Overview

classication vs. numeric prediction

Linear Models
Connectionist and Statistical Language Processing
Frank Keller
[email protected]

linear regression least square estimation evaluating a numeric model, correlation selecting a regression model linear regression for classication regression trees, model trees

Computerlinguistik Universit at des Saarlandes

Literature: Witten and Frank (2000: ch. 4, 6), Howell (2002: ch. 15).
Linear Models p.1/26 Linear Models p.2/26

Numeric Prediction
An instance in the data set has the following general form:

Example
Predict CPU performance from conguration data: cycle time (ns) 125 29 29 29 29 memory min (kB) 256 8000 8000 8000 8000 2000 512 1000 memory max (kB) 6000 32000 32000 32000 16000 8000 4000 4000 cache (kB) 256 32 32 32 32 0 32 0 chan min 16 8 8 8 8 2 0 0 chan max 128 32 32 32 16 14 0 0 performance 198 269 220 172 132 52 67 45
Linear Models p.4/26

a1 , i , a2 , i , . . . , ak , i , xi
where a1,i , . . . , ak,i are attribute values, and xi is the target value, for the i-th instance in the data set. So far we have only seen classication tasks, where the target value xi is categorical (represents a class). Techniques such as decision tree and Naive Bayes are not (directly) applicable if the target is numeric. Instead algorithms for numeric prediction can be used, e.g., linear models.

...
125 480 480

Linear Models p.3/26

Linear Regression
Linear regression is a technique for numeric predictions thats widely used in psychology, medical research, etc. Key idea: nd a linear equation that predicts the target value x from the attribute values a1 , . . . , ak :

Linear Regression
The regression equation computes the following predicted value xi for the i-th instance in the data set.
k

(2)

xi = w0 + w1 a1,i , w2 a2,i , . . . , wk ak,i = w0 + w j a j,i

j=1

(1)

x = w0 + w1 a1 + w2 a2 + . . . + wk ak

Here, w1 , . . . wk are the regression coefcients, w0 is called the intercept. These are the model parameters that need to be induced from the data set.

Key idea: to determine the coefcients w0 , . . . wk , minimize e, the squared difference between the predicted and the actual value, summed over all n instances in the data set:

(3)

e = (xi xi )2 = xi w0 w j a j,i
i=1 i=1 j=1

The method for this is called Least Square Estimation (LSE).

Linear Models p.5/26

Linear Models p.6/26

Least Square Estimation

We demonstrate how LSE works with the simple case of k = 1, dropping the intercept w0 . The error equation (3) simplies to (abbreviating w1 = w and a1 = a):

Least Square Estimation

To minimize the squared error for the data set, we therefore set the derivative in (5) equal to zero:

(4)

e = (xi wai )2 = (xi2 2wai xi + w2 a2 i)

i i

(6)

2 ai xi + 2w a2 i =0
i i

Now differentiate the error equation in (4) with respect to w:

By resolving this equation to w, we obtain a formula for computing the value of w that minimizes the error:

(5)

e 2 = (2ai xi + 2wa2 i ) = 2 ai xi + 2w ai w i i i

(7)

w=

i ai xi i a2 i

The derivative is the slope of the error function. The slope is zero at all points at which the function has a minimum.

This formula can be generalized to regression equations with more than one coefcient.

Linear Models p.7/26

Linear Models p.8/26

Example
Sample data set:

Evaluating a Numeric Model

a x 1 2 2 5 1 2 5 8
The t of a regression model can be visualized by plotting the predicted data values against the actual values.
10 8 6 target value 4 2 0 -2 -2
Linear Models p.9/26

actual value x predicted value x

Use Least Square Estimation to compute w for this data set:

(8)

xi ai 1 2 + 2 5 + (1)(2) + 5 8 w= i 2 = = 1.74 12 + 22 + (1)2 + 52 i ai

Regression equation: x = wa = 1.74a

-1

1 2 3 attribute value a

6
Linear Models p.10/26

Evaluating a Numeric Model

A suitable numeric measure for the t of a linear model is the mean squared error:

Example
Compute the mean squared error for the sample data set and the regression equation x = 1.74a:

(9)

1 n MSE = (xi xi )2 n i=1

Intuitively, this represents how much the predicted values diverge from the actual values on average. Note that the MSE is the quantity the LSE algorithm minimizes.

a x x (x x )2 1 2 1.74 0.068 2 5 3.48 1.346 1 2 1.74 0.068 5 8 8.70 0.490 MSE = 0.646
1 n i=1

(xi xi )2 = 1 4 (0.068 + 1.346 + 0.068 + 0.490) =

Linear Models p.11/26

Linear Models p.12/26

Correlation Coefcient
The correlation coefcient r measures the degree of linear association between predicted and the actual values:

Example
Compute the correlation coefcient for the example data set:

(10) (11) (12) SP =

SPA r= SP SA n )(xi x ) (x x SPA = i=1 i n1 )2 n i=1 (xi x n1 SA = )2 n i=1 (xi x n1

Here x and x are the means of the actual and predicted values, SP and SA their standard deviations. SPA is the covariance of the actual and predicted values.
Linear Models p.13/26

x = 3.25 x = 3.14 SPA = ((1.74 3.14)(2 3.25) + (3.48 3.14)(5 3.25)+ (1.74 3.14)(2 3.25) + (8.70 3.14)(8 3.25))/3 = 18.13 2 SP = ((1.74 3.14)2 + (3.48 3.14)2+ (1.74 3.14)2 + (8.70 3.14)2)/3 = 18.93 2 = ((2 3.25)2 + (5 3.25)2+ SA (2 3.25)2 + (8 3.25)2)/3 = 18.25 r = 18.13/( 18.93 18.25) = 0.975 r2 = 0.951
Linear Models p.14/26

Correlation Coefcient
Some important properties:
The correlation coefcient r ranges from

Partial Correlations
We can compute the multiple correlation coefcient that tells us how well the full regression model (with all attributes) ts the target values. We can also compute the correlation between the values of a single attribute and the target values. However this is not very useful as attributes can be intercorrelated, i.e., they correlate with each other (colinearity). We need to compute the partial correlation coefcient, which tells us how much variance is uniquely accounted for by an attribute once the other attributes are partialled out.

1.0 (perfect correlation) to 0 (no correlation) to 1.0 (negative correlation). r expresses how well the data points t on the straight line described by the regression model. r is signicant. Null hypothesis: there is no

Intuitively,

We can test if

linear relationship between predicted and actual values.

We can also compute r2 , which represents the amount of

variance accounted for by the regression model.

Linear Models p.15/26 Linear Models p.16/26

Selecting a Regression Model

We want to build a regression model that only contains the attributes that are predictive. Several methods to achieve this:
All subsets: compute models for all subsets of attributes

Testing on Unseen Data

We compute the regression weights and perform the model selection on the training data. To evaluate the resulting model, we compute model t on unseen test data using LSE or the correlation coefcient. Techniques for testing on unseen data (see last lecture):
Holdout: set aside a random sample of the data set for

and chose the one with the highest multiple r.

Backward elimination: compute a model for all attributes,

and then eliminate the one with the lowest partial r. Iterate until the multiple r deteriorates.
Forward selection: compute a model consisting only of the

testing, train on the rest.

attributes with the highest partial r. Then add the next best attribute. Stop when the multiple r doesnt improve. Different model selection algorithm can yield different models.
Linear Models p.17/26

k-fold crossvalidation: split the data in k random partition

and test on each one in turn.

leave-one-out: set

k to the number of instances in the data set, i.e., test on each instance separately.
Linear Models p.18/26

Linear Regression for Classication

Regression can be applied to classication:
Perform a separate regression on each class, set the

Linear Separability
Linear regression approximates a linear function. This means that the classes have to be linearly separable.

target value to 1 if an instance is in the class, and 0 if it is not in the class.

attribute a2

membership function for the class (1 for members, 0 for non-members).

To classify a new instance, compute the regression value

for each membership function, and assign the new instance the class with the highest value.
This procedure is called multiresponse linear regression.
Linear Models p.19/26

attribute a1

attribute a2

The regression equation then approximates the

attribute a1

For many interesting problems, this is not the case.

Linear Models p.20/26

Regression Trees
Regression trees are decision trees for numeric attributes. The leaves are not labeled with classes, but with the mean of the target values of the instances classied by a given branch.
To construct a regression tree, chose splitting attributes to minimize intrasubset variation for each branch. Maximize standard deviation reduction (instead of information gain):

Example

(13)

SDR = T
i

|Ti | T |T | i

Where T is the set of instances classied at a given node, T1 , . . . , Ti are the subset that T is split into, and is the standard deviation.

Linear Models p.21/26

Linear Models p.22/26

Model Trees
Model trees are regression trees that have linear regression models at their leaves, not just numeric values.
Induction algorithm:
Induce a regression tree using standard deviation

Example

reduction as the splitting criterion.

Prune back the tree starting from the leaves. For each leaf construct a regression model that accounts

for the instances classied by this leaf. Model trees get around the problem of linear separability by combining several regression models.

Linear Models p.23/26

Linear Models p.24/26

Summary
Linear regression models are used for numeric prediction. They t a linear equation that combines attributes values to predict a numeric target attribute.

References
Howell, David C. 2002. Statistical Methods for Psychology. Pacic Grove, CA: Duxbury, 5th edn. Witten, Ian H., and Eibe Frank. 2000. Data Mining: Practical Machine Learing Tools and

Techniques with Java Implementations. San Diego, CA: Morgan Kaufmann.

Least square estimation can be used to determine the coefcients in the regression equation so that the difference between predicted and actual values is minimal.
A numeric model can be evaluates using the mean squared error or the correlation coefcient. Regression models can be used for classication either directly in multiresponse regression or in combination with decision trees: regression trees, model trees.
Linear Models p.25/26 Linear Models p.26/26