Machine Learning in Practice Lecture 8

Machine Learning in PracticeLecture 8

Carolyn Penstein Rosé

Language Technologies Institute/ Human-Computer Interaction

Institute

Plan for the Day Announcements

Should be finalizing plans for term project Weka helpful hints Spam Dataset Overcoming some limits of Linear

Functions Discussing ordinal attributes in light of

linear functions

Weka Helpful Hints

Feature Selection Feature

selection algorithms pick out a subset of the features that work best Usually they

evaluate each feature in isolation

* Click here to start setting up feature selection

Feature Selection Feature

selection algorithms pick out a subset of the features that work best Usually they

evaluate each feature in isolation

* Now click here

Feature Selection

* Now click here.

Feature Selection

Feature Selection

* Now pick your baseclassifier just like before

Feature Selection

* Finally you will configurethe feature selection

Setting Up Feature Selection

* First click here.

Setting Up Feature Selection

* Select CHiSquaredAttributeEval

Setting Up Feature Selection

* Now click here.

Setting Up Feature Selection

* Select Ranker

Setting Up Feature Selection

* Now click here

Setting Up Feature Selection

* Set the number of featuresyou want

Setting Up Feature Selection The number you pick

should not be larger than the number of features available

The number should not be larger than the number of coded examples you have

Examining Which Features are Most Predictive

You can find a ranked list of features in the Performance Report if you use feature selection

* Predictiveness score

* Frequency

Spam Data Set

Spam Data Set

Word frequencies

Runs of $, !, Capitalization

All numeric Spam versus

NotSpam

* Which algorithm will work best?

Spam Data Set Decision Trees (.85 Kappa) SMO (linear function) (.79 Kappa) Naïve Bayes (.6 Kappa)

What did SMO learn?

Decision tree model

More on Linear Functions… exploring the idea of nonlinearity

Limits of linear functions

Numeric Prediction with the CPU Data Predicting CPU

performance from computer configuration

All attributes are numeric as well as the output

Numeric Prediction with the CPU Data Could discretize the output

and predict good performance, mediocre performance, or bad performance

Numeric prediction allows you to make arbitrarily many distinctions

Linear Regression

R-squared= .87

Outliers

** Notice that here it’s the really high values that fit the line the least well. That’s not always the case.

The two most highly weighted features

Exploring the Attribute Space

* Identify outliers with respect to typical attribute values.

The two most highly weighted features

Within 1 standard deviation

of the mean value

Trees for Numeric Prediction

Looks like we may need a representation that allows for a nonlinear solution

Regression trees can handle a combination of numeric and nominal attributes

M5P: computes a linear regression function at each leaf node of the treeLook at CPU performance data and compare a

simple linear regression (R = .93) with M5P (R = .98)

Results on CPU data with M5P

MoreData Here

BiggestOutliersHere

Results with M5P

MoreData Here

BiggestOutliersHere

Multi-Layer Networks can learn arbitrarily complex functions

Multilayer Perceptron

Best Results So Far

Forcing a Linear Function

Note that it weights the featuresdifferently than the linear regression

Partly because of normalization

Regression trees split on MMAXNN emphasizes MMIN

Review of Ordinal Attributes

Feature Space Design for Linear Functions

Often features will be numeric Continuous values May be more likely to generalize properly

with discretized valuesWe discussed the fact that you lose ordering

and distanceWith respect to linear functions, it may be more

important that you lose the ability to think in terms of ranges

Explicitly coding ranges allows for a simple form of nonlinearity

Ordinal Values Weka technically does not have ordinal

attributes But you can simulate them with “temperature coding”! Try to represent “If X less than or equal to .35”?

.2 .25 .28 .31 .35 .45 .47 .52 .6 .63

A B C D

A

A or B

A or B or C

A or B or C or D

Ordinal Values Weka technically does not have ordinal

attributes But you can simulate them with “temperature coding”! Try to represent “If X less than or equal to .35”?

.2 .25 .28 .31 .35 .45 .47 .52 .6 .63

A B C D

A

A or B

A or B or C

A or B or C or D

Ordinal Values Weka technically does not have ordinal

attributes But you can simulate them with “temperature coding”! Try to represent “If X less than or equal to .35”?

.2 .25 .28 .31 .35 .45 .47 .52 .6 .63

A B C D

A

A or B

A or B or C

A or B or C or D

Now how wouldyou represent

X <= .35?

Ordinal Values Weka technically does not have ordinal

attributes But you can simulate them with “temperature coding”! Try to represent “If X less than or equal to .35”?

.2 .25 .28 .31 .35 .45 .47 .52 .6 .63

A B C D

A

A or B

A or B or C

A or B or C or D

Now how wouldyou represent

X <= .35?

Feat2 = 1

Take Home Message Linear functions cannot learn interactions

between attributes If you need to account for interactions:

Multiple layersTree-like representationsAttributes that represent rangesLater in the semester we’ll talk about other

approaches