X-Attributes Classifier (XAC): A New Multiclass Classification Method by Using Simple Linear Regression and Its Geometrical Properties

Similar documents
Kernel-based Methods and Support Vector Machines

An Introduction to. Support Vector Machine

Support vector machines

Bayes (Naïve or not) Classifiers: Generative Approach

Binary classification: Support Vector Machines

Introduction to local (nonparametric) density estimation. methods

Solving Constrained Flow-Shop Scheduling. Problems with Three Machines

Applications of Multiple Biological Signals

Simple Linear Regression

Bayesian Classification. CS690L Data Mining: Classification(2) Bayesian Theorem: Basics. Bayesian Theorem. Training dataset. Naïve Bayes Classifier

CS 1675 Introduction to Machine Learning Lecture 12 Support vector machines

Functions of Random Variables

Machine Learning. knowledge acquisition skill refinement. Relation between machine learning and data mining. P. Berka, /18

Unsupervised Learning and Other Neural Networks

Study of Correlation using Bayes Approach under bivariate Distributions

Estimation of Stress- Strength Reliability model using finite mixture of exponential distributions

Econometric Methods. Review of Estimation

6. Nonparametric techniques

Lecture 7. Confidence Intervals and Hypothesis Tests in the Simple CLR Model

Correlation and Regression Analysis

A New Family of Transformations for Lifetime Data

Dimensionality Reduction and Learning

Chapter 4 (Part 1): Non-Parametric Classification (Sections ) Pattern Classification 4.3) Announcements

Midterm Exam 1, section 1 (Solution) Thursday, February hour, 15 minutes

Ordinary Least Squares Regression. Simple Regression. Algebra and Assumptions.

Comparison of Dual to Ratio-Cum-Product Estimators of Population Mean

Application of Calibration Approach for Regression Coefficient Estimation under Two-stage Sampling Design

Outline. Point Pattern Analysis Part I. Revisit IRP/CSR

ABOUT ONE APPROACH TO APPROXIMATION OF CONTINUOUS FUNCTION BY THREE-LAYERED NEURAL NETWORK

Chapter 14 Logistic Regression Models

Midterm Exam 1, section 2 (Solution) Thursday, February hour, 15 minutes

( ) = ( ) ( ) Chapter 13 Asymptotic Theory and Stochastic Regressors. Stochastic regressors model

2.28 The Wall Street Journal is probably referring to the average number of cubes used per glass measured for some population that they have chosen.

MEASURES OF DISPERSION

Chapter 5. Curve fitting

Summary of the lecture in Biostatistics

Research on SVM Prediction Model Based on Chaos Theory

Lecture Notes 2. The ability to manipulate matrices is critical in economics.

Statistics MINITAB - Lab 5

A tighter lower bound on the circuit size of the hardest Boolean functions

Pinaki Mitra Dept. of CSE IIT Guwahati

Part 4b Asymptotic Results for MRR2 using PRESS. Recall that the PRESS statistic is a special type of cross validation procedure (see Allen (1971))

Support vector machines II

Multiple Choice Test. Chapter Adequacy of Models for Regression

9.1 Introduction to the probit and logit models

ESS Line Fitting

Comparing Different Estimators of three Parameters for Transmuted Weibull Distribution

Dimensionality reduction Feature selection

Chapter Statistics Background of Regression Analysis

Feature Selection: Part 2. 1 Greedy Algorithms (continued from the last lecture)

Generative classification models

{ }{ ( )} (, ) = ( ) ( ) ( ) Chapter 14 Exercises in Sampling Theory. Exercise 1 (Simple random sampling): Solution:

Prof. YoginderVerma. Prof. Pankaj Madan Dean- FMS Gurukul Kangri Vishwavidyalaya, Haridwar

Objectives of Multiple Regression

Chapter 13, Part A Analysis of Variance and Experimental Design. Introduction to Analysis of Variance. Introduction to Analysis of Variance

Non-uniform Turán-type problems

Analysis of Variance with Weibull Data

PGE 310: Formulation and Solution in Geosystems Engineering. Dr. Balhoff. Interpolation

Rademacher Complexity. Examples

CS 2750 Machine Learning. Lecture 8. Linear regression. CS 2750 Machine Learning. Linear regression. is a linear combination of input components x

On Signed Product Cordial Labeling

12.2 Estimating Model parameters Assumptions: ox and y are related according to the simple linear regression model

Lecture 1. (Part II) The number of ways of partitioning n distinct objects into k distinct groups containing n 1,

Probability and. Lecture 13: and Correlation

Supervised learning: Linear regression Logistic regression

The number of observed cases The number of parameters. ith case of the dichotomous dependent variable. the ith case of the jth parameter

Derivation of 3-Point Block Method Formula for Solving First Order Stiff Ordinary Differential Equations

The equation is sometimes presented in form Y = a + b x. This is reasonable, but it s not the notation we use.

Chapter Business Statistics: A First Course Fifth Edition. Learning Objectives. Correlation vs. Regression. In this chapter, you learn:

STA 108 Applied Linear Models: Regression Analysis Spring Solution for Homework #1

Chapter Two. An Introduction to Regression ( )

Assignment 5/MATH 247/Winter Due: Friday, February 19 in class (!) (answers will be posted right after class)

Generalization of the Dissimilarity Measure of Fuzzy Sets

Introduction to Computer Design. Standard Forms for Boolean Functions. Sums and Products. Standard Forms for Boolean Functions (cont ) CMPT-150

Mechanics of Materials CIVL 3322 / MECH 3322

Line Fitting and Regression

ENGI 4430 Numerical Integration Page 5-01

DATE: 21 September, 1999 TO: Jim Russell FROM: Peter Tkacik RE: Analysis of wide ply tube winding as compared to Konva Kore CC: Larry McMillan

Overview. Basic concepts of Bayesian learning. Most probable model given data Coin tosses Linear regression Logistic regression

Chapter 13 Student Lecture Notes 13-1

L5 Polynomial / Spline Curves

. The set of these sums. be a partition of [ ab, ]. Consider the sum f( x) f( x 1)

About a Fuzzy Distance between Two Fuzzy Partitions and Application in Attribute Reduction Problem

Lecture Notes Types of economic variables

Lecture 8: Linear Regression

Double Dominating Energy of Some Graphs

KLT Tracker. Alignment. 1. Detect Harris corners in the first frame. 2. For each Harris corner compute motion between consecutive frames

ADAPTIVE CLUSTER SAMPLING USING AUXILIARY VARIABLE

STATISTICAL PROPERTIES OF LEAST SQUARES ESTIMATORS. x, where. = y - ˆ " 1

STA 105-M BASIC STATISTICS (This is a multiple choice paper.)

Unimodality Tests for Global Optimization of Single Variable Functions Using Statistical Methods

A New Development on ANN in China Biomimetic Pattern Recognition and Multi Weight Vector Neurons

f f... f 1 n n (ii) Median : It is the value of the middle-most observation(s).

ENGI 3423 Simple Linear Regression Page 12-01

Machine Learning. Introduction to Regression. Le Song. CSE6740/CS7641/ISYE6740, Fall 2012

Fault Diagnosis Using Feature Vectors and Fuzzy Fault Pattern Rulebase

Median as a Weighted Arithmetic Mean of All Sample Observations

Mean is only appropriate for interval or ratio scales, not ordinal or nominal.

hp calculators HP 30S Statistics Averages and Standard Deviations Average and Standard Deviation Practice Finding Averages and Standard Deviations

MIMA Group. Chapter 4 Non-Parameter Estimation. School of Computer Science and Technology, Shandong University. Xin-Shun SDU

Transcription:

Proceedgs o the World Cogress o Egeerg ad Computer Scece 05 Vol II WCECS 05, October -3, 05, Sa Fracsco, USA X-Attrbutes Classer (XAC): A New Multclass Classcato Method by Usg Smple Lear Regresso ad Its Geometrcal Propertes Jeremas T. Lals, Member, IAENG Abstract I ths paper, a ew multclass classcato method has bee proposed. Durg the trag process, the smple lear regresso was used to d the lear relatoshp betwee the pared varables as well as ts cetrod o every class. The three pots: based o lear uctos, cetrods, ad put values, were used to d the class membershp o the preseted ew obect by usg the ormula calculatg the area o tragle. Four stadard ad publc datasets take rom UCI mache learg repostory were used to evaluate the perormace o the proposed algorthm usg 5-old crossvaldato. Emprcal results show the satsactory perormace o XAC algorthm o learly ad olearly separable classes wth small trag sze ad/or hgh dmeso. Ide Terms data mg, multclass classcato, smple lear regresso, geometrc propertes U I. INTRODUCTION coverg hdde useul kowledge wth large datasets s the ma goal o data mg. It helps people makg proactve ad kowledge drve decsos. Hece, varous data mg techques emerged deret research topcs lke sequetal rules, patter recogto, clusterg, regresso ad classcato. Amog these topcs, data classcato became oe o maor researches due to ts wde applcatos [][3], such as bomedcal modelg, bologcal modelg ad etc. Classcato s a supervsed learg method that reers to the task o aalyzg a set o data cotag observatos order to lear a model or ucto that ca be used detyg the ew observato to oe o the predeed classes. It has bee a actve research topc ot oly the mache learg area, but also statstcs []. Early work o classcato ocused o dg whch varables dscrmate betwee two or more classes, or also kow as dscrmat ucto aalyss (DFA). The uderlyg dea DFA s to use the predctor varables rom the trag set to costruct the dscrmat uctos, lke lear uctos, that wll determe the group membershp o the usee Mauscrpt receved May, 05; revsed July 0, 05. Jeremas T. Lals s a Assstat Proessor the College o Computer Studes ad the Drector o the Isttutoal Research ad Publcato Oce o La Salle Uversty, Ozamz Cty, Phlppes, e-mal: eremas.lals@gmal.com obect. Moder classcato approaches ocused o automatc geerato o rule (e.g. decso tree), the use o codtoal probabltes ( e.g. Naïve Bayes), calculatg the dstaces the eature space (e.g. K-earest eghbor), ad eve through lear ad olear regresso (e.g. support vector mache) creatg more leble models. I ths paper, the researcher presets a ew ad smple classcato method based o smple lear regresso dg the lear relatoshps betwee the obect s attrbutes ad to use ts geometrcal propertes, area o tragle, calculatg the dstace o the ew obect rom the predetermed classes. Ths study also shows the applcablty o smple lear regresso lear ad olear separable multclass classcato problems. Four stadard datasets rom the UCI mache learg repostory were used to measure ad evaluate the perormace o the proposed algorthm. II. RELATED WORK A. Smple Lear Regresso Lear regresso s the task o dg the best-ttg straght le, or also kow as regresso le, through the eature space []. The ma dea ths techque s to reveal the lear relatoshp or to derve a lear ucto that lks varable ad y, deoted as y m b () where y s the crtero varable, m s the slope, s the predctor varable, ad b as y-tercept o the tredle. I ths case, the value o varable y s predcted based oly o varable, thus, t s called as smple lear regresso. There are some other lear ad bary classcato methods that apply ths techque to classy learly separable classes, such as perceptro ad support vector maches (SVM). However, the best-ttg le these methods s used to separate the two classes, where t s called as hyperplae. B. Lear Classcato va Hyperplae Regresso ad classcato are both learg techques data mg that are used to create predctve models based o the preseted data. However, these methods produce deret values or output varables, ad thus, used ISBN: 978-988-4047--5 ISSN: 078-0958 (Prt); ISSN: 078-0966 (Ole) WCECS 05

Proceedgs o the World Cogress o Egeerg ad Computer Scece 05 Vol II WCECS 05, October -3, 05, Sa Fracsco, USA deretly. Sce regresso takes cotuous values as output, the t s used to estmate or predct a respose. O the other had, classcato takes class labels as output so t s used to d the class membershp o the obect. However, there are some classcato methods, lke perceptro ad SVM, that adopted the cocept o lear regresso to classy obects but a deret ad ar more comple maer. Perceptro Perceptro s oe o the earlest algorthms or lear classcato veted by Frak Roseblatt at the Corell Aeroautcal Laboratory 957 [4]. It s also cosdered as a smple model o euro that has a set o eteral put that ca be o ay umber, a teral put b, ad oe Boolea output value. The ma dea ths method s to d the sutable values or the weghts w the separatg hyperplae, (), so that the trag eamples are correctly classed. The hyperplae s geometrcally deed as A. Trag Phase I order or ay classer to dety the correct class membershp o the ew obect, t should be traed rst usg the trag set ad create a predctve model. Fg. shows the block dagram o -attrbutes classer (XAC) trag procedure. ) Gve the trag dataset wth umber o attrbutes ad k tuples, d the lear relatoshp betwee the pars o attrbutes each class, w b 0 ( ) () 0 otherwse However, the separatg hyperplae s oly guarateed to be oud the learg set s learly separable, otherwse, the trag process wll ever stop. Ths maor drawback makes ths algorthm less applcable to may patter recogto problems. Support Vector Mache (SVM) Lke perceptro, support vector mache (SVM) s a hyperplae based classer, but t s backed wth sold theoretcal groudg [5]. The obectve ths method s to d a optmal hyperplae, w. + b = 0, that separates the two classes wth the largest marg. It meas that ths hyperplae has the largest mmum dstace to the trag set. The hyperplae ca be ormally deed as T ( ) sg ( w b) (3) where w s the weght vector ad b as the bas whch ca be computed based o the trag data pot by solvg a costraed quadratc optmzato problem. The al decso ca the be derved ad deed as ( ) sg N y ( ) b (4) Where ths ucto depeds o a o-zero support vectors α whch are ote a small racto o the orgal dataset. III. XAC ALGORITHM The ma obectve o ths study s to use the lear ucto, () = m. + b, classyg lear ad olear separable multclass obects. I geeral, the proposed algorthm has two stages, the trag phase ad classcato phase. Fg.. Block dagram o the proposed trag procedure o -attrbutes classer.... ( ( ( ) ) ) (5) where ( ) s the lear ucto betwee attrbutes ad +, α s the slope, ad β s the oset. The slope α ( ) ca be computed as: k ( ) (6) k whle the oset β ( ) s computed as: (7) k The resultg values o α ad β betwee the pared attrbutes each class wll the be used as teral puts to calculate the output value durg the classcato stage. ) Calculate the cetrod C o the pared varables ad + or each class, deoted as C (, ): ISBN: 978-988-4047--5 ISSN: 078-0958 (Prt); ISSN: 078-0966 (Ole) WCECS 05

Proceedgs o the World Cogress o Egeerg ad Computer Scece 05 Vol II WCECS 05, October -3, 05, Sa Fracsco, USA, (8) Fgs., 3 ad 4 llustrate the scatter plot o each par o attrbutes as well as ts correspodg regresso le or each class the Irs lower dataset. classcato process o XAC algorthm. To determe the class membershp o the put obect: ) Fd the rst pot o the tragle or every pared attrbutes o ts respectve class by usg the Fg.. Scatter plot ad lear relatoshps betwee sepal legth ad sepal wdth o the three classes. Fg. 3. Scatter plot ad lear relatoshps betwee sepal wdth ad petal legth 3 o the three classes. Fg. 4. Scatter plot ad lear relatoshps betwee petal legth 3 ad petal wdth 4 o the three classes. B. Classcato Phase Ater the trag process, the resultg model ca ow be used to classy the ew obect. Fg. 5 shows the Fg. 5. Block dagram o the classcato process o XAC. ISBN: 978-988-4047--5 ISSN: 078-0958 (Prt); ISSN: 078-0966 (Ole) WCECS 05

Proceedgs o the World Cogress o Egeerg ad Computer Scece 05 Vol II WCECS 05, October -3, 05, Sa Fracsco, USA prevously calculated lear uctos ( ), ( ), ( 3 ),, ( - ) ad ts correspodg put values,, 3,, -. The resultg y-coordates would be o the orm o (eteral put, teral put ( )). ) The prevously computed cetrod C o each pared attrbute class wll serve as the secod pot o the correspodg tragles, the orm o (teral put, teral put ). 3) Par the put values, e.g. (eteral put, eteral put ), to obta the thrd pot o the correspodg tragles class. 4) Use the three pots o each pared attrbutes class to calculate the area o ts correspodg tragles, Area ( ) ( ) 5) Calculate the dstace o the put obect rom the eature vectors every class by summg up all the correspodg Area o ts pared attrbutes, (9) dst Area (0) where s the umber o attrbutes. 6) The class that obtaed the least dstace wll be declared as the wer or the class membershp o the ew obect. IV. EXPERIMENTS A. Dataset To measure ad valdate the perormace o the proposed algorthm, our publc datasets rom UCI Mache Learg Repostory were cosdered: Irs Flower [6], Wheat Seed Kerel [7], Breast Tssue [8], Breast Cacer Wscos (Dagostc) [9], ad Oe Hudred Plat Speces Leaves [0]. Table I shows the characterstcs o each dataset used the epermets. B. Evaluato To evaluate the perormace o the proposed method, 5- old cross-valdato was used each epermet. The trag ad testg steps were perormed ve tmes by TABLE I DATASET CHARACTERISTICS Dataset Trag Sze Testg Sze # o Classes Dm Irs Flower 0 per class 40 per class 3 4 Wheat Seed 4 per class 56 per class 3 7 Breast Tssue 4 or class 7 or class 4 4 0 or class 3 or class 3 4 or class 4 39 or class or class 3 8 or class 4 Breast Cacer 7 or class 86 or class 30 4 or class 70 or class Leaves-Shape 3 per class 3 per class 5 64 parttog the dataset to ve mutually eclusve subsets or olds. Accuracy, precso, recall ad F score were also used to measure the correctess, eactess, completeess, ad retreval perormace, respectvely, o the model beg produced by XAC every epermet. V. RESULTS AND DISCUSSION The summary o epermets results usg the our datasets s reported Table II. As we ca see, the XAC algorthm perorms best wth the Irs lower dataset compared wth the other three datasets. The result proves the applcablty o smple lear regresso classyg ot oly learly separable, but Dataset TABLE II EXPERIMENTS RESULTS SUMMARY Mea Accuracy Mea Precso cludg olearly separable classes. Net to t are the results o the epermets coducted wth the breast cacer dataset havg a mea precso o 89.9. Note that the dvso o the dataset, trag ad testg, s slghtly mbalaced, where 6 s comg rom the beg class ad the rest s rom the malgat class. However, results rom the epermets usg the wheat seed dataset are ar more better terms o mea accuracy, mea recall, ad mea F-score compared to the results wth the breast cacer dataset. It s also otable that the algorthm was able to produce a acceptable result or leaves dataset terms o precso at 86.47 despte o the lmted umber o trag set, three per class, ad hgh dmesoalty. Addg to the dculty o the classcato problem ths dataset s that may o the sub speces resemble close appearace wth the other maor speces, ad may sub speces resemble a radcally deret appearace wth ts maor spece []. Furthermore, results also show the robustess o the approach by usg oly the shape-based dataset durg trag ad testg. However, results gve by XAC usg the breast tssue dataset gve the lowest result, especally terms o completeess at 67.88. Ths s due to the mbalace o the umber o trag ad testg sets each class, where, 48 o the total umber o t s comg rom oe class oly. I geeral, the proposed algorthm perorms satsactorly eve wth small umber o trag set at 0 o the total sze o each dataset. VI. CONCLUSION Mea Recall Mea F-Score Irs Flower 94.50 95.05 94.50 94.77 Wheat Seed 89.7 89.7 89.3 89.48 Breast Tssue 75.9 75.3 67.88 7.3 Breast Cacer 88.55 89.9 85.9 87.87 Leaves (Shape) 83.69 86.47 83.63 85.03 Ths paper has preseted a ew method that ca be used or multclass classcato problems wth learly ad olearly separable classes usg smple lear regresso whch s orgally desged or bary classcato ISBN: 978-988-4047--5 ISSN: 078-0958 (Prt); ISSN: 078-0966 (Ole) WCECS 05

Proceedgs o the World Cogress o Egeerg ad Computer Scece 05 Vol II WCECS 05, October -3, 05, Sa Fracsco, USA problem wth learly separable classes oly. Emprcal results rom the epermets coducted usg the our stadard ad publc datasets take rom UCI mache learg repostory showed the satsactory perormace o the proposed algorthm. For the uture work, several aveues or mprovemet ca stll be cosdered lke usg the olear regresso to cater those pared attrbutes wth olear relatoshp. REFERENCES [] V. S. M. Tseg ad C. Lee, Cbs: A ew classcato method by usg sequetal patters, Proc.005 SIAM Iteratoal Data Mg Coerece, CA, 005, pp.596-600. [] A. A, Classcato methods. CA: Idea Group Ic, 005, pp. -6. [3] A. Arakeya, L. Nersya, A. Gevorgya, ad A. Boyaya, Geometrc approach or Gaussa-kerel bolstered error estmato or lear classcato computatoal bology, Iteratoal Joural Iormato Theores ad Applcatos, vol., o., pp. 70-8-, 04. [4] F. Roseblatt, The perceptro-a percevg ad recogzg automato, Corell Aeroautcal Laboratory, New York, Report 85-460-, 957. [5] C. Cortes ad V. Vapk, Support-vector etworks, Mache Learg, vol. 0, o. 3, pp. 73-97, 995. [6] UCI Mache Learg Repostory, Irs data set, 988. [Ole]. Avalable: archve.cs.uc.edu/ml/datasets/irs [7] UCI Mache Learg Repostory, Seeds data set, 0. [Ole]. Avalable: archve.cs.uc.edu/ml/datasets/seeds [8] UCI Mache Learg Repostory, Breast tssue data set, 00. [Ole]. Avalable: archve.cs.uc.edu/ml/datasets/breast+tssue [9] UCI Mache Learg Repostory, Breast cacer wscos (dagostc), 995. [Ole]. Avalable: archve.cs.uc.edu/ml/datasets/breast+cacer+wscos+(dagost c) [0] UCI Mache Learg Repostory, Oe-hudred plat speces leaves data set, 0. [Ole]. Avalable: archve.cs.uc.edu/ml/datasets/oehudred+plat+speces+leaves+data+set [] C. Mallah, Probablstc Classcato rom a K-Nearest-Neghbor Classer, Computatoal Research, vol., o., pp. -9, 03. [] Wkpeda, Lear regresso, 05. [Ole]. Avalable: http://e.wkpeda.org/wk/lear_regresso ISBN: 978-988-4047--5 ISSN: 078-0958 (Prt); ISSN: 078-0966 (Ole) WCECS 05