PARENTS = VL_AIB(PCX) runs Agglomerative Information Bottleneck (AIB) on the class-feature co-occurrence matrix PCX and returns a vector PARENTS representing the sequence of compressed AIB alphabets.
PCX is the joint probability of the occurrence of the class label C and the feature value X. PCX has one row for each class label and one column for each feature value, non negative entires and sums to one. AIB iteratively merges the pair of feature values that decreases the mutual information I(X,C) the least. This compresses the alphabet of the discrete random variable X in such a way that the new variable is still informative about C.
Merge operations are represented by a binary tree. The nodes of the tree correspond to the original feature values and any other value obtained by merging.
The vector PARENTS represents the merge tree. The nodes are numbered in breadth-first order, starting from the leaves. The numbers associated to the tree leaves correspond to the original feature values (so the first leaf has number one and correspond to the first feature value). In total there are 2*M-1 nodes, where M is the number of feature values (the number of columns of PCX). The internal nodes are numbered according to the order in which AIB generates them. It is therefore possible to recover from the tree the state of the AIB algorithm at each step (see also VL_AIBCUT()). PARENTS is a UINT32 array with one element for each tree node storing the index of the parent node. The root parent is conventionally set to 1.
Feature values with null probability (null columns of the PCX matrix) are ignored by the AIB algorithm and the corresponding entries in the PARENTS vectors are set to zero. Notice that this causes the root of the tree to have index smaller of 2*M-1 (PARENTS has still 2*M-1 entries, but the last portion is zero-padded).
Alternatively, the option ClusterNull can be used to assign the null probability values to a special value. The result is similar to pretending that the null probability nodes have indeed very small probability, uniform across categories.
[PARENTS, COST] = VL_AIB(...) returns the values COST of the cost function being optimized by AIB (i.e. the mutual information I(X,C)). COST has M column. The first column is the initial value of the cost function. The others correspond to the cost after each of the M-1 merges. If less than M-1 merges are performed, the rest of the vector is filled with NaNs.
VL_AIB() accepts the following options:
If specified, increase verbosity level.
If specified, do not signal null nodes; instead cluster them.