Step by Step
D
Directed Acyclic Graph — the structure
A Bayesian network is structured as a directed acyclic graph (DAG) — nodes represent random variables, and directed edges represent direct probabilistic influence, with no cycles allowed.
Example: an edge from "Smoking" to "Lung Cancer" represents that smoking directly influences lung cancer risk.
C
Conditional Probability Tables — the numbers behind each node
Each node stores a Conditional Probability Table (CPT) specifying the probability of that node's value given each possible combination of its parent nodes' values.
Example: a "Lung Cancer" node's CPT might specify P(Cancer|Smoking=Yes) = 0.3 and P(Cancer|Smoking=No) = 0.02.
I
Conditional independence — the key structural property
Given its parents, a node is conditionally independent of all its non-descendants — meaning once you know a node's direct causes (parents), additional information about unrelated, non-descendant nodes doesn't change your belief about that node.
Example: once you know whether a patient smokes (the parent of lung cancer), learning about an unrelated non-descendant variable, like their favorite color, provides no additional information about their lung cancer probability.
Applied Walkthrough
1
A medical diagnosis Bayesian network includes nodes for Smoking, Lung Cancer, and Chest X-Ray Result, with edges Smoking → Lung Cancer → Chest X-Ray Result.
2
Given that we already know a patient's Smoking status (Lung Cancer's direct parent), does learning some entirely unrelated non-descendant fact change our belief about Lung Cancer? No — Lung Cancer is conditionally independent of unrelated non-descendants once we know its parent (Smoking).
3
However, the Chest X-Ray Result IS a descendant of Lung Cancer, so information about the X-ray result CAN still update our belief about Lung Cancer, since it isn't excluded by the conditional independence property.
4
This distinction — parents block the flow of irrelevant information from non-descendants, but descendants can still provide relevant evidence — is central to how inference works in Bayesian networks.
Exam Application
Exams test whether you understand the DAG structure (nodes as variables, edges as direct probabilistic influence, no cycles), the role of CPTs, and specifically the conditional independence property — that a node is independent of non-descendants given its parents, which is the mathematical foundation for efficient inference. Also expect a note that exact inference in Bayesian networks is NP-hard in general.
⚠ Common Trap
The most common trap is confusing "conditional independence given parents" with unconditional independence, or forgetting that this independence only applies to NON-DESCENDANTS — a node's descendants can still carry relevant information even after conditioning on its parents.
✓ Quick Self-Check
1. What structure does a Bayesian network use?
A Directed Acyclic Graph (DAG), with nodes as random variables and edges representing direct probabilistic influence.
Tap to reveal / hide
2. What does a node's Conditional Probability Table (CPT) specify?
The probability of that node's value given each possible combination of its parent nodes' values.
Tap to reveal / hide
3. What is the key conditional independence property in a Bayesian network?
Given its parents, a node is conditionally independent of all its non-descendants.
Tap to reveal / hide
4. Does a node's descendant provide relevant information about that node, even after conditioning on its parents?
Yes — the conditional independence property only excludes non-descendants, not descendants.
Tap to reveal / hide
5. Is exact inference in Bayesian networks generally considered computationally easy or hard?
Hard — exact inference is NP-hard in general.
Tap to reveal / hide