Minmaksnormalisering

Minmaksnormalisering (min-max normalization) is a data preprocessing technique used to rescale numerical features to a fixed range, typically [0, 1]. It preserves the relative ordering of values and the shape of the distribution while constraining the scale of each feature, which can improve the performance of many machine learning algorithms.

The standard formula for a value x in a feature with minimum xmin and maximum xmax is x' = (x − xmin) / (xmax − xmin). If xmin equals xmax, the feature has no variation and x' is commonly set to 0 (or another constant such as 0.5, depending on convention). In practice, extrema are computed from the training data and the same transformation is applied to all subsequent data, including test data.

Applications of minmaksnormalisering include algorithms that rely on distance measurements or gradient-based optimization, such as k-nearest neighbors, support vector machines, neural networks, and many regression models. It is also used to standardize features before principal component analysis or other dimensionality reduction techniques.

Advantages of minmaksnormalisering include its simplicity, the preservation of relationships among values, and the production of uniformly scaled features which can improve convergence in some learning algorithms. Disadvantages include sensitivity to outliers, which can skew the min and max and compress the majority of the data into a small interval. New data outside the original min–max range can lead to values outside the target interval, unless clipping is applied.

Variants and considerations: scaling to [-1, 1] can be achieved with a small modification to the formula. For datasets with significant outliers, robust scaling or standardization (z-score) may be preferred. In practice, ensure consistent scaling across training, validation, and test sets and address constant features that yield zero division.