SLT & ML

Friday, November 14, 2008

How to write and add your own algorithm to Weka

Recently，I am studying the weka .My purpose is that add my own algorithm to Weka.

I have collected some resource about these aspects from Weka website。 The links show as follow：

1. Frequently Asked Questions

In this webpage,it list many frequently asked questions by two catalogue：uing weka and developing weka。

2.How to write a new classifier to weka

Due to siginificant changes in the API beyond version 3.5.2, there are now two different HOWTOs available.

Please choose the version of Weka you want to develop for:

<= 3.5.2 (includes 3.4.x)
> 3.5.2

3.How do I add a new classifier, filter, kernel, etc.?

As of Weka 3.4.4, all the derived classes of superclasses that can be edited in the GenericObjectEditor, like subclasses ofweka.classifiers.Classifier for instance, can be determined dynamically at runtime. Read here for more information.
Note: Weka 3.5.8 turned the automatic discovery off by default.

The GenericObjectEditor is the core component in Weka for modifying schemes, like classifiers and filters in the GUI. It has to be configured correctly in order to show default and additional schemes. See the following articles for more details:

My Experience：

1.netbean+weka for developing english and chinese

2.if you want to use the weka3.5.8,due to Weka 3.5.8 turned the automatic discovery off by default.so you must make the adjust:

Automatic class discovery was turned off in 3.5.8. GUIs load faster

without it turned on and it doesn't work when Weka is loaded as a
plugin via a custom class loader. However, it is easily re-enabled. Just edit the
GenericPropertiesCreator.props file (located in weka/gui) and change
the "UseDynamic" property from "false" to "true".

3.you can be easy to add a new algorithm to weka.if you are famliar with chinese .the follow website is a good introduction:http://blog.chinaunix.net/u1/34560/showart_269880.html

4.when adding a new algorithm to weka.you may take attention to the problem,namely ,that your own algorithm can not display in the GUI(i.e Exprolor) even you have modify the "UseDynamic".For this problem.you should adjust your new package's position,For Instance:you want to add a new package"weka.classifier.cxk".you must configure the GenericPropertiesCreator.props again:

Original:

weka.classifiers.Classifier=\

weka.classifiers.bayes,\

weka.classifiers.functions,\

weka.classifiers.lazy,\

weka.classifiers.trees,\

weka.classifiers.meta,\

weka.classifiers.cxk,\

weka.classifiers.rules

Modify

weka.classifiers.Classifier=\

weka.classifiers.cxk,\

weka.classifiers.bayes,\

weka.classifiers.functions,\

weka.classifiers.lazy,\

weka.classifiers.trees,\

weka.classifiers.meta,\

weka.classifiers.rules

Monday, November 10, 2008

CRF 下载地址

CRF工具下载地址

CRF(Java):http://sourceforge.net/project/showfiles.php?group_id=105386

CRF++(C++):http://sourceforge.net/projects/crfpp

FlexCRF(C++):http://sourceforge.net/project/showfiles.php?group_id=184307

pocket CRF(C++): http://sourceforge.net/project/showfiles.php?group_id=201943

Wednesday, September 10, 2008

关于核矩阵的两种极端情况分析

核包括所有可能提供给学习机的关于输入在特征空间中的相对位置的信息。很自然的，如果要在数据集内发现几何结构，数据必须通过核矩阵展现这一结构。
如果这个核过于一般，对于具体类型的相似性没有给出足够的意义。用我们讨论先验知识的语言来说，这相当于为太多的分类赋予权重（可以看成是太多的坐标轴，即基太大）。所以，核把具有相等权值的任何一对输入看成相似（只有自己核自己相似）或者不相似（两两不同的向量是不相似的）。这样的核矩阵中除对角线以外的其他元素变得非常小，而同时对角线元素接近于1。因此核只能表示成单位矩阵的概念。这导致了过度拟合问题，因为我们把训练集正确分类很容易，但是核没有办法对新数据进行泛化处理。
另一种极端情况是，如果核矩阵是完全均一的，那么每一个输入都和每个其他的输入相似。这相当于每个输入都被映射到同一个特征向量，这导致数据欠拟，因为这是能够容易的表达出来的函数只有那些把所有的点都映射到同一个类的函数。
从几何上将，第一情形相当于把输入映射到特征空间中的标准正交点，而第二种情况，所有点都被合并到同一个映像。
这两种情况数据中都不存在非平凡的自然类，从而没有真正的结构可以泛化。