STAT 4051 Fall 2026: Introduction to Probabilistic Graphical Models
Overview
Prof Kazeem Adepoju introduces Bayesian networks as directed acyclic graphs that represent joint probability distributions compactly by using conditional independence and local probability tables. He demonstrates inference through factorization, marginalization, and Bayes' theorem, applying the methods to flu symptoms, a hypothetical COVID-19 diagnosis with a stated posterior of about 99.85%, tutoring and course outcomes, rain prediction, and other domains.
Key takeaways
- A Bayesian network factors a joint distribution into one local probability term per node, conditioned on that node's parents, which can make a high-dimensional model more compact.
- In the common-cause graph disease → fever and disease → cough, fever and cough may be associated marginally but are conditionally independent once disease status is known.
- Marginalization answers queries that omit variables by summing over their possible states, as in P(fever, cough) = Σdisease P(fever, cough, disease).
- In the flu example, P(fever|flu)P(cough|flu) = 0.8 × 0.7 = 0.56, illustrating the conditional-independence factorization for the two symptoms.
- Bayes' theorem combines evidence likelihoods with prior probabilities; the lecture's hypothetical COVID-19 example reports a posterior of about 99.85% while explicitly noting it is not a validated clinical estimate.
- Graph structures encode different dependencies: a chain models tutoring → mastery → passing, while a collider represents two causes converging on one effect.
Chapters
- Prof Kazeem Adepoju reports assignment scores averaged 98.16, compared with a midterm mean of 84.17 and median of 89.50.
- Midterm scores ranged from 39 to 100; Adepoju says he will use the median when describing the left-skewed score distribution.
- Students who missed one or two classes for a genuine reason can contact him about a possible three-point adjustment; he also encourages struggling students to attend office hours.
- Bayesian networks are introduced as probabilistic graphical models that use graph structure and probability theory to represent high-dimensional distributions compactly.
- The central task is inference: use observed variables to estimate an unobserved variable, such as the probability of burglary after two alarm systems sound.
- The network structure supports factorizing a joint distribution into local conditional probabilities.
- Adepoju explains conditional independence with a common-cause structure: C causes both A and B, written A ← C → B.
- Before observing C, A and B can be statistically associated; conditioning on C makes them independent in this structure.
- For this graph, the joint distribution factorizes as P(C)P(A|C)P(B|C), and P(A,B|C) = P(A|C)P(B|C).
- The medical example treats disease C as a cause of symptoms A and B, such as flu causing fever and cough.
- The joint probability of disease and both symptoms is P(A,B,C) = P(C)P(A|C)P(B|C).
- To find the probability of both symptoms without specifying disease status, sum over disease states: P(A,B) = Σc P(A,B,c).
- Conditioning on the known disease status removes the symptoms' association under the common-cause model.
- The worked example sets P(flu) = 0.2, P(fever|flu) = 0.8, P(cough|flu) = 0.7, P(fever|no flu) = 0.1, and P(cough|no flu) = 0.2.
- Marginal probabilities are obtained by summing over flu and no-flu cases; the joint probability of fever and cough is about 0.13, while fever alone is about 0.24.
- Fever and cough are not marginally independent because their joint probability differs from the product of their marginal probabilities.
- Given flu, the joint symptom probability is 0.8 × 0.7 = 0.56, matching the conditional-independence factorization.
- Adepoju models COVID-19 status as a cause of fever, cough, loss of smell, and a positive diagnostic test.
- The likelihood of the observed evidence is calculated separately for infected and uninfected cases by multiplying the relevant conditional probabilities in the network.
- Bayes' theorem combines those likelihoods with prior probabilities to estimate P(COVID-19 | symptoms and test result).
- The hypothetical calculation gives a posterior near 99.85%; Adepoju cautions that this is not a validated clinical estimate.
- A chain structure represents tutoring leading to mastery, which in turn affects whether a student passes: A → C → B.
- The joint distribution factorizes as P(A)P(C|A)P(B|C), following each node's parents in the graph.
- To calculate P(B|A), the example sums over both possible mastery states, C and not-C, rather than ignoring the intermediate variable.
- A collider occurs when two causes point to the same effect, represented as A → C ← B.
- The rain example links rain and sprinkler use to wet grass, with rain also influencing whether the sprinkler is used.
- To calculate P(wet grass), marginalize over the unobserved rain and sprinkler variables; conditional queries can retain whichever evidence is observed.
- Adepoju lists applications including COVID-19 and lung-disease diagnosis, where multiple clinical variables contribute evidence about a condition.
- A supply-chain network tracks three suppliers and the effects of local or global disruptions.
- Other examples include juvenile justice and repeat offending, the Simpson case, and weather forecasting.
- In the weather network, wind speed and humidity influence rain, while humidity at 9 a.m. can affect humidity at 3 p.m.; further calculations are planned for Wednesday.
Transcript excerpt
0:00 Good morning everyone. I want to welcome you to another um statistical machine learning class. Uh before I get started today, I want to say uh we've done with all uh grading and recording. Um I think uh 90 students out of 93 all wrote this uh M one exam and I still have like uh three students who have not um written the exam. I'm going to be writing them today. I want to hear from them. So I updated all results this morning. Uh if you check
0:47 your grade now, you're going to see that the grade are all are now available. I I did um you know match analysis, you know, I tried to investigate relationship uh between your assignment scored. You know, assignment is unprotoed and of course the midtime is proctored. So uh this is what I actually got. The mean in the assignment is 98.16 and the mean in the midtime is 84.17. And of course uh the median uh in the assignment is 100 and the median in the test
1:35 is um 89.50. So looking into um the box plots I realized that uh as for the midtime um the distribution is kind of skew to the left. So and that's why I'm going to go with the median instead of the mean. So it's a very good performance or even though I uh actually expected average to be hey but um I have to give kudos to every one of you. Um, of course, uh, by I know by the time we're going to ride Mid Time 2,
2:24 uh, you go, your performance is going to be better than this. But good job. You know, uh, I think I have a minimal minimum is 39 in the meantime and the maximum is 100.
This is the start of the transcript. Save this video free for the whole transcript, searchable and time-coded.
Related notes
Summary, takeaways, and chapters were generated by AI from the video's transcript and may contain errors. The video belongs to its creator, Prof Kazeem Adepoju.