Introduction to Bayesian Networks

Date: 1-9-2020

Main point of today: Why are Bayesian Networks useful? Why do we care?

You are smarter than your data. Data do not understand causes and effect; humans do. Judea Pearl, "Book of Why"

Textor applies BN in the development of immunotherapy-based cancer treatments.

Graphs in lecture shows a "bimodal" response to treatment, one group responds and the other does not. Then the question is: can we predict to which group the patient will belong?

Several slides show examples of past intelligent systems that were aiming to "replace" doctors in the sense that they could suggest treatments given symptoms. What's the catch?

Textor: I'm not aware of any major Tech companies AI for which this actually works.

Matthew Herper, Forbes: "IBM's Watson supercomputer recommended 'unsafe and incorrect' cancer treatments". IBM Watson's oncology became a red flag for how to not go to.

How could such a brilliant product in the field of knowledge fail in the other domain of oncology?

Perhaps we do not know enough about cancer yet in order to build a system like this.

Also: correlation does not imply causation. When drawing conclusions about data (in a more causal sense), we use knowledge about the problem. When the answer is not only in the data, we also need knowledge about the data (domain).

Keypoint: Bayesian Networks combine data with knowledge.

Q: What's the difference between data and knowledge/causal mechanisms.

A: One clear indicator of difference is that correlation is symmetric, but causation is not.

Take-away: a Bayesian network is both a

  • statistical model: a representation of data
  • mechanistic model: a representation of the world

Correlation is harder than you think

Correlation patterns can be complex: example in the slide shows where splitting on a simple variable (west vs east Germany) reverses the directionality of a regression line.

But how do we know that including the west vs. east divide drives the switch in direction?

Conditioning principle: (intuition that) the more information you add to something, should get you closer to the "truth". So this would be in favor of taking east vs. west is justified. But unfortunately, it doesn't always hold. The tricky part is: when does it hold, when does it not? How do we find out?

But we saw that adding more data can lead to a different conclusion.

The plot was an example of Simpson's Paradox.

Notice that the chance of being treated is not equal between males and females.

Solution of the Simpson's Paradox:

  • Confounders (in the example: sex) influence the outcome. So we should condition on it to remove this confounding effect.
  • Mediators (in the example: blood pressure): it's part of the causal effect we are trying to find, so if we condition on this we may obscure the relation we're interested in.

If we make assumptions, we can draw a network:

  • Cure cannot affect treatment
  • Treatment doesn't affect gender
  • Treatment can affect blood pressure

Key takeaway: we need to know the network to arrive at a correct solution. Data alone is not enough.

So to revisit the question about why Watson's oncology failed: just much data is not enough, we also need more information and knowledge about correct relations between variables.

Whereas IBM Watson knew the rules of jeopardy, it does not know the "rules" of cancer.

This shows a severe limitation of current AI, namely that it is too focused on data modeling alone and too little on mechanistic components.

Causal inference

No causes in - no causes out. Nancy Cartwright, "Hunting Causes and Using Them"

  • Statistics is the art of describing data.
  • Causal inference, instead, is the art of deriving causal insight from data for:
    • Policy recommendations
    • Intervention targets
    • Explanation

But causal inference can never be unambiguously performed by data alone. We always require extra information or assumptions.

The current hype of machine learning makes this mistake easily, as if no extra information would have to be modelled.

Whether you need causality depends on the problem you are trying to solve.

If it's about prediction stock values, machine learning may be enough. If it's about finding the right patient treatment, it's not. You need more understanding.

Summary

  • BNs are not just prediction algorithms
  • BNs fuse knowledge about the world with observations of the world (data)