Causal learning and modeling David Danks CMU Philosophy & Psychology 2014 NASSLLI.

Causal learning and modeling

David DanksCMU Philosophy & Psychology

2014 NASSLLI

High-level overview

Monday: History of causal inference Basic representation of causal structures

Tuesday: Inference & reasoning using graphical

models Interventions in causal structures

High-level overview

Wednesday: Basic principles of search & causal

discovery

Thursday: Challenges to causal discovery, and

responses Both principled and real-world

High-level overview

Friday: One of two possibilities: Singular / actual causation &

counterfactuals (in the causal graphical model framework)

Recent advances in causal learning & inference

Decided by a vote at end-of-class tomorrow (Tues)

Structure & assumptions

Mix of lecture & (group) problem-solving, so if you have questions/uncertainty,

Ask! If you’re confused, then someone else probably

is too…

Assuming basic knowledge of probabilities Focus is on conceptual/foundational issues, not

the technical details But ask if you want to know more about those

details!

A Brief History of Causal Discovery

“Big Picture” (very roughly)

Greeks - 1750: Unhelpful platitudes

1750 - 1950: Practical successes

1950 - present: Computers + Formal models = principled methods

Aristotle

384-322 BC Trying to answer:

“Why does X have A?” Four types of ‘cause’

Formal: Because of its structure Material: Because of its composition Efficient: Because of its development Final: Because of its purpose

But no systematic theory of inference

Francis Bacon

1561-1626 Novum Organum (1620)

For any phenomenon, construct: The table of presence (tabula praesentiae) The table of absence (tabula absentiae) The table of degrees (tabula graduum)

The cause of the phenomenon is the set of properties that explains every case on each of the three tables

John Stuart Mill

1806-1873 System of Logic (1843)

Algorithmic form of Bacon’s method (though unattributed) Method of agreement Method of difference Method of concomitant variation

David Hume

1711-1776 Causal inference cannot be

done using deduction It is always logically possible that future

“causes” will not be followed by the effect Actually a general argument about

induction But we do it by “custom or habit”

Had an evolutionary justification, but no framework in which to express it

Responses to Hume’s skepticism Hume’s arguments were quite influential

in philosophical circles And still matter in present-day philosophy

But in the sciences, people were starting to find methods that (sometimes) gave answers that at least seemed right…

Regression (Least Squares)

18th c. astronomy: find the “best” values for 6 unknowns given 75 observations Euler (1748)

Failed due to computational intractability Legendre (1805)

Developed the method of least squares Gauss (1795 / 1809)

Independent (earlier, unpublished) discovery & justification

Still the most common causal inference method…

Growth of statistics

Early theory of statistics emerges from probability theory throughout the 1800s

Laplace1749 1827

Quetelet1796 1874

Galton1822 1911

Pearson1857 1936

Spearman1863 1945

Yule1871 1951

1800 1900

Ronald A. Fisher

1890-1962 Essentially the father of modern

statistics, and developed: An array of statistical tests An analysis of various experimental designs The standard statistical and methodological

reference texts for a generation of scientists

Sewall Wright

1889-1988 Path analysis

Graphs encode high-levelstructure, and then regressioncan be used to estimate parameters

By mid-20th c., it had been adopted by a number of economists and sociologists

But no search procedures were provided Have to know the high-level structure

Causal graphical models

Developed by statisticians, computer scientists, and philosophers Dawid, Spiegelhalter, Wermuth, Cox,

Lauritzen, Pearl, Spirtes, Glymour, Scheines Represent both qualitative and

quantitative aspects of causation

Representing Causal Structures

Qualitative representation

We want a representation that captures many qualitative features of causality

Qualitative representation

We want a representation that captures many qualitative features of causality Causation occurs among variables ⇒

One node per variable

Qualitative representation

We want a representation that captures many qualitative features of causality Causation occurs among variables ⇒

One node per variable

Exercise

FoodEaten

Weight

Metabolism

Qualitative representation

We want a representation that captures many qualitative features of causality Asymmetry of causation ⇒

Need an asymmetric connection in the graph

Exercise

FoodEaten

Weight

Metabolism

Qualitative representation

We want a representation that captures many qualitative features of causality Asymmetry of causation ⇒

Need an asymmetric connection in the graph

Exercise

FoodEaten

Weight

Metabolism

Qualitative representation

We want a representation that captures many qualitative features of causality No (immediate) reciprocal causation ⇒

No cycles (without explicit temporal indexing)

Exercise

FoodEaten

Weight

Metabolism

Qualitative representation

We want a representation that captures many qualitative features of causality No (immediate) reciprocal causation ⇒

No cycles (without explicit temporal indexing)

Exercise

FoodEaten

Weight

Metabolism

Exercise

FoodEaten

Weight

Metabolism

Time t Time t+1

Directed Acyclic Graphs

More precise: DAG G = <V, E> V = set of nodes (for variables) E = set of edges (i.e., ordered pairs of nodes)

Path π = sequence of adjacent edges Directed path = path with all edges same direction

Acyclicity: No directed path from node A to itself

In general: We use genealogical & topological language to describe graphical relationships

Quantitative representation

DAGs alone can represent “A causes B”…but not “strength” or “form” of causation Need to represent the relationships

between the various variables states Exact quantitative representation will

depend on the type of variables being represented

Bayesian networks

All variables are discrete/categorical Represent quantitative causation using a

joint probability distribution I.e., a specification of the probability of any

combination of variable values, such as: P(E=Hi & FE=Lo & M=Hi & W=Hi) = 0.001;

P(E=Hi & FE=Lo & M=Hi & W=Lo) = 0.03;etc.

Note: Nothing inherently Bayesian about Bayes nets!

Structural Equation Models (SEMs) All variables are continuous/real-valued Represent quantitative causation using

systems of linear equations For example:

Exercise = a1FE + a2M + a3W + εE_noise

FE = b1E + b2M + b3W + εFE_noise

etc.

Connecting the pieces

DAG-based graphical model:

P(X) = P(X1) P(X2 | X1)

P(X3 | X1) P(X4 | X1,X2)

QuantitativeQualitative???

Connecting the pieces

Causal Markov assumption: Variables are independent of their non-

effects conditional on their direct causes Use the qualitative graph to constrain the

quantitative relationships

Encodes the intuition of “screening off” Given the values of the direct causes, learning

the value of a non-effect doesn’t help me predict

Connecting the pieces

Markov assumption for Bayes nets ⇒ Markov factorization of P(X1, X2, …):

Connecting the pieces

Markov assumption for Bayes nets: Markov factorization of P(X1, X2, …): Example:

Exercise

FoodEaten

Weight

Metabolism

P(E, FE, W, M) = P(E) * P(FE | E) *P(M | E) * P(W | M, FE)

⇒

Connecting the pieces

Markov assumption for SEMs: Markov factorization of joint probability

density:

Connecting the pieces

Markov assumption for SEMs: Markov factorization of joint probability

density: Example:

Exercise

FoodEaten

Weight

Metabolism

E = εE_noise

FE = a1E + εFE_noise

M = b1E + εM_noise

W = c1FE + c2M + εC_noise

⇒

Connecting the pieces

Causal Faithfulness assumption The only independencies are those predicted by

the Markov assumption Uses the quantitative relations to constrain the

qualitative graph Implication: No exactly counter-balancing causal

paths Exercise → Food Eaten → Weight and

Exercise → Metabolism → Weightdo not exactly offset one another

Implication: No perfectly deterministic relationships In particular, no variable is a mathematical function of

others

Causal vs. statistical models

Bayes nets and SEMs are not inherently causal models Markov and Faithfulness assumptions can

be expressed purely as graph-quant. constraints

Assuming a non-causal version of the assumptions ⇒ purely statistical model I.e., a compact representation of statistical

independencies among some set of variables

Causation and intervention

Causal claims support counterfactuals In particular, those about interventions

“If I had flipped the switch, the light would have turned on”

“If she hadn’t dropped the plate, then it would not have broken”

Etc.

Causation and intervention

One of the central causal asymmetries Interventions on a cause lead to changes in the

effect Flipping the switch turns off the light

In contrast, interventions on an effect do not lead to changes in the cause Breaking the light bulb doesn’t flip the switch

Some have argued that this is the paradigmatic feature of causation (Woodward, Hausman)

Looking ahead…

Have: Basic formal representation for causation

Need: Fundamental causal asymmetry (of

intervention) Inference & reasoning methods Search & causal discovery methods

Looking ahead…

Have: Basic formal representation for causation

Need: Fundamental causal asymmetry (of

intervention) Inference & reasoning methods Search & causal discovery methods