Meta & resources

Probability, statistics and data analysis

Machine Learning: concepts & procedures

Machine Learning: fundamental algorithms

Machine Learning: model assessment

Artificial neural networks

Natural language processing

The computer science appendix

The mathematics appendix

Powered By GitBook

(Some) mathematical measures

This page outlines, mainly for convenience, some useful measures/metrics utilised for several purposes.

The harmonic mean

It is the reciprocal of the mean of reciprocals:

$h = \frac{n}{\sum_{i=1}^{i=n} \frac{1}{x_i}}$

Common measures of distance/similarity

Euclidean

The euclidean distance of two vectors

$v_a = (x^i)_a$

and$v_b = (x^i)_b$

is the norm$l_2$

of the vector connecting them (it measures its length):$d = \sqrt{\sum_i (x^i_a - x^i_b)^2} = ||v_A - v_B||_2$

Hamming

The Hamming distance expresses the number of different elements in two lists/strings:

$A = 110101; B = 111001; d_{AB} = 2$

Jaccard (index)

Given two finite sets A and B, the Jaccard index gives a measure of how much they overlap, as

$J_{AB} = \frac{|A \cap B|}{|A \cup B|}$

Manhattan

Also called *cityblock*, the Manhattan distance between two points is the norm

$l_1$

of the shortest path a car would take between these two points in Manhattan (which has a grid layout):$d = \sum_i |u_i - v_i|$

Minkowski

The Minkowski distance is a generalisation of both the euclidean and the Manhattan to a generic p:

$d = \left(\sum_i |x_i - y_i|^p\right)^{1/p}$

Cosine

The cosine similarity is given by the cosine of the angle

$\theta$

spanned by the two vectors$d = \cos \theta = \frac{\bar u \cdot \bar v}{|\bar u| |\bar v|}$

So two perfectly overlapping vectors would have a cosine similarity of 1 and vectors at

$90^{\circ}$

would have a cosine similarity of 0.Chebyshev

It is also called *chessboard distance*. In the game of chess, the Chebyshev distance between the centers of the squares is the minimum number of moves a king needs to go from a square to another one.

$\max_i |u_i - v_i|$

See the figure here, it reports in red all the Chebyshev distance value from where the king (well, there's a drawing for it ...) sits to cell; note that the king can move horizontally, vertically and diagonally.

Last modified 6mo ago