The University of Southampton
Courses

# COMP3224 Causal Reasoning and Machine Learning

## Module Overview

### Aims and Objectives

#### Learning Outcomes

##### Knowledge and Understanding

Having successfully completed this module, you will be able to demonstrate knowledge and understanding of:

• Appreciate the difference between predictive ability and explanatory adequacy
• Identify the necessity of causal reasoning in application domains
• Distinguish between the roles of observational and experimental data
##### Subject Specific Intellectual and Research Skills

Having successfully completed this module you will be able to:

• Ability to construct and reason with deterministic and probabilistic models that represent hypothetical causal mechanisms
• Ability to demonstrate how such models capture changes of probability upon conditioning, upon performing actions or upon posing what-if scenarios.
• Evaluate models and algorithms proposed in the research literature to identify explanatory mechanisms behind data patterns
##### Transferable and Generic Skills

Having successfully completed this module you will be able to:

• Appreciate how working with patterns in data that have societal implications
##### Subject Specific Practical Skills

Having successfully completed this module you will be able to:

• Systematically work with data and within state-of-the-art software environments to learn patterns or concepts
• Create models for simulating data with different explanatory mechanisms

### Syllabus

* ML's limitations: review of relations between questions asked in machine learning and causality * Philosophical titbits: Asymmetry of cause and effect, co-ordination of effects due to hidden causes. * Machinery of probabilistic graphical models - Graphical Markov models; conditional independence and d-separation _ Structural equation modelling - Interventions and do-calculus - Simpson's paradox and confounders - Front-door and back-door criteria for identifying causal effects from observable data * Cause-effect , covariant shifts: If A and B are correlated, what is the direction of the arrow linking A and B? Independence of causal mechanism from input. Covariant shifts and regression modelling. * Representation learning and causality: Disentangling of representations via causal mechanisms and invariant risk minimisation. * Counterfactuals: The ability to answer ``what-if" questions requires a causal mechanism not mere correlations. Application example: eliminating spurious correlations in classification problems. * Potential outcomes, A/B testing and randomised trials: Explaining the relations between different approaches to and techniques in causal analysis. Applications to healthcare. * Fairness and bias: Fairness of algorithms from a process (disparate treatment) or an outcome perspective (disparate outcome). Fairness and bias from a causal lens and a counterfactual perspective.

### Learning and Teaching

#### Teaching and learning methods

Lectures, lab exercises, student-led presentations on specific topics

TypeHours
Specialist Laboratory 20
Lecture24
Total study time150

Judea Pearl and Dana Mackenzie (2018). The Book of Why.

J. Peters, D. Janzing, and B. Schoelkopf (2017). Elements of Causal Inference: Foundations and Learning Algorithms.

J. Pearl, M. Glymour, and N. P. Jewell (2016). Causal Inference in Statistics: A Primer.

### Assessment

#### Assessment Strategy

Coursework only: assessment based on presentations and reports.

#### Summative

MethodPercentage contribution
Analysis and report  (4000 words) 100%

#### Repeat Information

Repeat type: Internal & External