Improving Classifier Accuracy using Unlabeled Data..doc

IMPROVING CLASSIFIER ACCURACY USING UNLABELED DATA
Thamar I. Solorio Olac Fuentes

Department of Computer Science
Instituto Nacional de Astrofísica, Óptica y Electrónica
Luis Enrique Erro #1
Santa María Tonantzintla, Puebla, México

ABSTRACT Thus the question is: can we take advantage of the large
This paper describes an algorithm for improving classifier pool of unlabeled data? It would be extremely useful if we
accuracy using unlabeled data. This is of practical could find an algorithm that allowed improving
significance given the high cost of obtaining labeled data, classification accuracy when the labeled data are
and the large pool of unlabeled data readily available. The insufficient. This is the problem addressed in this paper.
algorithm consists of building a classifier using a very We evaluated the impact of incorporating unlabeled data
small set of previously labeled data, then classifying a to the learning process using several learning algorithms.
larger set of unlabeled data using that classifier, and Experimental results show that the classifiers trained with
finally building a new classifier using a combined data set labeled and unlabeled data are more accurate than the
containing the original set of labeled data and the set of ones trained with labeled data only. This is the result of
previously unlabeled data. The algorithm proposed here the overall averages from ten learning tasks.
was implemented using three well known learning
algorithms: feedforward neural networks trained with Even though the interest in learning algorithms that use
Backpropagation, the Naive Bayes Classifier and the C4.5 unlabeled data is recent, several methods have been
rule induction algorithm as base learning algorithms. proposed. Blum and Mitchell proposed a method for
Preliminary experimental results using 10 datasets from combining labeled and unlabeled data called co-training
the UCI repository show that using unlabeled data [1]. This method is targeted to a particular type of
improves the classification accuracy by 5% on average problem: classification where the examples can naturally
and that for 80% of the experiments the use of unlabeled be described using several different types of information.
data results in an improvement in the classifier's accuracy. In other words, an instance can be classified using
different subsets of the attributes describing that instance.
1. INTRODUCTION Basically, the co-training algorithm is this: two weak
classifiers are built, each one using different kind of
One of the problems addressed by machine learning is information, then, bootstrap from these classifiers using
that of data classification. Since the 1960’s, many unlabeled data. They focused on the problem of web-page
algorithms for data classification have been proposed. classification where each example can be classified using
However, all learning algorithms suffer the same the words contained in that page or using the links that
weakness: when the training set is small the classifier point to that page.
accuracy is low. Thus, these algorithms can become an
impractical solution due to the need of a very large Nigam et al. proposed a different approach, where a
training set. In many domains, unlabeled data are readily theoretical argument is presented showing that useful
available, but manual labeling is time-consuming, difficult information about the target function can be extracted
or even impossible. For example, there are millions of text from unlabeled data [2]. The algorithm learns to classify
documents available on the world-wide web, but, for the text from labeled and unlabeled documents. The idea in
vast majority, a label indicating their topic is not Nigam’s approach was to combine the Expectation
available. Another example is character recognition: Maximization algorithm (EM) with the Naive Bayes
gathering examples with handwritten characters is easy, classifier. They report an error reduction of up to 30%. In
but manual labeling each character is a tedious task. In this work we extended this approach, incorporating
astronomy, something similar occurs, thousands of spectra unlabeled data to three different learning algorithms, and
per night can be obtained with an automated telescope, evaluate it using several data sets form the UCI
but an astronomer needs several minutes to manually Repository [3].
classify each spectrum.
Unlabeled data have also been used for improving the
performance of artificial neural networks. Fardanesh and
Okan used the backpropagation algorithm, and the results

show that the classifier error can be decreased using The Naive Bayes classifier is a probabilistic algorithm
unlabeled data in some problem domains [4]. based on the simplifying assumption that the attribute
values are conditionally independent given the target
The paper is organized as follows: the next section values. Even though we know that in practice this
presents the learning algorithms. Section 3 describes how assumption does not hold, the algorithm's performance
unlabeled data are incorporated to the classifier's training. has been shown to be comparable to that of neural
Section 4 presents experimental results that compare the networks in some domains [6,7]. The Naive Bayes
performance of the algorithms trained using labeled and classifier applies to learning tasks where each instance x
unlabeled data to those obtained by the classifiers trained can be described as a tuple of attribute values <a1, a2, …
with labeled data only. Finally, some conclusions and an> and the target function f(x) can take on any value
directions for future work are presented. from a finite set V.
When a new instance x is presented, the Naive Bayes
2. LEARNING ALGORITHMS classifier assigns to it the most probable target value by
applying this rule:
Experiments in this work were made with three of the
most successful classification learning algorithms: F(x)=argmaxvj∈VP(vi)∏IP(ai|vj)
feedforward neural networks trained with
backpropagation, the C4.5 learning algorithm [5] and the To summarize, the learning task of the Naive Bayes is to
Naive Bayes classifier. build a hypothesis by estimating the different P(vi) and
P(ai|vj) terms based on their frequencies over the
2.1 Backpropagation and Feedforward training data.
Neural Networks

For problems involving real-valued attributes, Artificial
2.3 The C4.5 Algorithm
Neural Networks (ANNs) are among the most effective
learning methods currently known. Algorithms such as C4.5 is an extension to the decision-tree learning
Backpropagation use gradient descent or other algorithm ID3 [8]. Only a brief description of the method
optimization algorithm to tune network parameters to best is given here, more information can be found in [5]. The
fit a training set of input-output pairs. The algorithm consists of the following steps:
Backpropagation algorithm was applied in this work to a 1. Build the decision tree form the training set
feedforward network containing two layers of sigmoidal (conventional ID3).
units. 2. Convert the resulting tree into an equivalent set of
rules. The number of rules is equivalent to the
number of possible paths from the root to a leaf
node.
X1 HIDDEN LAYER 3. Prune each rule by removing any preconditions that
result in improving its accuracy, according to a
X2 validation set.
4. Sort the pruned rules in descending order according
to their accuracy, and consider them in this sequence
when classifying subsequent instances.
X3
Since the learning tasks used to evaluate this work involve
nominal and numeric values, we implemented the version
X4 WHO of C4.5 that incorporates continuous values.

WIH 3. INCORPORATING UNLABELED DATA
Figure 1. Representation of a feedforward neural network
with one hidden layer. The algorithm for combining labeled and unlabeled data is
described in this section. In the three learning algorithms
we apply this same procedure. First, the data set is divided
2.2 Naive Bayes Classifier randomly into several groups, one of these groups is
considered with its original classifications as the training
set, another group is separated as the test set and the

remaining data are the unlabeled examples. A classifier C1
is built using the training set and the learning algorithm As we can see in the three figures, the algorithm that
L1. Then, we use C1 to classify the unlabeled examples. shows the largest improvement with the incorporation of
With the labels assigned by C1 we merge both sets into unlabeled data is C4.5. In the ten learning tasks C4.5
one training set to build a final classifier C2. Finally, the presented an improvement average of 8% while the
test data are classified using C2. improvement averages for neural networks and Naive
Bayes were 5% and 3% respectively. Table 1 summarizes
The process described above was carried out ten times the results obtained in these experiments.
with each learning task and the overall averages are the
results described in the next section. 5. CONCLUSIONS AND FUTURE WORK

We have shown how learning from small sets of labeled
4. EXPERIMENTAL RESULTS training data can be improved upon with the use of larger
sets of unlabeled data. Our experimental results using
We used the following dataset form the UCI repository: several training sets and three different learning
wine, glass, chess, breast cancer, lymphography, balloons, algorithms show that for the vast majority of the cases,
thyroid disease, tic-tac-toe, ionosphere and iris. Figure 2 using unlabeled data improves the quality of the
compares the performance of C4.5 trained using the predictions made by the algorithms. This is of practical
labeled data only with the same algorithm using both significance in domains where unlabeled data are readily
labeled and unlabeled data as described in the previous available, but manual labeling may be time-consuming,
section. One point is plotted for each of the ten learning difficult or impractical. Present and future work includes:
tasks taken from the Irving repository of machine learning • Applying this methodology using ensembles of
datasets [2]. We can see that most points lie above the classifiers, where presumably the labeling of the
dotted line, which indicates that the error rate of the C4.5 unlabeled data and thus the final classifications
classifier trained with labeled and unlabeled data is assigned by the algorithm can be made more
smaller than the error of C4.5 trained with labeled data accurate.
only. Similarly, Figure 3 compares the performance of the • Experimental studies to characterize situations in
Naive Bayes classifier trained using labeled and unlabeled which this approach is not applicable. It is clear that
data to that obtained using only labeled data. Again, a when the set of labeled examples is large enough or
lower degree of error can be attained incorporating when the pseudo-labels can not be assigned
unlabeled data. Finally, Figure 4 shows the performance accurately, the use of unlabeled data can not improve
comparison of incorporating unlabeled data to a neural and may even decrease the overall classification
network's training to that using only labeled data. accuracy.

Naive Bayes
Classifier Neural Networks C4.5
C1 C2 Ratio C1 C2 Ratio C1 C2 Ratio
Wine 11.11 6.31 0.57 7.61 7.57 0.99 22.34 20.56 0.92
Glass 69.35 68.82 0.99 25.23 26.21 1.04 61.04 58.60 0.96
Chess 28.35 37.91 1.34 19.58 18.53 0.95
Breast 27.30 27.51 1.01 5.94 5.36 0.90 10.70 10.28 0.96
Lympho 27.82 36.72 1.32 40.17 38.28 0.95
Balloons 28.43 32.18 1.13 32.50 25.00 0.77
Tiroides 9.31 8.44 0.91 10.00 8.18 0.82 18.97 18.46 0.97
tic_tac_toe 34.87 32.98 0.95 20.94 19.81 0.95
ionosphere 64.04 64.04 1.00 12.52 12.47 1.00 25.64 21.81 0.85
Iris 4.93 2.64 0.54 9.80 9.42 0.96 18.10 16.76 0.93
Average 0.97 0.95 0.92

Table 1. Comparison of the error rates of the three algorithms. C1 is the classifier built using labeled data only, C2 is the
classifier built combining labeled and unlabeled data. Column Ratio presents results for C2 divided by the corresponding
figure for C1. In bold we can see lowest error for a given dataset and the largest reduction in error as a fraction of the original
error for each learning task. C4.5 shows the best improvement in 60% of the tasks. In 77% of the learning tasks the error was
reduced when using unlabeled data, and in 80% of the tasks the best overall results where obtained by a classifier that used
unlabeled data.

Figure 2. Comparison of C4.5 using labeled data only 6. ACKNOWLEDGEMENT: We would like to
with C4.5 using unlabeled data. Points above the diagonal thank CONACyT for partially supporting this work under
line exhibit lower error when the C4.5 is given unlabeled grant J31877-A.
data.
7. REFERENCES

[1] A. Blum, T. Mitchell, Combining Labeled and
Unlabeled Data with Co-Training. Proc. 1998
Conference on Computational Learning Theory, July
1998.

[2] K. Nigam, A. McCallum, S. Thrun , & T. Mitchell,
Learning to Classify Text from Labeled and
Unlabeled Documents, Machine Learning,
1999,1-22.

[3] C. Merz, & P. M. Murphy, UCI repository of
machine learning databases,
Figure 3. Comparison of Naive Bayes Classifier using http://www.ics.uci.edu./~mlearn/MLRepository.ht
labeled data only with Neural Network trained with a ml, 1996.
Naive Bayes Classifier using unlabeled data. Points above
the diagonal line exhibit lower error when the Neural
Network is given unlabeled data [4] M.T. Fardanesh and Okan K. Ersoy, "Classification
Accuracy Improvement of Neural Network
Classifiers by Using Unlabeled Data," IEEE
Transactions on Geoscience and Remote Sensing ,
Vol. 36, No. 3, 1998, 1020-1025.

[5] J. R. Quinlan, C4.5: Programs for Machine Learning
(San Mateo, CA: Morgan Kaufmann, 1993).

[6] D. Lewis, & M. Ringuette, A comparison of two 1997 International Conference on Machine
learning algorithms for text categorization, Third Learning.1997.
Annual Symposium on Document Analysis and
Information Retrieval, 1994, 81-93. [8] J. R. Quinlan, Induction of decision trees. Machine
Learning, 1(1), 1986, 81-106.
[7] T. Joachims, A probabilistic analysis of the Rocchio
algorithm with TFIDF for text categorization, Proc.
Figure 4.Comparison results of an ANN. Points above the
diagonal line exhibit lower error when the ANN is given
unlabeled data.

Improving Classifier Accuracy using Unlabeled Data..doc

More Related Content

What's hot (20)

Viewers also liked (8)

Similar to Improving Classifier Accuracy using Unlabeled Data..doc (20)

More from butest (20)

Improving Classifier Accuracy using Unlabeled Data..doc