lowcardinality

Low cardinality is a term used in statistics and data management to describe a categorical variable that has a relatively small number of distinct values compared with the size of the dataset. Cardinality is the count of unique values in a column. A low-cardinality feature may have a handful of categories such as gender, payment method, or regional codes, whereas high cardinality examples include user identifiers or precise timestamps.

In machine learning and data analysis, low-cardinality features are generally easier to encode and train with. One-hot encoding is commonly used for such features, producing a matrix with one column per category. This approach works well when there are only a few categories, but it can become unwieldy as cardinality grows. Alternatives for higher-cardinality data include label encoding, target encoding, or hashing tricks, which help manage dimensionality and sparsity.

From a database perspective, low cardinality columns tend to be less selective, which can reduce index effectiveness for some queries but may enable efficient compression and specialized indexing approaches such as bitmap indexes or histograms used by query optimizers.

Practical handling often involves grouping rare categories into an “Other” bucket, creating derived features from the category, or ensuring consistent encoding across training and deployment to avoid data leakage. The characterization of “low” cardinality is relative and depends on the data size and the modeling or querying task.

See also: high cardinality, categorical encoding, feature engineering.