| |
Ordinal Regression Perceptron
|
a=ord_reg_perceptron(hyperParam)
Generates a Ordinal Regression Perceptron (SP) object with given hyperparameters.
The ordinal regression perceptron learns a discriminative
function f(x_i) = such that
d(rank(x_i),rank(x_j)) > eps -> |f(x_i)-f(x_j)| > |g(x_i,x_j) * tau|
where d is a distance controlling the level of insensivity and g is a
function controlling the learning margin between two examples with
different ranks (g = margin_control).
New functions for d and g can be added to the end of training.m
(see this file for examples).
As the number of candidates for each example might be large, the training
data is given as follows:
X stores m times the base name (including path) of the file containing
the candidates. The actual files have a number which conincides with
the label Y. E.g. if X(i,:) = '../data/candidate', the corresponding
file would be ['../data/candidate' num2str(Y(i)) '.mat']. The name of the
variable in the file holding the feature vectors of the
candidates as row vectors is assumed to be "X".
ATTENTION: You have to give a name to your datasets as the first
parameter is interpreted as a name due to spider specification. If you do
not name your dataset spider will interpret the filenames as name of the
dataset.
Hyperparameters (with defaults)
eps -- epsilon for epsilon insensitive rank distance (see also a.d)
tau=1 -- "margin" by which the candidates shall be
separated
loops=100 -- maximal number of iterations of training
margin_control = 'inverse_index_difference'
-- function to control the learning margin
default is g(p,q) = p^{-1} - q^{-1}
d = 'abs_rank_diff' -- distance between two ranks
Model
alpha -- the weights
b -- the offset
Methods:
train, test
Example:
Say you have 100 files with candidates in ../demos/data/toy_candidate with
core name "candidate"
X = repmat(['../demos/data/toy_candidate/candidate'],100,1);
d = data('training',X,[1:100]')
a.eps = 30;
[r a]=train(a,d);
| |
Reference : Discriminative Reranking for Machine Translation |
| |
Author : Libin Shen, Anoop Sarkar and Franz Josef Och |