG VENTTUBE->PRESS VENTLUNG->MINVOL VENTLUNG->VENTALV EXPCO2 EXPCO2 VENTLUNG->EXPCO2 ARTCO2 ARTCO2 VENTALV->ARTCO2 VENTALV->PVSAT ARTCO2->EXPCO2 ARTCO2->CATECHOL PVSAT->SAO2 SAO2->CATECHOL HR HR CATECHOL->HR HR->HRBP HR->HRSAT HR->CO HR->HREKG CO->BP
PyAgrum inline image

Lazy Propagation, like several other inference algorithms, uses a junction tree to propagate information.

You can show the junction tree used by Lazy Propagation with pyAgrum:

jt = ie.junctionTree()
gnb.showJunctionTree(bn, size="12")

svg

## another representation of the junction, more convenient for investigating the flow of data in the jt
## the size/width of cliques and separators are proportionnal to the number of nodes in the factor.
jt.map()
0 0~16 0--0~16 1 1~32 1--1~32 2 2~33 2--2~33 3 3~4 3--3~4 4 4~22 4--4~22 5 5~22 5--5~22 6 6~23 6--6~23 7 7~26 7--7~26 8 8~17 8--8~17 10 10~14 10--10~14 11 11~16 11--11~16 12 12~13 12--12~13 13 13~30 13--13~30 14 14~26 14--14~26 16 16~17 16--16~17 17 17~24 17--17~24 19 19~27 19--19~27 20 20~33 20--20~33 22 22~33 22--22~33 23 23~27 23--23~27 23~31 23--23~31 24 24~26 24--24~26 26 26~27 26--26~27 27 30 30~31 30--30~31 31 31~32 31--31~32 32 32~33 32--32~33 33 19~27--27 12~13--13 2~33--33 23~27--27 22~33--33 11~16--16 24~26--26 31~32--32 10~14--14 26~27--27 13~30--30 5~22--22 7~26--26 20~33--33 16~17--17 32~33--33 23~31--31 8~17--17 1~32--32 3~4--4 4~22--22 17~24--24 14~26--26 30~31--31 6~23--23 0~16--16

One can easily walk through the junction tree.

for n in jt.nodes():
print([bn.variable(n).name() for n in jt.clique(n)])
['CVP', 'LVEDVOLUME']
['FIO2', 'VENTALV', 'PVSAT']
['ARTCO2', 'EXPCO2', 'VENTLUNG']
['VENTMACH', 'MINVOLSET']
['VENTMACH', 'DISCONNECT', 'VENTTUBE']
['PRESS', 'KINKEDTUBE', 'INTUBATION', 'VENTTUBE']
['ANAPHYLAXIS', 'TPR']
['HRBP', 'ERRLOWOUTPUT', 'HR']
['LVFAILURE', 'HISTORY']
['HREKG', 'HR', 'ERRCAUTER']
['PCWP', 'LVEDVOLUME']
['PAP', 'PULMEMBOLUS']
['SHUNT', 'INTUBATION', 'PULMEMBOLUS']
['HRSAT', 'HR', 'ERRCAUTER']
['LVFAILURE', 'HYPOVOLEMIA', 'LVEDVOLUME']
['HYPOVOLEMIA', 'STROKEVOLUME', 'LVFAILURE']
['CO', 'BP', 'TPR']
['INTUBATION', 'VENTLUNG', 'MINVOL']
['KINKEDTUBE', 'INTUBATION', 'VENTTUBE', 'VENTLUNG']
['INSUFFANESTH', 'TPR', 'ARTCO2', 'SAO2', 'CATECHOL']
['CO', 'STROKEVOLUME', 'HR']
['CO', 'CATECHOL', 'HR']
['CO', 'TPR', 'CATECHOL']
['INTUBATION', 'SHUNT', 'PVSAT', 'SAO2']
['INTUBATION', 'ARTCO2', 'PVSAT', 'SAO2']
['ARTCO2', 'VENTALV', 'INTUBATION', 'PVSAT']
['VENTALV', 'INTUBATION', 'ARTCO2', 'VENTLUNG']
for e in jt.edges():
print(f"Separator for {e} : {jt.clique(e[0]).intersection(jt.clique(e[1]))}")
Separator for (13, 30) : {18, 2}
Separator for (2, 33) : {26, 22}
Separator for (3, 4) : {16}
Separator for (26, 27) : {34, 30}
Separator for (7, 26) : {31}
Separator for (12, 13) : {4}
Separator for (31, 32) : {27, 26, 2}
Separator for (23, 31) : {26, 28}
Separator for (5, 22) : {0, 2, 20}
Separator for (17, 24) : {13}
Separator for (19, 27) : {34, 14}
Separator for (24, 26) : {34, 31}
Separator for (32, 33) : {25, 2, 26}
Separator for (6, 23) : {14}
Separator for (23, 27) : {14, 30}
Separator for (11, 16) : {15}
Separator for (10, 14) : {7, 31}
Separator for (8, 17) : {9}
Separator for (0, 16) : {15}
Separator for (1, 32) : {25, 27}
Separator for (20, 33) : {2, 22}
Separator for (4, 22) : {20}
Separator for (14, 26) : {31}
Separator for (22, 33) : {2, 22}
Separator for (30, 31) : {2, 27, 28}
Separator for (16, 17) : {1, 9}
jt.hasRunningIntersection()
True

The junction tree created by a LazyPropagation is optimized for the query (see RelevanceReasonning notebook). But you can also introspect a junction tree directly from a BN or a graph using the JunctionTreeGenerator’s class.

bn = gum.fastBN("0->1->2<-3->4->5->6<-2->7")
jtg = gum.JunctionTreeGenerator()
gnb.sideBySide(
bn,
jtg.junctionTree(bn),
jtg.eliminationOrder(bn),
jtg.binaryJoinTree(bn),
captions=[
"A Bayesien network",
"a junction tree for this BN",
"its elimination order",
"an (optimized) binary join tree",
],
)
anonymousBN 0 0 1 1 0->1 2 2 1->2 6 6 2->6 7 7 2->7 3 3 3->2 4 4 3->4 5 5 4->5 5->6
A Bayesien network
0 2-5-6 0~4 2-5 0--0~4 1 2-7 1~4 2 1--1~4 2 0-1 2~3 1 2--2~3 3 1-2-3 3~4 2-3 3--3~4 4 2-3-5 4~5 3-5 4--4~5 5 3-4-5 2~3--3 4~5--5 0~4--4 1~4--4 3~4--4
a junction tree for this BN
[6, 7, 0, 1, 2, 3, 4, 5]
its elimination order
0 2-5-6 0~6 2-5 0--0~6 1 2-7 1~6 2 1--1~6 2 0-1 2~3 1 2--2~3 3 1-2-3 3~4 2-3 3--3~4 4 2-3-5 4~5 3-5 4--4~5 4~6 2-5 4--4~6 5 3-4-5 6 2-5 2~3--3 4~5--5 3~4--4 1~6--6 0~6--6 4~6--6
an (optimized) binary join tree

junction tree from graphs (using uniform domainSize)

Section titled “junction tree from graphs (using uniform domainSize)”
## creating a dag slightly different
dag = bn.dag()
dag.addArc(0, 3)
dag.addArc(0, 7)
gnb.sideBySide(
dag,
dag.moralGraph(),
jtg.junctionTree(dag),
jtg.eliminationOrder(dag),
jtg.binaryJoinTree(dag),
captions=[
"A DAG",
"its moral graph",
"a junction tree for this dag (with partial order)",
"its elimination order (with partial order)",
"an (optipmized) binary jointree (with partial order)",
],
)
0 (0) 0 1 (1) 1 0->1 3 (3) 3 0->3 7 (7) 7 0->7 2 (2) 2 1->2 6 (6) 6 2->6 2->7 3->2 4 (4) 4 3->4 5 (5) 5 4->5 5->6
A DAG
0 (0) 0 1 (1) 1 0--1 2 (2) 2 0--2 3 (3) 3 0--3 7 (7) 7 0--7 1--2 1--3 2--3 5 (5) 5 2--5 6 (6) 6 2--6 2--7 4 (4) 4 3--4 4--5 5--6
its moral graph
0 2-5-6 0~4 2-5 0--0~4 1 0-1-2-3 1~4 2-3 1--1~4 1~2 0-2 1--1~2 2 0-2-7 4 2-3-5 4~5 3-5 4--4~5 5 3-4-5 4~5--5 0~4--4 1~4--4 1~2--2
a junction tree for this dag (with partial order)
[6, 1, 7, 0, 2, 3, 4, 5]
its elimination order (with partial order)
0 2-5-6 0~4 2-5 0--0~4 1 0-1-2-3 1~4 2-3 1--1~4 1~2 0-2 1--1~2 2 0-2-7 4 2-3-5 4~5 3-5 4--4~5 5 3-4-5 4~5--5 0~4--4 1~4--4 1~2--2
an (optipmized) binary jointree (with partial order)
## creating an undigraph slightly different
ug = bn.dag().moralGraph()
ug.addEdge(0, 7)
gnb.sideBySide(
ug,
jtg.junctionTree(ug),
jtg.eliminationOrder(ug),
jtg.binaryJoinTree(ug),
captions=[
"A undigraph",
"a junction tree for this undigraph",
"its elimination order",
"an (optipmized) binary jointree",
],
)
0 (0) 0 1 (1) 1 0--1 7 (7) 7 0--7 2 (2) 2 1--2 3 (3) 3 1--3 2--3 5 (5) 5 2--5 6 (6) 6 2--6 2--7 4 (4) 4 3--4 4--5 5--6
A undigraph
0 2-5-6 0~2 2-5 0--0~2 1 3-4-5 1~2 3-5 1--1~2 2 2-3-5 2~3 2-3 2--2~3 3 1-2-3 3~4 1-2 3--3~4 4 1-2-7 4~5 1-7 4--4~5 5 0-1-7 2~3--3 4~5--5 0~2--2 1~2--2 3~4--4
a junction tree for this undigraph
[6, 4, 5, 3, 2, 1, 7, 0]
its elimination order
0 2-5-6 0~2 2-5 0--0~2 1 3-4-5 1~2 3-5 1--1~2 2 2-3-5 2~3 2-3 2--2~3 3 1-2-3 3~4 1-2 3--3~4 4 1-2-7 4~5 1-7 4--4~5 5 0-1-7 2~3--3 4~5--5 0~2--2 1~2--2 3~4--4
an (optipmized) binary jointree

Using partial order to specify the elimination order

Section titled “Using partial order to specify the elimination order”
## adding a partial order for the elimination order
po = [[1, 2, 3], [0, 4, 7], [5, 6]]
gnb.sideBySide(
bn,
jtg.junctionTree(bn, po),
jtg.eliminationOrder(bn, po),
jtg.binaryJoinTree(bn),
captions=[
"A Bayesien network",
"a junction tree for this BN using partial order",
"its elimination order following partial order",
"an (optimized) binary join tree",
],
)
anonymousBN 0 0 1 1 0->1 2 2 1->2 6 6 2->6 7 7 2->7 3 3 3->2 4 4 3->4 5 5 4->5 5->6
A Bayesien network
0 1-2-3-4 0~1 1-2-4 0--0~1 1 0-1-2-4 1~2 0-2-4 1--1~2 2 0-2-4-5-6-7 0~1--1 1~2--2
a junction tree for this BN using partial order
[3, 1, 2, 7, 0, 4, 6, 5]
its elimination order following partial order
0 2-5-6 0~6 2-5 0--0~6 1 2-7 1~6 2 1--1~6 2 0-1 2~3 1 2--2~3 3 1-2-3 3~4 2-3 3--3~4 4 2-3-5 4~5 3-5 4--4~5 4~6 2-5 4--4~6 5 3-4-5 6 2-5 2~3--3 4~5--5 3~4--4 1~6--6 0~6--6 4~6--6
an (optimized) binary join tree
## adding a partial order for the elimination order also for the graphs
po = [[0, 4, 7], [1, 3], [5, 6, 2]]
## creating a dag slightly different
dag = bn.dag()
dag.addArc(0, 3)
dag.addArc(0, 7)
gnb.sideBySide(
dag,
dag.moralGraph(),
jtg.junctionTree(dag, po),
jtg.eliminationOrder(dag, po),
jtg.binaryJoinTree(dag, po),
captions=[
"A DAG",
"its moral graph",
"a junction tree for this dag (with partial order)",
"its elimination order (with partial order)",
"an (optimized) binary jointree (with partial order)",
],
)
0 (0) 0 1 (1) 1 0->1 3 (3) 3 0->3 7 (7) 7 0->7 2 (2) 2 1->2 6 (6) 6 2->6 2->7 3->2 4 (4) 4 3->4 5 (5) 5 4->5 5->6
A DAG
0 (0) 0 1 (1) 1 0--1 2 (2) 2 0--2 3 (3) 3 0--3 7 (7) 7 0--7 1--2 1--3 2--3 5 (5) 5 2--5 6 (6) 6 2--6 2--7 4 (4) 4 3--4 4--5 5--6
its moral graph
0 0-2-7 0~1 0-2 0--0~1 1 0-1-2-3 1~4 2-3 1--1~4 2 3-4-5 2~4 3-5 2--2~4 4 2-3-5 4~5 2-5 4--4~5 5 2-5-6 0~1--1 4~5--5 2~4--4 1~4--4
a junction tree for this dag (with partial order)
[7, 0, 4, 1, 3, 2, 6, 5]
its elimination order (with partial order)
0 0-2-7 0~1 0-2 0--0~1 1 0-1-2-3 1~4 2-3 1--1~4 2 3-4-5 2~4 3-5 2--2~4 4 2-3-5 4~5 2-5 4--4~5 5 2-5-6 0~1--1 4~5--5 2~4--4 1~4--4
an (optimized) binary jointree (with partial order)