Plugging objects together: e.g perform cross validation on the following system: greedy backward feature selection on a fast base algorithm, training on those features with an SVM for each output in a one-against-the-rest multi-class system, choosing all hyperparameters with a model selection method.
他在工具箱中提供的接口都是模仿者对象的思想,任何方法和数据都看成对象,例如:用一个SVM对象去训练一个Data数据对象:
X=rand(50)-0.5; Y=sign(sum(X,2)); d=data(X,Y) % make simple data
[res alg]= train(svm,d) % train a support vector machine
and that's it - the svm is trained! To test it on more data:
[res]= test(alg,d)
这里面对象有多少呢?看下面的列表,包括算法模型和数据:
Basic library objects.
+
Storing input data and output results
+
Implementation of data object that limits memory overhead
+
Generic algorithm object
+
Groups sets of objects together (algorithms or data)
+
Evaluates loss functions
+
Takes mean loss over groups of algs
+
Builds chains of objects: output of one to input of another
+
To train and test different hyperparameters of an object
+
Cross validation using objects given data
+
Evaluates and caches kernel functions
+
Evaluates and caches distance functions
Statistical Tests objects.
-
Wilcoxon test of statistical significance of results
-
Corrected resampled t-test - for dependent trials
Dataset objects.
+
Spiral dataset generator.
+
Generator of dataset with only a few relevant features
+
Simple 2d Gaussian problem generator
+
Linear Regression with o outputs and n inputs
Pre-Processing objects
+
Simple normalization of data
+
General user specified mapping function of data
Density Estimation objects.
+
Parzen's windows kernel density estimator
-
Density estimator which assumes feature independence
+
Classifer based on density estimation for each class
+
Normal distribution density estimator
Pattern Recognition objects.
+
Support Vector Machine (svm)
-
C4.5 for binary or multi-class
+
k-nearest neighbours
-
Conditional Probability estimation for margin classifiers
-
Multi-Kernel LP-SVM
-
Minimize the a-norm in alpha space using kernels
-
Local and Global Consistent Learner
+
Bagging Classifier
+
ADABoost method
-
Hidden Markov Model
-
Leave One Out Machine
-
Minimize l1 norm of w for a linear separator
-
Kernel Dependency Estimation: general input/output machine
-
Kernel Perceptron
-
Ordinal Regression Perceptron (Shen et al.)
-
Splitting Perceptron (Shen et al.)
-
Sparse, online Pereceptron (Crammer et al.)
-
Random Forest Decision Trees WEKA-Required
-
J48 Decision Trees for binary WEKA-Required
Multi-Class and Multi-label objects.
+
Voting method of one against the rest (also for multi-label)
+
Voting method of one against one
-
Multi-class Support Vector Machine by J.Weston
-
C4.5 for binary or multi-class
+
k-nearest neighbours
Feature Selection objects.
+
Generic object for feature selection + classifier
-
SVM Bound-based feature selection
+
Recursive Feature Elimination (also for the non-linear case)
-
Dual zero-norm minimization (Weston, Elisseeff)
-
Primal zero-norm based feature selection (Mangasarian)
-
Fisher criterion feature selection
-
selection algorithm of Friedman (greedy selection)
-
Multi-class feature selection using spectral clustering
-
Mutual Information for feature selection.
Regression objects.
+
Support Vector Regression
+
Gaussian Process Regression
-
Relevance vector machine
+
(possibly multi-dimensional) ridge regression
-
Multivariate Regression via Stiefel Constraints
+
k-nearest neighbours
+
meta method for independent multiple output regression
-
kernel matching pursuit
-
kernel partial least squares
-
least mean squared regression [now obselete due to multi_rr]
-
Radial Basis Function Network (with moving centers)
-
Reduced Error Pruning Tree WEKA-Required
-
Structure Output Learning using Joint Kernel Method
Model Selection objects.
+
select parameters from a grid of values
-
Selecting SVM parameters by generalization bound
+
Bayessian parameter selection
Unsupervised objects.
+
One class SVM
+
K means clustering
+
Kernel Vector Quantization
+
Kernel Principal Components Analysis
-
Probabilistic Principal Component Analysis
-
Non-negative Matrix factorization
-
Spectral clustering
-
Manifold ranking
-
Probabilistic PCA
Reduced Set and Pre-Image objects.
-
Calculate Pre-Images based on multi-dimensional scaling
-
Calculate Pre-Images based on learning and ridge regression
-
Bottom Up Reduced Set; calculates reduced set based on gradient descent
-
Bottom Up Reduced Set; calculates reduced set for rbf with fixed-point iteration schemes
-
Top Down Reduced Set; calculates reduced set with multi-dimensional scaling
-
Top Down Reduced Set; calculates reduced set with ridge regression
-
Reduced Set Selection via L1 penalization
-
Reduced Set Selection via L0 penalization
+
Reduced Set Selection via matching pursuit