Parallel muiticategory Support Vector Machines (PMC-SVM) for Classifying Microarray Data 研究生 : 許景復 單位 : 光電與通訊研究所

Outline  Introduction  SMO-SVM  Parallel Muiticategory SVM  Parallel Implementation and Environment  Parallel Evaluation and Analysis  Classifying Microarray Data  Conclusions

Introduction Biologists want to separate the data into multiple categories using a reliable cancer diagnostic model. Based on a comprehensive evaluation of several muiticategory classification methods, it is found that support vector machines (SVM) are the most effective classifiers for performing accurate cancer diagnosis form gene expression. In the paper, we developed new parallel muiticategory support vector machines (PMC-SVM) based on the sequential minimum optimization-type decomposition methods for support vector machines (SMO-SVM) of LibSVM term that needs less memory.

SMO-SVM The basic idea behind SVM is to separate two point classes of a training set, by using a decision function optimization by solving a convex quadratic programming optimization problem of the form Subject to (1)

SMO-SVM entriesare defined as wheredenotes a kernel function, such as polynomial kernel or Gaussian kernel. where is a constant. and is a vector of all ones.is the symmetric positive semidefinite matrix. (3)

SMO-SVM The subset, denoted as B, is called working set. If B is restricted to have only two elements, this special type of decomposition method is the Sequential Minimal Optimization (SMO).

Step2:IfIs a stationary point of (2), stop. Otherwise, find a two-element working set Define, andand as subvector ofcorresponding toand,respectively. There are four steps to implement SMO: Findas the initial feasible solution. SetStep1:

Step3: If Solve the following sub-problem with the variable : subject to else solve subject to constraints of (4) Step4:Setto be the optimal solution of (4) and and go to step 2.. Set (4) (5)

Parallel Muiticategory SVM(PMC-SVM)  In muiticategory classification of support vector machines, the algorithm will generate sub models for categories.  Generating models is the most time consuming task in this algorithm so it is desirable to distribute all the sub models onto multiple processors and each processor perform a subtask to improve the performance.

Example : We have 4 processors and k=16, that means we have to generate k(k-1)/2 models, which are total 120 models. whereis the total number of the processors andthe number of categories.

Parallel Implementation and Environment One is the sharedmemory SGI Origin 2800 Supercomputers(sweetgum) equipped with 128 CPUs, 64 gigabytes of memory, and 1.6 Terabytes of fiberchannel disk. The other is a distributed memory Linux cluster (mimosa) with 192 nodes.

Parallel Evaluation and Analysis PMC-SVM is tested on both sweetgum and mimosa platforms using the above two datasets. Dataset 1: Letter_scale classes: 26 trainig size: 16,000 features: 16 Dataset 2: Mnist_scale classes: 10 training size: 21,000 features: 780

Figure 2. The speedup of PMC- SVM on sweetgum with Dataset 1 (Letter_scale ) Figure 3. The speedup of PMC- SVM on mimosa with Datasets 1 (Leetter_scale)

Figure 4. The speedup of PMC-SVM on swetgum with Datasets 2 (Mnist_problem) Figure 4. The speedup of PMC-SVM on mimosa with Datasets 2 (Mnist_problem)

Classifying Microarray Data Dataset 3: 14_Tumors(40Mb) Human tumor types: 14 normal tissue types: 12 Dataset 4: 11_Tumors(18Mb) Human tumor types: 11 In the work, two microarray datasets were to demonstrate the performance of PMC-SVM, as listed below:

#of PEsTime (s)Speedup #of PEsTime (s)Speedup Table 6: Performance on sweetgum (Dataset 3) Table 7: Performance on sweetgum (Dataset 4)

Conclusions PMC-SVM has been developed for classifying large datasets based on SMO-type decomposition method. The experimental results show that the high performance computing techniques and parallel implementation can achieve a significant speedup.

Thanks for your attendance!