Glossary

MCAR / MAR / MNAR

Missing Completely At Random (the gap is unrelated to anything), Missing At Random (related only to observed values), Missing Not At Random (related to the missing value itself). The emergency dataset is largely MCAR, which makes imputation harder to exploit but fairer to evaluate.

Class imbalance

Some labels appear far more often than others. A model can score high overall while being useless on the rare classes — hence macro-averaged metrics.

Fidelity (synthetic data)

How well synthetic data reproduces the real data’s distributions and the relationships between columns.

Membership-inference / privacy risk

Whether an attacker can tell that a specific real record was in the training set — the main privacy test for a generative model.

Embedding

A learned vector representation of something discrete (a word, a category) so that similar items sit near each other.

Latent space

The compressed internal representation a model works in — e.g. an autoencoder’s bottleneck, or a GAN’s noise input.

Epoch / batch

One epoch is a full pass over the training data; a batch is the small chunk processed before each weight update.

Overfitting / regularisation

Overfitting is learning the training set’s noise instead of its signal; regularisation (dropout, weight decay, early stopping) is any technique that discourages it.

Mode collapse

A GAN failure where the generator produces only a narrow slice of the real variety.

Coverage ratio

In this work, the fraction of the actions that should have been recommended for a case that actually appear somewhere in the model’s ranked lists.

References: missing-data taxonomy — Rubin, “Inference and missing data,” Biometrika 63(3), 1976; membership inference — Shokri et al., IEEE S&P 2017, arXiv:1610.05820; mode collapse — Salimans et al., NeurIPS 2016, arXiv:1606.03498; embedding / latent space / epoch / regularisation — Goodfellow, Bengio & Courville, Deep Learning (2016); coverage ratio — my TeleNotArzt paper, doi:10.5220/0014481000004052.