Scaling:
Some machine learning algorithms are very sensitive to distance and tuned to work well in specific ranges e.g. [-1, 1] range. For example, scaling to this range for gradient descent optimizer makes the loss surface smooth. Therefore, models tend to converge faster, and the [-1, 1] range offers the highest floating point precision.
Some of Scaling techniques: Min-max scaling: The minimum value that the input can take is scaled to –1 and the maximum possible value to 1: x_scaled = (2*x - max_x - min_x)(max_x - min_x). Clipping:The numeric value is linearly scaled between these two reasonable bounds, and then clipped to lie in the range [–1, 1]. Z-score normalization: scaling the input using the mean and standard deviation estimated over the training dataset: x_scaled = (x - mean_x)/(stddev_x)
This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.