normalizer

A normalizer is a component or process that converts data or signals into a standard form, enabling consistent processing, comparison, or interpretation. The term is used across disciplines such as statistics, machine learning, linguistics, and signal processing.

In statistics and machine learning, normalization refers to rescaling numeric features to a common range or distribution. Common methods include min-max scaling, which maps features to a desired interval (often 0 to 1); z-score standardization, which centers data at zero with unit variance; and more robust or max-abs variants. The goal is to reduce bias due to differing feature scales and to improve the performance and convergence of learning algorithms. Normalizers may be applied per feature or per sample, depending on the task.

In vector spaces, normalization typically means converting a vector to a unit vector. This is done by dividing the vector by its norm, most often the L2 norm. Normalized vectors are useful in similarity measures such as cosine similarity and in stabilizing computations that depend on direction rather than magnitude. Other norms, like the L1 or Linf norms, yield different unit-like representations.

In text and data representation, normalization transforms content into canonical forms. Examples include case folding (converting to lowercase), Unicode normalization (such as NFC or NFD), removing diacritics, and stripping or standardizing punctuation. This improves matching, indexing, and cross-system interoperability.

In signal processing and audio, normalization adjusts amplitude to a reference level to maintain consistent loudness or avoid clipping. In databases and ETL pipelines, data normalization involves converting disparate data into standardized formats and canonical representations to ensure data integrity and comparability.