Anti-causal learning

As explained in Causal AI

Anti-causal learning is a learning setting in which features are caused by the label or target variable.

Anti-causal learning predicts a label from features that the label causes: Y → X, rather than the causal-learning direction X → Y. This distinction matters because the usefulness of unlabeled feature observations depends on the causal structure. section 5.3.2

For anti-causal prediction, learning can be viewed as estimating P(X|Y) and P(Y), then obtaining P(Y|X) through Bayes' rule. Unlabeled observations of X can improve the representation of P(X), helping infer P(Y|X). section 5.3.2

The practical implication is a causal test for semi-supervised learning:

• Expect unlabeled data to help when the label causes the features. section 5.3.2
• Treat unlabeled feature data as unlikely to improve the predictive mechanism when features cause the label. section 5.3.2
• Use causal knowledge to avoid spending data, compute, and engineering effort on mismatched algorithms. section 5.4

Handwritten digits

If Y is the handwritten digit and X is its image, the image is treated as an effect of the label. Unlabeled images can therefore contribute to learning a model for predicting the digit from the image. section 5.3.2

Sunscreen and sunburn

The prevalence of sunburn P(O) can help estimate sunscreen use P(C|O) when a sunburn is observed, illustrating how information about an effect can support reverse-direction prediction. section 5.3.2

Figure 5.16 In causal learning, the features cause the label. In anti-causal learning, the label causes the features.

Compare the directed graphs X → Y and X ← Y to identify whether prediction follows causal or anti-causal direction. section 5.3.2

How Causal AI covers it

How deep it goes

Explains distribution decomposition, Bayes' rule, and causal contrasts, while noting that real structures may include unobserved common causes.

Examples

Uses handwritten digit images and sunscreen–sunburn reasoning to illustrate labels causing features and reverse-direction prediction.

What it assumes

Assumes probability notation and Bayes' rule while contrasting feature-to-label and label-to-feature causal directions.

What you can do after

Evaluate whether semi-supervised learning is likely to benefit from unlabeled features given a causal relationship between X and Y.

How it connects

Connects anti-causal learning to causal learning, semi-supervised learning, Bayes' rule, and independence of mechanism.

liveBookmanning.com

Read Causal AI on liveBook Get the book at manning.com

Related concepts

• causal learningcontrasts with · Causal learning is a learning setting in which features or inputs cause the label or target variable.
• semi-supervised learningproduces · Semi-supervised learning is a machine-learning approach that trains models using both labeled and unlabeled data.

This explainer was generated by Manning's knowledge graph from the text of Causal AI. The section links point to the parts of the book it draws on.