Document Classification Using Expectation Maximization with Semi Supervised Learning

International Journal on Soft Computing ( IJSC ) Vol.2, No.4, November 2011
DOI : 10.5121/ijsc.2011.2404 37
Document Classification Using Expectation
Maximization with Semi Supervised Learning
Bhawna Nigam1
, Poorvi Ahirwal2
, Sonal Salve3
, Swati Vamney4
Department of Information Technology, IET, DAVV
1
bhawnanigam@gmail.com 2
poorvi.ahirwal@yahoo.in 3
salve_sonal@yahoo.com
4
svamney@gmail.com
Abstract
As the amount of online document increases, the demand for document classification to aid the analysis
and management of document is increasing. Text is cheap, but information, in the form of knowing what
classes a document belongs to, is expensive. The main purpose of this paper is to explain the expectation
maximization technique of data mining to classify the document and to learn how to improve the accuracy
while using semi-supervised approach. Expectation maximization algorithm is applied with both supervised
and semi-supervised approach. It is found that semi-supervised approach is more accurate and effective.
The main advantage of semi supervised approach is “DYNAMICALLY GENERATION OF NEW CLASS”.
The algorithm first trains a classifier using the labeled document and probabilistically classifies the
unlabeled documents. The car dataset for the evaluation purpose is collected from UCI repository dataset
in which some changes have been done from our side.
Keywords
Data mining, semi-supervised learning, supervised learning, expectation maximization, document
classification.
1.Introduction
Data mining [2][3] is the extraction of useful knowledge from large amount of data. Data mining
tools can provide solution to the business problems that were to too time consuming when done
manually. Classification[11] is one of the important aspect which comes under data mining and
is a predictive modeling technique.
It is used to group the data or documents in some pre-defined classes i.e. to classify them
according to attribute matching approach. Most companies have huge amount of data available
which has to be refined, to apply classification on them. Data mining techniques can be
implemented rapidly on existing software and hardware platforms to enhance the value of
existing information resources, and can be integrated with new products and systems as they are
brought on-line. Classification techniques [3] are used in various real world problems with
respect to application domain as well as for various research purposes relevance to today’s
business environment as well as a basic description of how data warehouse architectures can
evolve to deliver the value of data mining to end users. Data mining derives its name from the
similarities between searching for valuable business information in a large database — for
example, finding linked products in gigabytes of store scanner data — and mining a mountain for

38
a vein of valuable ore. Both processes require either sifting through an immense amount of
material, or intelligently probing it to find exactly where the value resides.
In general it is observed that nothing can be gained from the unlabeled data. But the fact is that
they provide the join probability distribution of words.
This white paper represents EXPECTATION MAXIMIZATION ALGORITHM[6] with semi
supervised approach that takes advantage of both labeled and unlabeled data. EM is a class of
iterative algorithm for maximum likelihood and maximum posterior estimation in problem with
unlabeled data.
2. Techniques Used for Classification
1. Artificial neural network.
2. Decision tree.
3. Expectation maximization algorithm.
4. Naïve baye’s algorithm.
2.1. Description of document classification using Expectation algorithm
2.1.1.Overview of Expectation Maximization.
Create an initial model, θ0. Arbitrarily, randomly, or with a small set of training examples. Use
the model θ’ to obtain another model θ such that
Σi log Pθ(yi) > Σi log Pθ’(yi)
Repeat the above step until reaching a local maximum. Guaranteed to find a better model after
each iteration.
2.1.2. Steps for preprocessing and classifying document[7] can be summarized as follows:
• Remove periods, commas, punctuation, stop words. Collect words that have occurrence
frequency more than once in the document. We called this collection of words as
vocabulary.
• View the frequent words as word sets by matching the words which are in the vocabulary
as well as training set documents.
• Search for matching word set(s) or its subset(containing items more than one) in the list
of word sets collected from training data with that of subset(s) (containing items more
than one) of frequent word set of new document.
• Collect the corresponding probability values of matched word set(s) for each target class.
• Calculate the probability
• Apply z score algorithm to calculate the range in which the attributes must be lying.
• Calculate the probability class by applying expectation maximization algorithm.
• Categorize the document in the class having maximum probability.
2.1.3. BASIC EM
The Expectation-Maximization algorithm is used in maximum likelihood estimation where the
problem involves two sets of random variables of which one, X, is observable and other, Z, is
hidden. In simpler words algorithm works in following two steps:

39
E-step
Estimates the expectation of the missing value i.e. unlabeled class information. This step
corresponds to performing classification of each unlabeled document. Probability distribution is
calculated using current parameter.
M-step
Maximizes the likelihood of the model parameter using the previously computed expectation
of the missing values as if were the true ones.
Step 1-
Given:
X - Labeled data
Z- Missing values
θ - Unknown parameter
L(θ;X,Z) = p(X,Z|θ) - likelihood function (probability).
L(θ|X) ϵ {α p(X| θ): α >0} , p(X| θ) * p(Z|θ)
Step 2:
With the given variables X, Z and θ, the maximum likelihood estimation of the unknown
parameters is calculated by the marginal likelihood of the observed data. The value obtained is
not tractable.
Finding maximum likelihood
L(θ;X) = p(X|θ) = (X,Z|θ)
Step 3- caculate expected value of log likelihood function.
Q(θ|θ^(t)) = [log L(θ;X,Z)]
Step 4- find the parameters that maximizes this quantity
θ^(t+1) = arg max Q(θ|θ^(t))
The observed data points X, may be discrete which are finite or taken from countably infinite set
or it may be continous which are taken from uncountably infinite set.
The EM algorithm can be applied to other data models when one of the parameters Z or θ is
known. This shows the iterative nature of algorithm that continues to predict the values until
converges to a particular and specific value.

40
3. Applying Expectation Maximization algorithm with Semi Supervised
Approach
We applied semi supervised approach for increasing the accuracy and to reduce the human efforts
for classifying the document. Until now EM is applied with supervised approach, but the
disadvantage of this is that the whole data must be labeled, and no dynamical generation of new
classes would be there, which is too time consuming and leads to reduction in speed of
classification.
In semi supervised approach we have both labeled and unlabeled data. From labeled data we give
training to data set and then the unlabeled are classified. Submitted document is categorized in
some pre defined classes and incase the document fails to lie in those defined classes then
dynamically a new class is generated and document is categorized in that class and that new class
is updated in database.
4. Implementation:
We are taking “car evaluation” dataset from UCI repository. We are considering 5 attributes
buying, maintenance, , price, mileage and safety and three pre-defined classes unacceptable,
good and very good. There are 1500 rows in the dataset. Half of the data is considered to be the
training data and remaining the testing. These 5 attributes are used to classify the document
submitted by user. Before the unlabeled document is read, two check are made. First the format
of the document is checked, if its .txt, then only its is processed further otherwise declared as
invalid. If the document passes the first check it must also lie only in the car domain to be
classified. If it is within domain then further steps of classification are done else process is ended
there only. The document to be classified is read and the classifier searches these five attributes in
that document. After mining these attributes from document the classifiers calculates the
probability of that document to belong to our pre-defined classes. We make a set of probabilities
which is the initial set. According to the probability calculated the classifier expects the class of
document i.e. E –step is done. Now use that set to create a subset of that provided the probability
of subset is maximum then the previous set. This procedure is done iteratively until and unless we
reaches at the maximum probability of falling any document in pre-defined classes. Classification
is predicated on the idea that electronic documents can be classified based on the probability that
certain keywords will correctly identify a piece of text document to its annotated category.
Document is classified in the correct class and if the probability does not match with any of these
classes, a new class is dynamically generated and updated in the dataset.
5. Comparison of Expectation Maximization with Supervised and Semi
Supervised Approach
Disadvantage of supervised approach with expectation maximization algorithm is that, first of all
the data should be labeled hence increases the efforts and time consuming and the unlabeled data
is useless here.
In case the document fails to lie in pre defined classes the approach is not able to classify that
document hence leading to decrease in efficiency of classifier.
In semi supervised approach unlabeled data is not useless, it has been used to train the classifier
and in case the document fails to lie in pre defined classes it leads to dynamically generation of
new class to categorize that document and the database is updated automatically.

41
6. Performance Analysis
For measuring the performance, the following definitions of precision, recall, accuracy and F1-
measure are used to find the effectiveness of document classifier. Accuracy of classifier is the
percentage of documents correctly classified by classifier
Accuracy = Number of correct prediction
Total number of prediction
Recall is determined by number of documents retrieved that are relevant with respect to total
number of documents that are relevant. Thus,
Recall=TP/(TP+FN)
Precision is determined by number of documents retrieved that are relevant with respect to total
number of documents that are retrieved. Thus,
Precision=TP/(TP+FP)
Thus,
F1-measure=(2*recall*precision)/(recall+precision)
Where,
TP (True Positive): The number of documents correctly classified to that class.
TN (True Negative): The number of documents correctly rejected from that class.
FP (False Positive): The number of documents incorrectly rejected from that class.
FN (False Negative): The number of documents incorrectly classified to that class.
8. Experimental Results
The work of classifying a new document depends mainly on the matching of values of attributes
with those of the values of predefined classes. We studied and implemented both the approaches
discussed above. For performance evaluation, we used the Accuracy, Precision and Recall
metrics that were presented in the previous section. In the following two figures both the
techniques are compared with each other. The comparison is done on the basis accuracy and F1-
Measure calculated for both the techniques. Accuracy and F1-Measure are compared with respect
to number of documents given to classifier.
As observed in the experiments, semi supervised approach of classification is ahead of supervised
method. Also, the dynamic generation of new class and updation in the database increases the
accuracy and precision of the result.
The results obtained on performing the experiments are:

42
Fig1. Experimental results
0
20
40
60
80
100
0 10 20 30
precision
0
20
40
60
80
0 10 20 30
Number of documents
Recall
recall
0
10
20
30
40
50
60
70
80
0 10 20 30
Numberof documents
Accurac
y
semisupervised
supervised
0
20
40
60
80
100
0 10 20 30
Number of documents
F1-Measure
Semisupervise
d
Supervised

43
9. Conclusion and Future Work
This paper has presented an algorithm that address the question of when and how unlabeled data
may be used to supplement scarce labeled data, especially when learning to classify text
document s. This is an important question in text learning because of the high cost of hand
labeling data and because of the availability of large volumes of unlabeled data. We have
presented an algorithm that takes advantage of it and experimental results that show significant
improvements by using unlabeled documents for training classifier in three real-world document
classification tasks.
When our assumption of data generation are correct, basic EM can effectively in corporate
information from unlabeled data. However, the full complexity of real-world text data cannot be
completely captured by known statistical models. It is interesting then, to consider the
performance of a classifier based on generative models that make incorrect assumptions about the
data. In such cases, when the data is inconsistent with the assumptions of the model, our method
for adjusting the relative contribution of the unlabeled data prevents the unlabeled data for
degrading classification.
References
[1] Data Mining: An AI Perspective, Xindong Wu1, Senior Member, IEEE.
[2] Jiawei Han and Micheline Kamber, 2001. “Data Mining: Concepts and Techniques”, Morgan
Kaufmann Publisher: CA.
[3] Arthur Dempster, Nan Laird and Donald Rubin, “Likelihood from incomplete data via the EM
algorithm”. Journal of the Royal Statistical Society, Series B, 39(1): 1-38, 1977.
[4] Carlos Ordonez, Edward Omiecinski. “FREM: Fast and Robust EM Clustering for Large Data Sets”,
Conference on Information and Knowledge Management, Proceedings of the 11th International
Information and Knowledge Management, Virginia, USA 2002, pg. 590-599.
[5] Kamal Nigam, Andrew McCallum, Sebastian Thrun Tom Mitchell, "Text Classification from Labeled
and Unlabeled Documents using EM ", Machine Learning Springer Netherlands, Vol. 39, Numbers
2-3/May 2000, pg. 103-134.
[6] Ruocco, A., & Frieder, O. (1997). Clustering and Classification of Large Document Bases in a
Parallel Environment. Journal of the American Society for Information Science , 48 (10).
[7] McCallum, A., & Nigam, K. (1998). A comparison of event models for Naive Bayes text
classification.
[8] Kamal Paul Nigam, “Using Unlabeled Data to Improve Text Classification“, Ph.D. Dissertation,
Carnegie Mellon University, PA, USA, May 2001.
[9] Mohammed Aljlayl, Ophir Frieder, & David Grossman, “On Arabic-English Cross-Language
Information Retrieval: A Machine Translation Approach”, IEEE Computer Society, ITCC 2002.
[10] Wee Sun Lee, Bing Liu, Philip S. Yu, Xiaoli Li, “ Partially Supervised Classification of Text
Documents”, Machine Learning-International Workshop Then Conference-, 2002
[11] Yoshimasa Tsuruoka, and Jun’ichi Tsujii," Training a Naive Bayes Classifier via the EM Algorithm
with a Class
[12] Nigam, Kamal, McCallum, Andrew, Thrun, Sebas-tian, and Mitchell, Tom M. Text classi¯ cation
fromlabeled and unlabeled documents using em. Machine Learning, 39(2/3):103{134, 2000.

44
[13] Mann, Gideon S. and McCallum, Andrew. General-ized expectation criteria for semi-supervised
learn-ing with weakly labeled data. Journal of MachineLearning Research, 11:955{984, 2010.
[14] Rayid Ghani, "Combining labeled and unlabeled data for text classification with a large number of
categories", Proceedings of the IEEE International Conference on Data Mining 2001.
[15] Chawla, Nitesh V. and Karakoulas, Grigoris J. Learn-ing from labeled and unlabeled data: An
empirical study across techniques and domains. J. Artif. In-tell. Res. (JAIR), 23:331{366, 2005.

Document Classification Using Expectation Maximization with Semi Supervised Learning

More Related Content

What's hot (19)

Similar to Document Classification Using Expectation Maximization with Semi Supervised Learning (20)

Recently uploaded (20)

Document Classification Using Expectation Maximization with Semi Supervised Learning