Description
Fit a mixture model where some cases have known group labels and the rest do not, so the many unlabelled cases sharpen a classifier built from a few labelled ones.
Labelling is expensive: coding open-ended responses, tickets or sessions by hand. In a semi-supervised mixture the EM algorithm keeps the known labels fixed for the labelled cases and uses posterior membership probabilities for the unlabelled ones, so both contribute to estimating each class's distribution. How the mixing proportions are estimated depends on how the labelled cases were chosen: if drawn from the same population they inform proportions, if sampled class by class they do not. More than p labelled cases per class removes likelihood singularities, and the labelled-only estimates give natural starting values. A key case is nonrandom labelling, where only cases above some threshold get a definitive label: fitting the model to all cases remains valid, whereas fitting to labelled cases alone is biased. Gains are largest when classes are close together.
Attributes
- Discipline
- Computational & Machine Learning
- Loss type
- Siloed
- Goal
- Synthesise data, Segment users
- Stage
- Discovery, Post-launch
- Data
- Mixed
- Effort
- Days
Steps
- Assemble labelled and unlabelled cases measured on the same features.
- Record how the labelled cases were selected (random sample, by class, or by a threshold rule).
- Fit the class-conditional models on the labelled cases to obtain starting values.
- Run EM with labels fixed for labelled cases and posterior probabilities for unlabelled ones, estimating proportions according to the selection scheme.
- Classify new cases by the plug-in Bayes rule and check accuracy on a held-out labelled set.
Sources
- Finite Mixture Models (McLachlan & Peel, 2000)
Also known as
- Partial classification
- EM with labelled and unlabelled data
- Updating a classifier with unclassified cases
- Partial nonrandom classification
- Semi-supervised mixture model
Categories
- Analytics & Behaviour
- Computational & Machine Learning