Classification Part 4. Model Evaluation

Similar documents
Evaluation & Credibility Issues

A general framework for estimating similarity of datasets and decision trees: exploring semantic similarity of decision trees

Solution for Assignment 1 : Intro to Probability and Statistics, PAC learning

Driving Cycle Construction of City Road for Hybrid Bus Based on Markov Process Deng Pan1, a, Fengchun Sun1,b*, Hongwen He1, c, Jiankun Peng1, d

Hidden Markov Models

Performance Evaluation

Riemann is the Mann! (But Lebesgue may besgue to differ.)

Data Mining Models and Evaluation Techniques

Stein-Rule Estimation and Generalized Shrinkage Methods for Forecasting Using Many Predictors

Section 11.5 Estimation of difference of two proportions

How do we solve these things, especially when they get complicated? How do we know when a system has a solution, and when is it unique?

5.1 How do we Measure Distance Traveled given Velocity? Student Notes

7.2 Riemann Integrable Functions

Behavior-based Authentication Systems. Multimedia Security

Section 6: Area, Volume, and Average Value

8 Laplace s Method and Local Limit Theorems

1 Online Learning and Regret Minimization

Hybrid Group Acceptance Sampling Plan Based on Size Biased Lomax Model

= x x 2 = 25 2

Lecture INF4350 October 12008

Jack Simons, Henry Eyring Scientist and Professor Chemistry Department University of Utah

Probabilistic Investigation of Sensitivities of Advanced Test- Analysis Model Correlation Methods

Best Approximation in the 2-norm

Discrete Mathematics and Probability Theory Spring 2013 Anant Sahai Lecture 17

u( t) + K 2 ( ) = 1 t > 0 Analyzing Damped Oscillations Problem (Meador, example 2-18, pp 44-48): Determine the equation of the following graph.

Math 426: Probability Final Exam Practice

Fast Frequent Free Tree Mining in Graph Databases

Normal Distribution. Lecture 6: More Binomial Distribution. Properties of the Unit Normal Distribution. Unit Normal Distribution

4.5 JACOBI ITERATION FOR FINDING EIGENVALUES OF A REAL SYMMETRIC MATRIX. be a real symmetric matrix. ; (where we choose θ π for.

Designing Information Devices and Systems I Discussion 8B

Lecture 1: Introduction to integration theory and bounded variation

5: The Definite Integral

A signalling model of school grades: centralized versus decentralized examinations

Math& 152 Section Integration by Parts

Population bottleneck : dramatic reduction of population size followed by rapid expansion,

Quantum Physics II (8.05) Fall 2013 Assignment 2

Motion. Acceleration. Part 2: Constant Acceleration. October Lab Phyiscs. Ms. Levine 1. Acceleration. Acceleration. Units for Acceleration.

9.1 Day 1 Warm Up. Solve the equation = x x 2 = March 1, 2017 Geometry 9.1 The Pythagorean Theorem 1

Jin-Fu Li. Department of Electrical Engineering National Central University Jhongli, Taiwan

Monte Carlo method in solving numerical integration and differential equation

( dg. ) 2 dt. + dt. dt j + dh. + dt. r(t) dt. Comparing this equation with the one listed above for the length of see that

A Brief Review on Akkar, Sandikkaya and Bommer (ASB13) GMPE

#6A&B Magnetic Field Mapping

Lecture 14: Quadrature

UvA-DARE (Digital Academic Repository) A robust Xbar control chart Schoonhoven, M.; Does, R.J.M.M.

How do we solve these things, especially when they get complicated? How do we know when a system has a solution, and when is it unique?

Physics 201 Lab 3: Measurement of Earth s local gravitational field I Data Acquisition and Preliminary Analysis Dr. Timothy C. Black Summer I, 2018

A Signal-Level Fusion Model for Image-Based Change Detection in DARPA's Dynamic Database System

Sufficient condition on noise correlations for scalable quantum computing

DATA Search I 魏忠钰. 复旦大学大数据学院 School of Data Science, Fudan University. March 7 th, 2018

Advanced Calculus: MATH 410 Notes on Integrals and Integrability Professor David Levermore 17 October 2004

Online Short Term Load Forecasting by Fuzzy ARTMAP Neural Network

Reinforcement Learning

SEMIPARAMETRIC INFERENTIAL PROCEDURES FOR COMPARING MULTIVARIATE ROC CURVES WITH INTERACTION TERMS

Discrete Mathematics and Probability Theory Summer 2014 James Cook Note 17

p-adic Egyptian Fractions

Operations with Polynomials

Bayesian Networks: Approximate Inference

Non-Linear & Logistic Regression

Chapter 3 Solving Nonlinear Equations

Goals: Determine how to calculate the area described by a function. Define the definite integral. Explore the relationship between the definite

Detection and Estimation Theory

Math 113 Fall Final Exam Review. 2. Applications of Integration Chapter 6 including sections and section 6.8

Math 61CM - Solutions to homework 9

Chapter 9: Inferences based on Two samples: Confidence intervals and tests of hypotheses

Recitation 3: More Applications of the Derivative

Multiscale Fourier Descriptor for Shape Classification

Fig. 1. Open-Loop and Closed-Loop Systems with Plant Variations

LUMS School of Science and Engineering

1 Error Analysis of Simple Rules for Numerical Integration

Geometric Sequences. Geometric Sequence a sequence whose consecutive terms have a common ratio.

Predict Global Earth Temperature using Linier Regression

Synoptic Meteorology I: Finite Differences September Partial Derivatives (or, Why Do We Care About Finite Differences?

Math 520 Final Exam Topic Outline Sections 1 3 (Xiao/Dumas/Liaw) Spring 2008

The graphs of Rational Functions

different methods (left endpoint, right endpoint, midpoint, trapezoid, Simpson s).

{ } = E! & $ " k r t +k +1

ADVANCEMENT OF THE CLOSELY COUPLED PROBES POTENTIAL DROP TECHNIQUE FOR NDE OF SURFACE CRACKS

We partition C into n small arcs by forming a partition of [a, b] by picking s i as follows: a = s 0 < s 1 < < s n = b.

Chapter 4: Dynamic Programming

Results of quality assessment tasks of the optical -electronic complex using complex instrumentation background and astroclimatic conditions

First Midterm Examination

Chapter 7 Notes, Stewart 8e. 7.1 Integration by Parts Trigonometric Integrals Evaluating sin m x cos n (x) dx...

3.4 Numerical integration

Temporally-Biased Sampling for Online Model Management

Chapter Direct Method of Interpolation More Examples Civil Engineering

CBE 291b - Computation And Optimization For Engineers

APPROXIMATE INTEGRATION

A Robust Feature-Based Digital Image Watermarking Scheme

Lecture 21: Order statistics

Vyacheslav Telnin. Search for New Numbers.

Chapters 4 & 5 Integrals & Applications

The Regulated and Riemann Integrals

MAT 772: Numerical Analysis. James V. Lambers

LECTURE NOTE #12 PROF. ALAN YUILLE

SUMMER KNOWHOW STUDY AND LEARNING CENTRE

N 0 completions on partial matrices

1 The Riemann Integral

Definition of Continuity: The function f(x) is continuous at x = a if f(a) exists and lim

A-Level Mathematics Transition Task (compulsory for all maths students and all further maths student)

Transcription:

Clssifiction Prt 4 Dr. Snjy Rnk Professor Computer nd Informtion Science nd Engineering University of Florid, Ginesville Model Evlution Metrics for Performnce Evlution How to evlute the performnce of model Methods for Performnce Evlution How to obtin relible estimtes Methods for Model Comprison How to compre the reltive performnce mong competing models Dt Mining Snjy Rnk Fll 2003 2

Metrics for Performnce Evlution Focus on the predictive cpbility of model Rther thn how fst it tkes to clssify or build models, sclbility, etc. Confusion Mtrix: PREDICTED c b d : TP (true positive) b: FN (flse negtive) c: FP (flse positive) d: TN (true negtive) Dt Mining Snjy Rnk Fll 2003 3 Metrics for Performnce Evlution PREDICTED (TP) c (FP) B (FN) d (TN) Most widelyused metric: d TP TN Accurcy = = b c d TP TN FP FN Dt Mining Snjy Rnk Fll 2003 4 2

Cost Mtrix PREDICTED C(i j) C(Yes Yes) C(No Yes) C(Yes No) C(No No) C(i j): Cost of misclssifying clss j exmple s clss i Accurcy is useful mesure if C(Yes No)=C(No Yes) nd C(Yes Yes)=C(No No) P(Yes) = P(No) (clss distribution re equl) Dt Mining Snjy Rnk Fll 2003 5 Cost vs. Accurcy Cost Mtrix PREDICTED C(i j) 00 0 Model M PREDICTED Model M 2 PREDICTED C(i j) 50 60 40 250 C(i j) 250 5 45 200 Accurcy = 80% Cost = 390 Accurcy = 90% Cost = 4255 Dt Mining Snjy Rnk Fll 2003 6 3

CostSensitive Mesures Precision (p) = c Recll (r) = b 2rp 2 F mesure (F) = = r p 2 b c Precision is bised towrds C(Yes Yes) & C(Yes No) Recll is bised towrds C(Yes Yes) & C(No Yes) Fmesure is bised towrds ll except C(No No) w w d 4 Weighted Accurcy = w w b w c w d Dt Mining Snjy Rnk Fll 2003 7 2 3 4 Methods for Performnce Evlution How to obtin relible estimte of performnce Performnce of model my depend on other fctors besides the lerning lgorithm: Clss distribution Cost of misclssifiction Size of trining nd test sets Dt Mining Snjy Rnk Fll 2003 8 4

Lerning Curve Lerning curve shows how ccurcy chnges with vrying smple size Requires smpling schedule for creting lerning curve Arithmetic smpling Geometric smpling Effect of smll smple size Bis in the estimte Vrince of the estimte Dt Mining Snjy Rnk Fll 2003 9 Methods for Estimtion Holdout Reserve 2/3 for trining nd /3 for testing Rndom subsmpling Repeted holdout Cross vlidtion Prtition dt into k disjoint subsets kfold: trin on k prtitions, test on the remining one Leveoneout: k=n Strtified smpling Oversmpling vs. Undersmpling Bootstrp Smpling with replcement Dt Mining Snjy Rnk Fll 2003 0 5

Receiver Operting Chrcteristic (ROC) Developed in 950s for signl detection theory to nlyze noisy signls Chrcterize the trdeoff between positive hits nd flse lrms ROC curve plots TP (on the yxis) ginst FP (on the xxis) Performnce of ech clssifier represented s point on the ROC curve chnging the threshold of lgorithm, smple distribution or cost mtrix chnges the loction of the point Dt Mining Snjy Rnk Fll 2003 ROC Curve dimensionl dt set contining 2 clsses (positive nd negtive) Any point locted t x > t is clssified s positive At threshold t: TP=0.5, FN=0.5, FP=0.2, FN=0.88 Dt Mining Snjy Rnk Fll 2003 2 6

(TP,FP): (0,0): declre everything to be negtive clss (,): declre everything to be positive clss (,0): idel ROC Curve Digonl line: Rndom guessing Below digonl line: prediction is opposite of the true clss Dt Mining Snjy Rnk Fll 2003 3 Using ROC for Model Comprison No model consistently outperforms the other M is better for smll FPR M2 is better for lrge FPR Are under the ROC curve Idel, re = Rndom guess, re = 0.5 Dt Mining Snjy Rnk Fll 2003 4 7