Propagation of affinity

In data-processing Programming, the propagation of affinity is a recent algorithm of Partitionnement of data, or clustering , which makes it possible to find the elements of a unit which are most representative - a criterion of resemblance being given - of the unit.

Description of the algorithm

It is an iterative algorithm which rests on the division of “affinities”:

Each element C reference mark in its vicinity an element which resembles to him sufficiently, and increases its affinity for this element;
the following stages consist in “propagating” this affinity:

Each element C the reference mark that for which it with largest affinity, noted m ;
It adds to its own affinities those of m ;
This stage is repeated a certain number of times, or until the number of elements passes in lower part of a certain threshold - or when this stage does not bring any more any change.

There are then three cases:

the element considered has a maximum affinity for another element: it resembles to him;
the element considered has a maximum affinity for itself: it is “exemplary” ( exemplar );
the element considered has a null affinity: it “is insulated”.

The number of exemplary elements depends on many parameters and a priori cannot be given.

One obtains at the conclusion of the algorithm a complete Arbre, connecting the similar elements which could be identified like such.

References

Random links: