a = mrank(hyperParam)
Generates an mrank object with given hyperparameters.
Hyperparameters (with defaults)
child=kernel('rbf') -- the kernel is stored as a member called "child"
alpha=0.99 -- parameter governing convergence rate
iterations=600 -- number of iterations of algorithm
use_edge_graph=0 -- if turned on, sparsify kernel using edge graph
normalize=0 -- normalize in the input space? 1 yes, 0 no
standardize=[] -- centre and standardize data? 1 yes, 0 no
default behaviour is to do so, except for
linear and polynomial-based kernels
(good for maintaining sparsity).
Methods:
TRAIN, TEST, PLOT
There is no real need to split the algorithm into TRAIN and TEST
phases - this is done for consistency with other algorithm classes.
If labels are supplied in DAT.y, a single call to TRAIN is sufficient.
If DAT.y is empty, TRAIN merely stores the data in the algorithm
object, waiting for query points to be supplied by TEST.
[DAT, A] = TRAIN(A, DAT);
Points with Y=0 within DAT are considered unlabelled, and points with
non-zero Y are the query points. The unlabelled points are ranked
according to their similarity to the query points, computed via an
iterative spreading activation network algorithm. A.result on exit
contains [I S R] where I is the original index, S the ranking score
and R the rank of each point. Results can be plotted with PLOT(A).
DAT is unchanged.
DAT = TEST(A, DAT)
If DAT contains labels, throw away data stored in A and procede using
DAT as for TRAIN. If not, use DAT as the query points, and use any
unlabelled points from the data stored in A as the unlabelled points
(stored labelled points, i.e. those that were used as query points in
the previous run, are removed). On exit, the y field of DAT contains
ranking scores. The results can be plotted with PLOTRANKING(DAT).
PLOT(A) -- see HELP MRANK/PLOT
d = gen(spiral({'m=100','n=0.5','noise=0.35'}));
labelled = [50];
d.Y = double(ismember([1:length(d.Y)]', labelled));
a = mrank;
a.child = kernel('rbf',0.1);
a.use_edge_graph = 1;
[dd aa] = train(a, d);
plot(aa, 'etrc')
(plots [e]dge graph, minimal spanning [t]ree, [r]anking and [c]onvergence)
| |
Reference : Ranking on Data Manifolds |
| |
Author : Dengyong Zhou , Jason weston , Arthur Gretton , OlivierBousquet and Bernhard Schölkopf |