Enclosing Machine Learning Paradigm

libsvm-plus matlab interface 2.88 is out

2009-02-28T20:27:00.000-08:00

After seriously working on the libsvm-plus code, now I have made a matlab interface for it. And Now several new features are supported:

(1) One-Vs-One MC-SVC
(2) One-Vs-Rest MC-SVC
(3) DDAG MC-SVC

Plz keep an eye on my blog.

libsvm plus MATLAB interface

2008-07-29T20:32:00.000-07:00

Finally, I have compiled an interface for libsvm plus. I should thank Nemo for providing us a free version of libsvm plus with neat coding. Also, I should thank Hsuan-Tien and Nemo for the help in the initialization methods of my QP solver. A new feature like CCMEB, is now inluded as a solver for the general center-constrtained minimum enclosing ball problem. For an experimental version, you could send your request to my email: xunkaidotweiatgmaildotcom

New QP solver will be included!

2008-06-06T06:59:00.000-07:00

Ha, after several weeks' work, I have compiled a new QP solver for general QP problem, the new solver is based on LIBSVM and Coreset MEB. I use LIBSVM-SMO as a core solver, and use coreset for further speedup. Note it only solves QP with bounded inequality and equality constraints.

Enclosing Machine Learning for PR

2008-05-24T07:53:00.000-07:00

Since this term, I'm working towards improving EML learning performance. Till now, Everything goes well. I think I can solve two difficult problems existed in SVM learning. One is multiple class SVM, the other one is multivariate SVR. Both are chanlleging problems. Actually, I got ideas for these two problems in 2005. But now they seems more important than before. Since I need more excellent journal papers for completing my two project. So today might be the new start point.

Work Hard!

Enclosing Machine Learning Toolbox In preparation

2008-04-11T01:10:00.000-07:00

After almost one year's work, I have just completed the initial Enclosing machine learning toolbox. The whole toolbox is written in MATLAB with patially C-code for speedup. Now the supported features are:

(1) L2SVC

(2) L2OCSVM

(3) RKHSMEB

(4) L2SVR

(5) Pending features include: Ranking, Manifold Learning, etc.

and Currently the supported solvers are:

(1) CPLEX

(2) MATLAB Solver

(3) GMNP Solver by V. F

(4) Pending Solvers include Libsvm etc.

I think I will give away all the source codes when I'm graduated in June this year!

So if you interest in it, just leave me a message and I will e-mail you when available!

RKHS for vector value function learning

2007-10-10T05:32:00.000-07:00

This idea is quite straightforward for generalization. Just need to redefine according definition and derive related theorems. Yet this work is quite hard.

New idea

2007-10-09T22:50:00.000-07:00

Ha,

Recently, we have read a new paper called ellipsoidal kernel machine, this paper gives us many hints about Mahalanobis SVM. We can have similar ideas towards improving the VC bound analysis of SVM. I think this point is important. Yet, the remain is also difficult. Because We have to derive the theoretical proof of it.
One Possible way is that firstly by defining a Mahalanobis transform, we also get a new space, we only need to prove that this new space is also a RKHS, then similar conclusions may be applied.

Anybody interested may give some suggestions.

Problems

2007-03-09T05:00:00.000-08:00

Several problems occur both in my life and my project. Maybe this is a wrong end with a beautiful start. I don't know whether I could suvive this tortue reality. But I have to face all the difficulties. Because I have to go my way, a way without stop. Nobody can block my progress. Nobody can wipe away the fire from my heart. Even though I think the god is fair. I can handle with all the difficulties.

Mahalanobis SVM in Progress Now

2007-01-31T03:26:00.000-08:00

As we mentioned before, We have been planning this new project several months ago. Now this project seems urgent. The idea is quite straightforward. Obviously, this new idea is more data dependent. This is because of its tight relyon the covariance matrix. As this is reported by several scholars. The related papers include minimax probability machine and Mahalanobis one-class support vector machines.

Our work is to propose a more general Mahalanobis distance based SVM.

I think we will spend about half of a year to complete this tight work. Also learning bounds for this works should be given.

Two conference papers are Now ready for Publication

2007-01-30T03:32:00.000-08:00

This year is really a fruitful year for me. A new project was approved by the 863 National High Technology Program. And Aha I am the head of the project. It is an exploring type project. Recently, My two papers were accepted by ISNN2007, all are accepted as oral presentation. Moreover, A paper was selected into a high level international journal Neurocomputing. Until toady, I did manage to revise both papers.

There are many peoples who help me during writing the two papers. First of all, Dr Guang-bin Huang really help me a lot both in study and life. He is like a brother. He is patient with me. He continuously encourage me. And he brings with me invaluable chances. Several days ago, He invite me as a reviwer for Neurocomputing. This is really an invaluable chance for me. Here, I cordially thank you for your big help.

Next is Dr Johan Loefberg. Yes, he is really a genius in both Optimization and Matlab programming. His great Matlab toolbox Yalmip is really a big works. You should never miss it.

The last is my girl friend Yue Feng. For her support and tolerance, I can have enough time to prepare my papers. Yes baby, I love you very much.

Although the work comes to an end, another new work comes along. I will prepare another new paper about Mahalanobis SVM. And this work is cooperated with Dr Huang.

Concept Innovation Vs Theory Innovation ?

2007-01-26T18:50:00.000-08:00

These two may be confusing. But who can tell me which is more important ?

How about one class classification under bayes framework ?

2006-12-10T08:03:00.000-08:00

This problem might be quite difficult for me. As this is not my major. But this is quite useful, Since one class classification stemmed from density estimation.

Till now, I'm still puzzled about several questions. Dr D.J.Tax's works might give me much hint.
Dr Debie and Professor Alex's MVEE in RKHS also should be a good reference.

Some Idea about Multiple Class Classification via One Class Classification

2006-11-20T12:42:00.000-08:00

Classification based on One class cognitive cognizing is always our leading idea. Indeed there are some related works about this subject. Such as training each class a cognizer and then detect whether a new sample is inside its boundary, Set covering machine might also be a good choice.

We also plan to use convex hull ensembles to learning the one class.

We still face with many obstacles yet.

Mahalanobis Support Vector Machines (In Plan)

2006-10-22T20:40:00.000-07:00

Actually mahalanobis distance is very useful, and illustrates superior performances in classifier design and pattern recognition applications. Also, Traditional SVMs are based on Euclid distance, If we use Mahalanobis distance then which conclusions can we make ? This is the main idea.

Next, actually is the implementation. How to derive this solution and which kind of optimization problem it can be simplified as ? How to make use of the covariance matrix ? How to estimate it ?
The minimax probability machine might provide a method for kernelizing the Mahalanobis SVMs.

Again the problem that might be the most difficult is that how to learn the kernel matrix. Generally there are two methods, the one is to learn the matrix directly, which is actually some kind of KPCA related. The other is to use an appoximation method, such as affine or convex combinations of some base kernels.

We may apply it to all the forms of SVMs.

Enclosing Machine Learning; Concepts and Algorithms

2006-09-10T10:43:00.000-07:00

How to implement two class classification with an ellipsoid ?

2006-09-05T17:09:00.000-07:00

With the help of Prof. Alexander Dolia, I did succeed to obtain this wonderful solution. This solution is actully a Minimum Enclosing and Maximum Excluding algorithm. It did not require that both class samples are balanced. It suits well for imbalanced samples case. To see the detail, you can just have a look at following conclusions.

Enclosing Machine Learning Paradigm

2006-01-13T22:17:00.000-08:00

We are pleased to announce that Our Project are now supported by China Nature Science Foundation. This new project is something new about machine learning concepts. We mainly focus on the cognitive class cognizing and recognizing. And we will mainly develop fast implementation techiniques for supporting the project.

Today is a new day, Yet we still need time to clarify each important concept we made.

About My Blog name!

2005-08-08T15:27:00.000-07:00

My blog is now active. I'm always puzzled about its name. You may think that the name is only a name without meaning. But for me, I think it important. It reflects one's mind more further one's soul.

Why did i call myself a uique scaler. This might be due to too much reasons. Yes, I'm unique, Although everybody is unique indeed. I indeed love research and communicating with others. But I found there is nobody that i can turn to for help. Sometime I feel lonely, and sad.

But I could not withdraw, I must march forward and try to approach my dream step by step.

I insist that one day my name will appear in the whole world !