Overview

training set, validation set, test set

Evaluation
Connectionist and Statistical Language Processing
Frank Keller
[email protected]

holdout, stratication crossvalidation, leave-one-out comparing machine learning algorithms comparing against a baseline precision and recall evaluating numeric prediction

Computerlinguistik Universit t des Saarlandes a

Literature: Witten and Frank (2000: ch. 6), Mitchell (1997: ch. 5).

Evaluation p.1/21

Evaluation p.2/21

Training and Test Set

For classication problems, we measure the performance of a model in terms of its error rate: percentage of incorrectly classied instances in the data set. We build a model because we want to use it to classify new data. Hence we are chiey interested in model performance on new (unseen) data. The resubstitution error (error rate on the training set) is a bad predictor of performance on new data. The model was build to account for the training data, so might overt it, i.e., not generalize to unseen data.

Test and Validation Set

Use two data set: the training set (seen data) to build the model (determine its parameters) and the test set (unseen data) to measure its performance (holding the parameters constant). Sometimes, we also need a validation set to tune the model (e.g., for pruning a decision tree). The validation set cant be used for testing (as its not unseen). All three data set have to be representative samples of the data that the model will be applied to.

Evaluation p.3/21

Evaluation p.4/21

Holdout
If a lot of data data are available, simply take two independent samples and use one for training and one for testing. The more training, the better the model. The more test data, the more accurate the error estimate. Problem: obtaining data is often expensive and time consuming. Example: corpus annotated with word senses, experimental data on subcat preferences. Solution: obtain a limited data set and use a holdout procedure. Most straightforward: random split into test and training set. Typically between 1/3 and 1/10 held out for testing.

Stratication
Problem: the split into training and test set might be unrepresentative, e.g., a certain class is not represented in the training set, thus the model will not learn to classify it. Solution: use stratied holdout, i.e., sample in such a way that each class is represented in both sets. Example: data set with two classes A and B. Aim: construct a 10% test set. Take a 10% sample of all instances of class A plus a 10% sample of all instances of class B. However, this procedure doesnt work well on small data sets.

Evaluation p.5/21

Evaluation p.6/21

Crossvalidation
Solution: k-fold crossvalidation maximizes the use of the data.
Divide data randomly into Train the model on

Crossvalidation
Example: data set with 20 instances, 5-fold crossvalidation test set training set

k folds (subsets) of equal size.

k 1 folds, use one fold for testing. k times so that all folds are used for

Repeat this process

testing.
Compute the average performance on the k test sets.

This effectively uses all the data for both training and testing. Typically k = 10 is used. Sometimes stratied k-fold crossvalidation is used.
Evaluation p.7/21

i1 , i2 , i3 , i4 i5 , i6 , i7 , i8 i9 , i10 , i11 , i12 i13 , i14 , i15 , i16 i17 , i18 , i19 , i20

compute error rate for each fold then compute average error rate

Evaluation p.8/21

Leave-one-out
Leave-one-out crossvalidation is simply k-fold crossvalidation with k set to n, the number of instances in the data set.
This means that the test set only consists of a single instance, which will be classied either correctly or incorrectly. Advantages: maximal use of training data, i.e., training on n 1 instances. The procedure is deterministic, no sampling involved. Disadvantages: unfeasible for large data sets: large number of training runs required, high computational cost. Cannot be stratied (only one class in the test set).

Comparing Algorithms
Assume you want to compare the performance of two machine learning algorithms A and B on the same data set. You could use crossvalidation to determine the error rates of A and B and then compute the difference. Problem: sampling is involved (in getting the data and in crossvalidation), hence there is variance for the error rates. Solution: determine if the difference between error rates is statistically signicant. If the crossvalidations of A and B use the same random division of the data (the same folds), then a paired t -test is appropriate.

Evaluation p.9/21

Evaluation p.10/21

Comparing Algorithms
Let k samples of the error rate of algorithms A be denoted by x1 , . . . , xk and k samples of the error rate of B by y1 , . . . , yk . Then the t for a paired t -test is:

Comparing against a Baseline

An error rate in itself is not very meaningful. We have to take into account how hard the problem is. This means comparing against a baseline model and showing that our model performs signicantly better than the baseline. The simplest model is the chance baseline, which assigns a classication randomly. Example: if we have two classes A and B in our data, and we classify each instance randomly as either A or B, then we will get 50% right just by chance (in the limit). In the general case, the error rate for the chance baseline is 1 1 for an n-way classication. n
Evaluation p.12/21

(1)

t=

d 2 /k d

Where d is the mean of the differences xn yn , and d is the standard deviation of this mean.

Evaluation p.11/21

Comparing against a Baseline

Problem: a chance baseline is not useful if the distribution of the data is skewed. We need to compare against a frequency baseline instead. A frequency baseline always assigns the most frequent class. Its error rate is 1 fmax , where fmax is the percentage of instances in the data that belong to the most frequent class. Example: determining when a cow is in oestrus (classes: yes, no). Chance baseline: 50%, frequency baseline: 3% (1 out of 30 days). More realistic example: assigning part-of-speech tags for English text. A tagger that assigns to each word its most frequent tag gets 20% error rate. Current state of the art is 4%.
Evaluation p.13/21

Confusion Matrix
Assume a two way classication. Four classication outcomes are possible, which can be displayed in a confusion matrix: predicted yes actual class yes no true positive false positive class no false negative true negative

True positives (TP): class members classied as class members True negatives (TN): class non-members classied as non-members False positives (FP): class non-members classied as class members False negatives (FN): class members classied as class non-members

Evaluation p.14/21

Precision and Recall

The error rate is an inadequate measure of the performance of an algorithm, it doesnt take into account the cost of making wrong decisions. Example: Based on chemical analysis of the water try to detect an oil slick in the sea.

Precision and Recall

Measures commonly used in information retrieval, based on true positives, false positives, and false negatives:

Precision: number of class members classied correctly over total number of instances classied as class members.

(2)

Precision =

|TP| |TP| + |FP|

False positive: wrongly identifying an oil slick if there is none. False negative: fail to identify an oil slick if there is one.
Here, false negatives (environmental disasters) are much more costly than false negatives (false alarms). We have to take that into account when we evaluate our model.
Evaluation p.15/21

Recall: number of class members classied correctly over total number of class members.

(3)

Recall =

|TP| |TP| + |FN|

Evaluation p.16/21

Precision and Recall

Precision and recall can be combined in the F -measure:

Evaluating Numeric Prediction

Error rate, precision, recall, and F -measure are mainly used for classication tasks. They are less suitable for tasks were a numeric quantity has to be predicted. Examples: predict the frequency of subcategorization frames. We want to measure how close the model is to the actual frequencies.

(4)

F=

2|TP| 2 recall precision = recall + precision 2|TP| + |FP| + |FN|

Example: For the oil slick scenario, we want to maximize recall (avoiding environmental disasters). Maximizing precision (avoiding false alarms) is less important. The F -measure can be used if precision and recall are equally important.

p1 , . . . , pn : predicted values of the for instances 1, . . . , n a1 , . . . , an : actual values of the for instances 1, . . . , n
There are several measure that compare ai and pi .

Evaluation p.17/21

Evaluation p.18/21

Mean Squared Error

Mean squared error measures the mean difference between actual and predicted values. Used e.g. for connectionist nets:

Summary
No matter how the performance of the model is measured (precision, recall, MSE, correlation), we always need to measure on the test set, not on the training set. Performance on the training only tells us that the model learns what its supposed to learn. It is not a good indicator of performance on unseen data. The test set can be obtained using an independent sample or holdout techniques (crossvalidation, leave-one-out). To meaningfully compare the performance of two algorithms for a given type of data, we need to compute if a difference in performance is signicant. We also need to compare performance against a baseline (chance or frequency).
Evaluation p.20/21

(5)

1 n MSE = (ai pi )2 n i=1

Often also the root mean squared error is used:

(6)

RMSE =

1 n (ai pi)2 n i=1

Advantage: has the same dimension as the original quantity.

Evaluation p.19/21

References
Mitchell, Tom. M. 1997. Machine Learning. New York: McGraw-Hill. Witten, Ian H., and Eibe Frank. 2000. Data Mining: Practical Machine Learing Tools and

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

Evaluation p.21/21