A Performance Analysis of CLMS and Augmented CLMS Algorithms for Smart Antennas
SASA Presentation 2013
1. 1
2013
Hildegard Erasmus & Morné Lamont
Hildegard Erasmus & Morné Lamont
University of Stellenbosch
Department Statistics & Actuarial Science
South African Statistical Association Conference 2013
Prediction Accuracy Estimation:
Applications to Linear Regression,
Regression Trees &
Support Vector Regression
2. Introduction
Classical linear regression
Alternative techniques
- Regression Trees
- Support Vector Regression
Advantages & Disadvantages
Simulation study
Conclusion
Prediction Accuracy Estimation SASA 2013
3. Regression as a scientific method first appeared around
1885 (Izenman, 2008)
Since then: regression evolved into variety of forms,
including linear, non-linear, parametric &
nonparametric
Objective of study:
- theoretical considerations regarding different regression
techniques
- highlight advantages & disadvantages
- assess performance of different techniques
- identify conditions for good predictive performance
Prediction Accuracy Estimation SASA 2013
Introduction
Classical linear
regression
Alternative techniques
Regression Trees
Support Vector
Regression
Advantages &
Disadvantages
Simulation study
Conclusion
4. Method of Least Squares (LS):
- Originated: Astronomy (1805)
- Legendre: developed LS method to determine the
orbits of planets
- Gauss & Laplace: Gaussian curve to describe error
component, crucial to success of the LS method
- Gauss: claimed to have used method of estimating
coefficients of set of linear equations by
minimizing error sum of squares since 1809
- Galton: develop ideas of regression & correlation
: fail to link LS with regression
- Yule: replace Gaussian error curve assumption with
assumption of linearly related variables (1897)
: proof that LS could be applied in regression
Prediction Accuracy Estimation SASA 2013
Introduction
Classical linear
regression
Alternative techniques
Regression Trees
Support Vector
Regression
Advantages &
Disadvantages
Simulation study
Conclusion
5. Classical Linear Regression:
- Model:
with : (n x 1) dependent or response variable
: unknown parameters
: design matrix with jth row , , … ,
: (n x 1) error term
- Assumptions (error term):
1.
2. ′
3. Normally distributed
Prediction Accuracy Estimation SASA 2013
Introduction
Classical linear
regression
Alternative
techniques
Regression Trees
Support Vector
Regression
Advantages &
Disadvantages
Simulation study
6. - Parameter estimation:
- Regression coefficients ( ) and error variance ( )
- Method of Least Squares
- Estimation from data:
′ and ′
Application inR:
-Function: lm,predict
150 200 250 300 350
051015202530
Linear regression
x
y
Prediction Accuracy Estimation SASA 2013
Introduction
Classical linear
regression
Alternative techniques
Regression Trees
Support Vector
Regression
Advantages &
Disadvantages
Simulation study
Conclusion
7. Alternative techniques:
1. Regression Trees
- Bagging
- Boosting
- Random Forest
2. Support Vector Regression
Prediction Accuracy Estimation SASA 2013
Introduction
Classical linear
regression
Alternative techniques
Regression Trees
Support Vector
Regression
Advantages &
Disadvantages
Simulation study
Conclusion
8. Regression Trees
Main idea:
- Nonparametric method to predict response variable y
from known input variables (Izenman, 2008)
- Classification and Regression Trees (CART) algorithm:
use recursive partitioning of input space into non-
overlapping rectangular (r=2) or cubic (r>2) regions &
fit simple prediction model within each partition
- Constant value assigned to each region as prediction
- Tree: graphical representation of partitioning
Setup:
- Learning data: , , 1,2, … , ,
Prediction Accuracy Estimation SASA 2013
Introduction
Classical linear
regression
Alternative techniques
Regression Trees
Support Vector
Regression
Advantages &
Disadvantages
Simulation study
Conclusion
9. |
ZN < 16.57
ZN < 5.75
CRIM < 48.75 ZN < 10.7
ZN < 18.84
ZN < 20.735-2.501-3.541-1.410-2.559
2.257
0.690-1.292
0 20 40 60 80 100
0510152025
Partitioning of Housing Data
CRIM
ZN
-2.50 -3.54
-1.41
-2.56
2.26
0.69
-1.29
Prediction Accuracy Estimation SASA 2013
Introduction
Classical linear
regression
Alternative techniques
Regression Trees
Support Vector
Regression
Advantages &
Disadvantages
Simulation study
Conclusion
10. Aspects to consider:
- Choose splitting conditions at each node
- Decision rule for when a node should be terminal
(node that does not split into two daughter nodes)
- Rule for assigning a predicted response value to every
terminal node
Application in R:
- Package: rpart
- Function: rpart, predict
Prediction Accuracy Estimation SASA 2013
Introduction
Classical linear
regression
Alternative techniques
Regression Trees
Support Vector
Regression
Advantages &
Disadvantages
Simulation study
Conclusion
11. Bagging (Bootstrap aggregating):
- Procedure combines an ensemble of learning algorithms
to improve performance over a single algorithm
(Breiman, 1996)
- Designed to reduce variance & improve stability
- Independently construct trees using bootstrap samples;
simple majority vote taken for prediction
Boosting:
- Reduce high bias of predictors that under fit the data
- Enhance accuracy of a “weak” (slightly >50% accuracy)
binary classification learning algorithm
- Successive trees give extra weight to incorrectly
predicted points; weighted vote taken for prediction
Prediction Accuracy Estimation SASA 2013
Introduction
Classical linear
regression
Alternative techniques
Regression Trees
Support Vector
Regression
Advantages &
Disadvantages
Simulation study
Conclusion
12. Random Forest:
- Add additional layer of randomness to bagging
(Breiman, 2001)
- Construct each tree using a different bootstrap sample
- Different tree construction:
split each node using the best among a randomly
chosen set of predictors (Liaw & Wiener, 2002)
- Robust against overfitting
- Very user-friendly: only two parameters
: number of variables in subset
& number of trees in forest
Prediction Accuracy Estimation SASA 2013
Introduction
Classical linear
regression
Alternative techniques
Regression Trees
Support Vector
Regression
Advantages &
Disadvantages
Simulation study
Conclusion
13. Support Vector Regression (SVR)
Main idea:
- “...computation of a linear regression function in a high
dimensional feature space where the input data are
mapped via a nonlinear function.”
(Basak, Pal, & Patranabis, 2007)
- Involve optimization of a convex loss function or
equivalently using quadratic optimization under
given constraints
Setup:
- Training data: , , … , ,
where : space of input patterns, say
: consist of predictors , … , , each dim 1
: number of variables
Prediction Accuracy Estimation SASA 2013
Introduction
Classical linear
regression
Alternative techniques
Regression Trees
Support Vector
Regression
Advantages &
Disadvantages
Simulation study
Conclusion
14. Goal:
- Find a function that deviates from actual responses with
at most distance while ensuring small coefficient values
- Tube formed around true regression function that
contains most of the data points
- Points falling outside of the tube: described by
introducing slack variables (Smola & Schölkopf, 2003)
- Approximate training data with linear function:
〈 , 〉
- Find function that will
minimize ‖ ‖ ∑ ∗ℓ
subject to
〈 , 〉
〈 , 〉 ∗
Prediction Accuracy Estimation SASA 2013
Introduction
Classical linear
regression
Alternative techniques
Regression Trees
Support Vector
Regression
Advantages &
Disadvantages
Simulation study
Conclusion
∗
0
15. - SV expansion in input space ( :
∑ ∗ 〈 , 〉
∗
: obtain via quadratic optimization
Kernel functions:
- Kernel: function K: → , such that for all , ,
, 〈Φ , Φ 〉
- Used to compute inner products of the form 〈Φ , Φ 〉
in feature space using nonlinear kernel in input space
- High dimensionality makes it computationally
expensive or impossible
- SV expansion in feature space :
∑ ∗
,
Prediction Accuracy Estimation SASA 2013
Introduction
Classical linear
regression
Alternative techniques
Regression Trees
Support Vector
Regression
Advantages &
Disadvantages
Simulation study
Conclusion
16. - Examples of kernel functions:
Name Kernel function Dim
Polynomial
, 〈 , 〉 1 , ∈
!
! !
Gaussian
, , ∈
∞
ANOVA ,
,
∞
Application in R:
- Package: kernlab
- Function: ksvm, predict
Prediction Accuracy Estimation SASA 2013
Introduction
Classical linear
regression
Alternative techniques
Regression Trees
Support Vector
Regression
Advantages &
Disadvantages
Simulation study
Conclusion
17. LINEAR
REGRESSION
REGRESSION TREES SVR
Advantages:
Easy to estimate
parameters
Conceptually simple Generalized
performance
Computationally easy Highly interpretable Wide, real-world
application
Handle missing values
well
Sparsity of support
vectors
Resistant to outliers Perform well when
p>>n
No normality
assumptions
No normality
assumptions
Disadvantages:
Assumptions Trees: unstable, high
variance
How to estimate
parameters
Outliers/influential
values
Lack of smoothness Parameters estimation
are computationally
intense
Multicollinearity Stepfunction: values not
always accurate
Which kernels to
choose
Variable selection
needed if p>>n
Prediction Accuracy Estimation SASA 2013
Introduction
Classical linear
regression
Alternative techniques
Regression Trees
Support Vector
Regression
Advantages &
Disadvantages
Simulation study
Conclusion
18. -0.6 -0.4 -0.2 0.0 0.2 0.4 0.6
161820222426
Fitted models
X-values
Y-values
Linear
Tree
Random Forest
-0.6 -0.4 -0.2 0.0 0.2 0.4 0.6
161820222426
Fitted models
X-values
Y-values
SVR:Polynomial
SVR:Gaussian
SVR:ANOVA
Prediction Accuracy Estimation SASA 2013
19. Software: R
Data sets: (Johnson & Wichern, 2007)
Parameter estimation: R functions optimize and optim
Prediction accuracy measures:
| |
Cross-validation: 100 simulations
: 70% training data, 30% test data
Bootstrap: 1000 bootstrap samples
Prediction Accuracy Estimation SASA 2013
Introduction
Classical linear
regression
Alternative techniques
Regression Trees
Support Vector
Regression
Advantages &
Disadvantages
Simulation study
Conclusion
20. Results (Natural gas data):
MSE MAPE
Technique Mean Sd Mean Sd
Linear
Regression CV 398.2812 92.7482 5.2777 0.6665
BS 335.9571 19.6197 4.8414 0.1375
Regression Tree CV
1006.6497 431.5992 8.7364 1.9983
BS
628.2681 122.0066 6.7289 0.6826
Random Forest CV
441.2309 124.8267 5.59 0.9815
BS
223.4301 51.7075 3.3958 0.3513
SVR: Polynomial CV
692.964 487.3593 6.2902 1.6811
BS
456.7031 333.1714 3.6937 0.7804
SVR: Gaussian CV
378.6259 102.3776 5.3426 0.8103
BS
297.2053 32.399 4.6845 0.2609
SVR: ANOVA CV
4363.471 860.2588 20.332 3.3114
BS
4044.828 766.1202 19.2571 1.0407
CV: Cross-validation
BS : Bootstrap
Prediction Accuracy Estimation SASA 2013
Introduction
Classical linear
regression
Alternative techniques
Regression Trees
Support Vector
Regression
Advantages &
Disadvantages
Simulation study
Conclusion
21. Results (Pulp & Paper data):
MSE MAPE
Technique Mean Sd Mean Sd
Linear
Regression CV 2.8228 1.3047 5.7085 1.2441
BS
2.3586 0.4082 5.2373 0.472
Regression Tree CV
2.9881 0.9126 6.1803 1.1149
BS
1.7547 0.26 4.6158 0.3316
Random Forest CV
1.4502 0.5845 4.3711 0.9242
BS
0.7125 0.223 2.5465 0.2797
SVR: Polynomial CV
1.7355 1.0912 4.7764 1.2274
BS
2.3586 0.4082 5.2373 0.472
SVR: Gaussian CV
1.4788 0.6458 4.409 0.9694
BS
1.9256 6.4087 3.8799 1.1653
SVR: ANOVA CV
5.6161 1.09 9.1277 1.2696
BS
0.7727 0.2732 2.7881 0.3638
Prediction Accuracy Estimation SASA 2013
Introduction
Classical linear
regression
Alternative techniques
Regression Trees
Support Vector
Regression
Advantages &
Disadvantages
Simulation study
Conclusion
22. Conclusion:
- SVR with Gaussian kernel performed the best for the
Natural gas data set
- Random Forest technique performed the best for the
Pulp and Paper data set
- SVR with ANOVA kernel performed the worst for both
data sets
- Linear regression expected to perform well when linear
relations exist; alternative techniques expected to
outperform when more complex relationships are present
- Corresponding results were obtained by Cross-validation
and Bootstrap methods in calculation of MSE and MAPE
measures
- Could add Out-of-Bag procedure for additional validation
Prediction Accuracy Estimation SASA 2013
Introduction
Classical linear
regression
Alternative techniques
Regression Trees
Support Vector
Regression
Advantages &
Disadvantages
Simulation study
Conclusion
23. References:
- Basak, D., Pal, S., & Patranabis, D. C. (2007). Support Vector Regression.
Neural Information Processing, 203-218.
- Breiman, L. (1996). Bagging predictors. Machine Learning, 123-140.
- Breiman, L. (2001). Random Forests. Machine Learning, 5-32.
- Izenman, A. J. (2008). Modern Multivariate Statistical Techniques. New York:
Springer Science+Business Media.
- Johnson, R. A., & W., W. D. (2007). Applied Multivariate Statistical Analysis.
NJ: Pearson Education, Inc.
- Johnson, R. A., & Wichern, D. W. (2007). Applied Multivariate Statistical
Analysis. NJ: Pearson Education, Inc.
- Liaw, A., & Wiener, M. (2002). Classification and Regression by randomForest. R
News.
- Smola, A. J., & Schölkopf, B. (2003). A Tutorial on Support Vector Regression.
Statistics and Computing, 199-222.
Prediction Accuracy Estimation SASA 2013