Text categorization using compression models
 Eibe Frank, Chang Chui and Ian H. Witten Department of Computer Science
 University of Waik ato
 Hamilton, New Zealand
 feibe, ckc1, ihwg@cs.waikato.ac.nz 1 Introduction
 Text categorization, or the assignment of natural language texts to predefined
categories based on their content, is of growing importance as the volume
of information available
 on the internet continues to overwhelm us. The use of predefined categories
implies a \\supervised learning" approach to categorization, wherealready-classified
articles|which effectively define the categories|are used as \\training data"
to build a model that can be used for classifying new articles that comprise
the \\test data." This contrasts with
 \\unsupervised" learning, where there is no training data and clusters of
like documents are sought amongst the test articles. With supervised learning,
meaningful labels {such as
 keyphrases}are attached to the training documents, and appropriate labels
can be assigned
 automatically to test documents depending on which category they fall into.
 Text categorization is a hot topic in machine learning. Typical approaches
extract
 \\features" from articles, and use the feature vectors as input to a machine
learning scheme
 that learns how to classify articles. The features are generally words.
Because there are so
 manyof them, a selection process is applied to determine the most important
ones, and the
 remainder are discarded. This \\bag of words" model neglects word order
and contextual effects. It also raises some problems: how to define a \\word,"
what to do with numbers
 andother non-alphabetic strings, and whether to apply stemming. It has often
been observed that compression seems to provide a very promisingalter-
 native approach to categorization. The overall compression of an article
with respect to
 different models can be compared to see which one it fits most closely.
Such a scheme has several potential advantages: 
*  it yields an overall judgement on the document as a whole, rather than
discarding
 information by pre-selecting features;
 
*  it avoids the messy and rather artificial problem of defining word boundaries;
 
*  it deals uniformly with morphological variants of words;
 
*  depending on the model {and its order}, it can take account of phrasal
effects that
 span word boundaries;
 
*  it offers a uniform way of dealing with different types of documents|for
example, arbitrary files in a computer system; 
*  it generally minimizes arbitrary decisions that inevitablyneed to be taken
to render
 any learning scheme practical. Furthermore, a compression-based approach
to text categorization does offer potential im-
 provements in compression performance, by selecting the most suitable model
for each
 1text on an individualbasis and transmitting its identity to the receiver|although
it is
 categorization, not compression performance, that is our primary motivator.
W e have performed extensive experiments on the use of compression models
for cat-
 egorization using a standard dataset. This has involved working out how
to deal with
 the {normal} situation where a document may belong to several categories
{not merely
 choosing the one that it fits best}. We report some encouraging results
on two-category
 situations, and the results on the general problem seem reasonably impressive|in
one case
 outstanding. Compression-based methods certainly succeed in categorizing
the ma jority of
 documents correctly, and compare quite well with simple machine learning
schemes.
 However, we find that compression-based methods do not compete with the
published
 state of the art in the use of machine learning for text categorization
{although, as men-
 tioned in Section 2.1, the art is rather difficult to replicate because
it is not fully described in the literature}. Some reasons why this is the
case are discussed in the closing section. We have two overall conclusions.
First, a negative result: we do not recommend the use
 of compression models for text categorization if one seeks the best possible
categorization performance. Second, a methodological point: results in this
area should be evaluated
 comparatively with respect to the state of the art|it is too easy to give
a positive impression
 by avoiding direct, quantitative, comparison with other work.
 2 Existing approaches to text categoriza tion
 T ext categorization is a supervisedlearning task where a test document
is classified into
 categories using a mapping derived from a set of labeled training documents.
This is
 a standard setting in machine learning, and there is a host of learning
algorithms for
 such problems {Witten and Frank, 2000}|many of which have also been applied
to text
 categorization. They all require documents to be transformed into feature
vectors before
 learning can take place. In the following we briefly review how this is
done, and which
 supervised learning schemes have been applied.
 2.1 Dat a prepara tion
 Standard approaches to text categorization using supervised learning represent
each docu-
 ment by the set of words it contains. Generally, each word is a binary feature
{Dumais et al.,
 1998}, although more complicated procedures based on combinations of term
frequencies
 and inverse document frequencies are also possible {Yang and Pedersen, 1997}.
The low-level problems of word identification and extraction are generally
brushed un- der the carpet. For example, Dumais et al. {1998} state only
that \\text files are processed
 using Microsoft's Index Server." Yang {1999} uses the SMART system {Salton,
1989} for
 removing stop words and stemming. Yet it is possible that text categorization
results are
 quite sensitive to the precise details of word extraction. For example,
financial articles
 may be distinguished by a prevalence of numeric dollar figures, which may
well be dis-
 carded wholesale by a preprocessor. There have been no studies of the robustness
of text
 categorization to changes in these low-level decisions.
 Most schemes perform feature extraction prior to learning by selecting a
small number
 of words to participate in the learning phase and discarding the rest {Yang
and Pedersen,
 1997}. For learning schemes that are sensitive to irrelevant features, this
improves perfor-
 mance markedly; if a scheme's computational complexity depends heavily on
the number
 of features, it may be the only way to make learning feasible. For example,
Dumais et
 2al. {1998} selected between 50 and 300 features for each category, based
on a mutual infor-
 mation measure between a feature and a category.
 2.2 Learning schemes Many supervisedlearning methods have been applied to
the problem of text categorization.
 Information retrieval metrics , used by full-text retrieval systems to allow
users to sharpen their queries using relevance feedback {Rocchio, 1971},
have been used by imagining a
 query that contains all the words in the test document and using weights
derived from the documents in each class {Dumais et al., 1998}. Naive Bayes
classifiers estimate the
 probability of each feature given each category from the training data {Langley
et al .,
 1992}, assumingstatistical independence of the features {which is why the
method is called \\naive"}. The Bayes net technique {Sahami, 1996} models
limited dependence between
 different features, and has also been applied to text categorization {Dumais
et al., 1998}. Near est-neighb or classifiers assign to a test document the
class of the training document
 that most closely matches it. A \\divide-and-conquer" approach leads naturally
to a decision
 tr e e {Lewis and Ringuette, 1994}. A linear model assigns weights to the
features during the
 training phase, and sums them for each feature that appears in the test
document {Lewis et al., 1996}. Neural nets use multi-stage combinations of
simple non-linear models {Ng
 et al., 1997}. Linear support vector machines select a small number of critical
boundary
 instances {i.e. documents} from each category and build a linear discriminant
function that separates them as widely as possible. The best results for
text categorization have been obtained using support vector ma- chines {Dumais
et al., 1998}, committees of decision trees {Apte et al., 1998}, and nearest-
 neighbor classifiers that consider k nearest neighbors instead of only one
{Yang, 1999}. 3 Text categoriza tion using PPM
 All these approaches to text categorization share the disadvantage that
input documents
 must be converted into feature vectors before they can be processed. This
involves many arbitrary decisions, making experimental results hard to replicate.
The effect of these deci-
 sions has never been thoroughly investigated. Pre-processing requires language-dependent
 mechanisms like stemming that may not be readily available for the language
in ques-
 tion. Finally, if text categorization is considered from a broader point
of view|where a \\text" can be any character stream|word-based approaches
will necessarily exhibit de- ficiencies. Ideally, a text categorization scheme
should be able to classify arbitrary files,
 not just English-language documents. Its success should depend only on the
availability of
 sufficient training data, not on the type of documents to which it is applied.
 In contrast to general-purpose classification methods that require extensive
data prepa- ration, compressiontechniques deal with arbitrary sequences of
characters. Hence they
 offer the prospect of a uniform approach to text categorization. The question
is whether
 they can be successfully applied to the task of discriminatingbetween classes
of documents.
 In the following we investigate this question using the PPM compression
scheme with order
 2 and escape method C {Bell et al ., 1990}. Other orders were tried, but
both lower and
 higher choices were found to degrade performance in almost all cases|presumably
because the amount of training data available is insufficient to justify
more complex models.
 3Training data Test dataArticles Text {Kb} Articles Text {Kb}corn181 210
56 81corporate acquisitions1650 1307 719 542crude oil389 522 189 206earnings2877
1460 1087 457grain433 478 149 166interest347 329 131 147money market538 610
179 211shipping197 212 89 94trade issues369 569 117 180wheat212 235 71 77Table
1: Corpus of Reuters articles used in experiments
 3.1 The benchmark dat a
 All our results are based on the Reuters-21578 collection of newswire stories,
 1
 divided
 into training and test documents using the ModApte split|the standard testbed
for the
 evaluation of text categorization schemes. In total there are 12,902 stories,
averaging 200
 words each, thathave been classified into 118 categories. However, the distribution
of
 stories among categories is highly skewed: the ten largest contain 75\045
of stories. These
 ten categories|earnings, corporateacquisitions, money market, grain, crude
oil, trade issues, interest, shipping,wheat, and corn|are shown in Table
1, along with the number
 of training and test stories that each one contains. A story does not necessarily
belong
 to only one category; manystories are assigned to multiple categories, and
some are not
 assigned to any category at all.
 3.2 Experiments using pair wise discrimination
 Application of a straightforward compression methodology to the problem
of text catego-
 rization quickly yields encouraging results. Consider the two-class case.
To distinguish documents of class A from documents of class B, we form separate
compression models M
 A
 and M
 B
 from the training documents of each class. Then, given a test document {different
 from the training documents}, we compress it according to each model and
calculate the
 gain in per-symbol compression obtained by using M
 A
 instead of M
 B
 . We assign the docu- ment to one or the other class depending on whether
this difference is positive or negative,
 on the principlethat M
 A
 will compress documents of class A better and similarly for M
 B
 .
 Figure 3.2 shows results for ten pairs of categories from the Reuters data,
using the
 ModApte split from Table 1. 2
 The graphs show, onthe vertical axis, the difference in compression. The
vertical line to the left of each plot shows the test documents of one class,
 andthe vertical line to the right shows the test documents of the other.
The fact that almost
 all the points in the lefthand line lie above the zero line, and almost
all in the righthand line lie below it, indicates that almost all test documents
are classified correctly. Superimposed
 on each vertical line is a box whose center indicates the average compression
difference
 for that class, and whose extent indicates the standard deviation of the
distribution. The
 lefthand boxes lie comfortably above the line and the righthand ones comfortably
below it.
 In Figure 3.2a just one article is miscategorized, and that by only a small
margin. In
 Figures 3.2b and 3.2c there are no miscategorizations. Figure 3.2d also
shows very few1
 The collection is publicly available at www.research.att.com/\030lewis/reuters21578.html.
 2
 However, for legibility Figure 3.2 only shows one-half of the test results.
 4atendatend Helvetica-2-1012-2-1012atendatend Helveticaatendatend Helveticaatendatend
Helveticaatendatend Helveticaatendatend Helveticaatendatend Helveticaatendatend
Helveticaatendatend Helveticaatendatend Helvetica{a} {b} {c} {d} {e} {f }
{g} {h} {i} {j}
 Figure 1: Pairwise discrimination using compression: {a} money market vs
shipping; {b}
 grain vs interest; {c} earnings vs wheat; {d} corporate acquisitionsvs trade
issues; {e}
 corporateacquisitions vs grain; {f } crude oil vs earnings; {g}corporate
acquisitions vs earnings; {h} corn vs wheat; {i} grain vs wheat; {j} corn
vs grain.
 errors,but one is by a rather large margin. Figures 3.2d and 3.2e show a
new phenomenon: an article {just one in each case} that is assigned to both
categories, displayed in the
 middle.Clearly a pairwisediscrimination policy cannot handle such cases.
In Figure 3.2d
 the doubly-classified article is near the zero point, while in Figure 3.2e
it is far above it.
 Figure 3.2f{j show some less satisfactory results. In Figure 3.2f, several
{19} earnings articles are misclassifiedas crude oil , one crude oil article
is miscategorized by a large mar-
 gin, and there is one article that belongs to both categories. Results would
be improved by
 choosing a small positive number, instead of zero, as the threshold. In
Figure 3.2g many
 {44}earnings articles are misclassified as corp or ate acquisitions and
two corp or ate acquisi-
 tions articles are assigned to the category earnings ; inaddition, there
are two articles that
 belong to both. Again, results would be improved by choosing a small positive
threshold. Figure3.2h shows an article being misclassified by a rather large
margin|the compression
 metric assigns one of the wheat articles to the corn category by a very
large margin; in fact,this article is more corn-like {one hesitates to say
\\corny"} than most corn articles.
 A significant number of articles belong to both categories. Figures 3.2i
and 3.2j show an
 extreme situation where all but one of the wheat articles, and all of the
corn articles, also
 belong to the grain category. Wheat and corn evidently form a subclass of
grain and cannot be distinguishedusing this methodology.
 T able 2 summarizes the situation for all possible pairwise discriminations.
We com-
 press articles corresponding to the row but not the column according to
{a} the model
 corresponding to the row and {b} the model corresponding to the column,
and present the
 mean compression difference in bits/character, {b} { {a}, averaged over
the test articles.
 Themeans are positive, indicating that they lie on the correct side of the
zero line. The
 only exception is that corresponding to the right-hand \\bar" of Figure
3.2i, wheat vs grain , where there is only one article on wheat that is not
also on grain , and that article is classified
 5corn corp. crude earn- grain inter- money ship- trade wheat
 acq. oil ings est market ping issuescorn -  0.43 0.39 0.53 0.00 0.62 0.50
0.38 0.45 0.09corp. acq.0.79  -  0.40 0.31 0.59 0.65 0.63 0.47 0.61 0.68crude
oil0.45 0.26  -  0.35 0.37 0.48 0.43 0.37 0.38 0.46earnings1.47 0.99 1.15
-  1.18 1.53 1.48 1.34 1.39 1.27grain0.13 0.36 0.33 0.47  -  0.54 0.44 0.33
0.41 0.11interest0.65 0.26 0.33 0.36 0.46  -  0.17 0.52 0.55 0.59money market0.57
0.39 0.37 0.50 0.44 0.18  -  0.46 0.32 0.55shipping0.27 0.19 0.16 0.32 0.20
0.38 0.29  -  0.31 0.24trade issues0.36 0.32 0.25 0.45 0.26 0.33 0.20 0.35
-  0.34wheat0.11 0.37 0.35 0.49 {0.06 0.58 0.46 0.33 0.45  - Table 2: Mean
difference in compression between model corresponding to row and model
 corresponding to column, for articles corresponding to row but not to column
 incorrectly. These results paint a generally encouraging picture, but underline
the fact that the pairwise methodology needs extending to cope with multiply-classified
articles.
 3.3 Building positive and negative models
 For multiply-classifiedarticles, we decide whether a model belongs to a
particular category independently of whether it belongs to any other category.
We build positive and negative
 models for each category, the first from all articles that belong to the
category andthe second from those that do not. For a particular category
C , call these models M
 P
 and M
 N
 respectively.
 Given a new article A, denote its length when compressed according to these
models by L[AjM
 P ] and L[AjM N
 ]. From these lengths, the article's probability given the categories C
and
 \026
 C is:
 P r[AjC ] = 2
 \000L[AjM
 P
 ]
 P r[Aj \026
 C ] = 2 \000L[AjM N
 ]
 Bayes' formulagives the probability that a particular article A belongs
to category C :
 P r[C jA] =
 P r[AjC ]P r[C ]P r[AjC ]P r[C ] + P r[Aj \026
 C ]P r[
 \026
 C ]
 The prior probability of C is the proportion of articles belonging to that
category, and the
 denominator is the prior probability of article A. 3.3.1 Setting the threshold
Now we turn to the question of deciding whether a new article should in fact
be assigned to
 C or not. This presents the tradeoff between making the decision liberally,
increasing the chance that a category-C article is correctly identified but
also increasing the number of
 \\false positives"; or conservatively , reducing the number of false positives
but also increasing
 the number of \\false negatives." This tradeoff is familiar in information
retrieval, where a search engine must decide how long a list of articles
to present to the user, balancing the
 disadvantage of too many false positives {irrelevant documents that are
displayed} if the
 list is too long against too many false negatives {relevant documents that
are not displayed}
 if it is too short. Following standard usage, we quantify this tradeoff
in terms of recall and precision. In
 order to allow comparison of our results with others, we strive to maximize
the average of
 recall and precision|a figure that is called the \\breakeven point." 6PPM
Naive LSVMOverlapDumais et al. Number of featuresBayes{1998} 5 50 300corn54.2
65.3 90.30.04965.3 83.3 61.2 57.8corporate acquisitions91.0 87.8 93.60.03087.8
66.2 84.5 85.7crude oil80.7 79.5 88.90.04479.5 76.9 82.6 83.5earnings96.3
95.9 98.00.02095.9 91.1 95.1 96.4grain74.6 78.8 94.60.03878.8 84.3 82.2 78.9interest60.4
64.9 77.70.04564.9 55.4 59.8 52.8money market76.3 56.6 74.50.05356.6 50.9
61.2 61.0shipping81.9 85.4 85.60.03985.4 71.3 83.1 83.7trade issues65.0 63.9
75.90.04763.9 63.4 67.0 57.0wheat64.9 69.7 91.80.05169.7 85.3 74.9 68.2Table
3: Recall/precision breakeven point for compression-based categorization
compared
 with Naive Bayes and linear support vector machines; also {on the right},
subsidiary results
 for Naive Bayes
 The basic strategy is to compare the predicted probability P r[C jA] with
a predeter-
 mined threshold t, and declare A to have classification C if the probability
exceeds the threshold. The threshold is chosen individually, for each class,
to maximize the average of
 recall and precision for that class. To this end the training data is further
divided into a
 new training set {2/3 of the training data} and a validation set {1/3 of
the training data}.
 The threshold t is chosen to maximize the average of recall and precision
for the category {the breakeven point} on the validation set. Once it is
obtained, maximum utility is made
 of the training data by rebuilding the models M
 P and M
 N
 based on the full training data.
 As an additional benefit, threshold selection automatically adjusts for
the fact that M
 P
 and M
 N
 are formed from different amounts of training data. In general, one expects
to
 achieve better compression with more training data. On the other hand, the
results in Figure 3.2 indicate {to our surprise} that differing amounts of
training data do not have a
 strong influence on pairwise discrimination: it does not seem essential
for good performance
 to compensate for training set size.
 3.3.2 Results
 T able 3 shows the breakeven points obtained from our experiments, and compares
them
 with the results obtained by Dumais et al. {1998} for the Naive Bayes and
Linear Support
 Vector Machine methods. {Ignore the rightmost block of figures; we return
to them in
 Section 4}. PPM performs better than Naive Bayes on the six largest categories
{grain is
 theonly exception} and worse on the four smallest ones. It is almost uniformly
inferior to
 the support vector method, money market being the only exception. Compared
to the support vector method, PPM produces particularly bad results on the
categories wheat and corn . These two categories are {almost} proper subsets
of the
 category grain . This is because articles in grain summarize the result
of harvesting grain
 products|for example, by listing the tonnage obtained for each crop. These
articles use
 very similar terminology. Consequently the model for wheat is very likely
to assign a high
 score to every article in grain .
 It is the occurrence of the term \\wheat" that is the only notable difference
between an
 article in grain that belongs to wheat and one that does not. The presence
of a single word
 is unlikely to have a significant effect on overall compression of an article,
and this is why
 PPM performs poorly on these categories.
 Support vector machines perform internal feature selection, and can focus
on a single
 7corn corp. crude earn- grain inter- money ship- trade wheat
 acq. oil ings est market ping issuescorn0 0 0 0 43 0 0 0 2 29corporate acquisitions0
0 10 18 0 0 0 0 0 0crude oil0 5 0 2 2 0 1 14 2 0earnings1 8 0 0 1 0 0 0 0
0grain0 0 2 0 0 0 0 4 8 0interest2 48 11 15 5 0 108 3 20 1money market0 0
0 0 0 43 0 0 19 0shipping3 0 8 0 7 0 0 0 0 4trade issues0 3 4 0 7 16 38 5
0 1wheat13 0 0 0 26 0 0 0 2 0Table 4: \\False positive" confusion matrix
for the predictions made by PPM
 word if that is the only discriminatingfeature of a category. In comparison,
Naive Bayes performs badly on the same categories as PPM {money marketis
the only exception}. This
 is because, like PPM, it has no mechanism for internal feature selection.
Section 4 presents
 empirical evidence for the importance of feature selection in text categorization.
 PPM performs badly on wheat and corn because the category grain occurs in
substantial numbers in both the positive and the negative training data for
these two categories. The
 effect can be quantified by computing the entropy of the distributionof
grain articles among the positive and negative trainingarticles. The same
entropy figure can be computed for all
 other categories. The sum of these entropies, weighted according to the
prevalence of the
 correspondingcategory in the training data, represents a coarse measure
of the \\overlap,"
 or similarity, between the positive and negative training data for a category.
3
 The Overlap column of Table 3 shows this measure. It correlates well with
the perfor- mance difference between the support vector method and PPM. The
only exception is the category money market , and we conjecture from Naive
Bayes's poor performance on it that
 this category isan outlier that occurs because it is poorly suited to the
word-based ap-
 proach. Excluding money market , the correlation coefficient for the difference
in breakeven
 performance between LSVM and PPM on the one side, and the entropy measure
on the
 other,is 0.71, and the correlation is statistically significant with a p-value
of 0.03.
 Table 4 summarizes some of the errors made by PPM on the test data. It shows
how the
 false positives associated with the category corresponding to a row are
distributedamong the categories correspondingto the columns. Most false positives
occur when articles belong
 to related categories. This is particularly striking for wheat and corn
: the first row shows
 that 29 articles belonging to corn are incorrectly identified as wheat;
thelast row shows that 13 wheat articles are incorrectly assigned to corn
. Most false positives for the wheat and corn models belong to the category
grain , which comprises wheat , corn , andseveral
 smaller categories {oat , rice , etc.} This adds further support to the
argument that PPM
 performs poorly with overlapping categories. A similar \\false negative"
confusionmatrix
 confirms that several articles belonging to both wheat and corn are not
identified as wheat
 {5 out of 15 false negatives}.3 When calculating the entropy, we divide
the weight of an article by the number of categories it belongs to, giving
every article a weight of 1 in the final sum.
 83.3.3 Modific ations The results in Table 3 were obtained quickly, and
we found them encouraging. Subsequently
 we made many attempts to improve them, all of which met with failure.
 In order to force PPM to buildmodels that are more likely to discriminate
successfully
 between similarcategories, weexperimented with a more costly approach. Instead
of building one positive and one negative model, we built one positive model
and 117 negative
 ones for each of the 118 categories. For each negative model we only used
articles belonging
 to the corresponding category that did not occur in the set of positive
articles. During
 classification, an article was assigned to a category if the positive model
compressed it more
 than all negative models did. Results were improved slightly for categories
like wheat and corn . However, the support vector method still performed
far better. Moreover, compared to the standard PPM method, performance deteriorated
on some other categories.
 We also experimented with the following modifications of the standard procedure,
none of which produced any significant improvement over the results reported
above:
 
*  not rebuilding the two models from the full training data;
 
*  using the same number of stories for buildingM
 P
 and M
 N {usually there are far more
 stories available for building M
 N
 };
 
*  priming the models with fresh Reuters data from outside the training and
test sets; 
*  priming the models with the full training data {positive and negative
articles};
 
*  artificially increasing the counts for the priming data compared with
those for the
 training data and vice versa ;
 
*  using only a quarter of the original training data for validation;
 
*  using escape method A instead of C; 
*  using a word model of order 0, escaping to a character model of order
2 for unseen
 words.
 4 The import ance of feature selection
 In order to test the importance of feature selection we performed experiments
with the
 Naive Bayes learningscheme, varying the number of features it had access
to. We employed the standard word-based approach where the occurrence of
a particular word is treated as a binary feature. Before the input text was
split into words, we removed all non-letter
 characters and stop words. We didnot perform stemming. Naive Bayes does
not incorporate any mechanism for feature selection. Before it is
 applied, features must be pre-selected according to their influence on category
membership.
 Following Dumais et al. {1998}, weused a different set of features for each
category, choosing
 the k features that had the greatest mutual information with the category.
The rightmost block of Table 3 shows the breakeven performance for the ten
largest
 categories for three different numbers of features: k =5, k =50 and k =300.
It also
 includesthe results obtained by Dumais et al. {1998} using fifty features.
Although our
 results are similar overall, they differ slightly for some categories, possibly
because we did
 not perform stemming. The results show that the optimum number of features
varies significantly among cat-
 egories. For several categories {earnings , corp or ate acquisitions , crude
oil and shipping } a
 large number of features is best. However, for grain , wheat and corn performance
peaks with only five features. Moreover, forwheat and corn the breakeven
point increases dra- matically when five features are used instead of fifty|and
in fact for wheat it increases
 9further when just one feature is used. This is consistent with the conjecture
above that
 often the occurrence of just a few words is sufficient to predict category
membership.
 5 Conclusions Compared to state-of-the-art machinelearning techniques for
categorizing Englishtext, PPM produces inferior results because it is insensitive
to subtle differences between articles
 that belong to a category and those that do not. We do not believe our results
are specific to
 PPM. If the occurrence of a single word determines whether an article belongs
to a category
 or not, any compression scheme will likely fail to classify the article
correctly. Machine
 learning schemes fare better because they automatically eliminate irrelevant
features.
 Compared to word-based approaches, compression-based methods avoid ad hoc
deci-
 sions when preparing the text for the actual learning task. Morever, compression-based
 methods apply immediately to the categorization of arbitrary documents,
not just English
 text. However, itis hard to see how efficient feature selection could be
incorporated into
 PPM. Hence it seems appropriate to abandon this method and to move to a
classical ma-
 chine learning setting where, instead of using words, each n -gram is treated
as a separate feature for the learning algorithm. Anecdotal evidence indicates
that the idea of using compression to classify documents is one that has
been reinvented many times. One of us {IHW} investigated its use for
 document classification in the mid-1980s. We know of few records of such
investigations,
 although Teahan {1998} concludes, based on some experiments, that compression
methods
 are capable of ascribing authorship and identifying language dialects. We
are less sanguine, and tend to believe that compression-based methods will
not compete with other, state of the art, methods for such problems. Given
our interest in the use of compression for text
 mining{Witten et al., 1999}, we would like to be proved wrong.
 References
 Bell,T.C., Cleary, J.G. and Witten, I.H. {1990} Text compression . Prentice
Hall, Englewood Cliffs, NJ.
 Dumais, S., Platt, J., Heckerman, D. and Sahami, M. {1998} \\Inductive learning
algorithms and represen-
 tations for text categorization." Proc Int Conf on Info and Knowledge Management
, pp. 148{155.
 Langley, P., Iba, W. and Thompson, K. {1992} \\An analysis of Bayesian classifiers.
In Proc National Con-
 ference on Artificial Intelligence , pp. 223{228. Lewis, D.D. and Ringuette,
M. {1994} \\A comparison of two learning algorithms for text categorization."
 Proc Annual Symposium on Document Analysis and Information Retrieval , pp.
37{50.
 Ng, H.T., Goh, W.B. and Low, K.L. {1997} \\Feature selection, perceptron
learning, and a usability case
 study for text categorization." Proc ACM SIGIR, pp. 67{73. Rocchio, J.J.
Jr. {1971} \\Relevance feedback in information retrieval." In The SMART Retrieval
System: Experiments in automatic document processing , edited by G. Salton,
pp. 313{323. Prentice Hall.
 Sahami, M. {1996} \\Learning limited dependence Bayesian classifiers." Proc
Int Conf on Knowledge Dis-
 covery and Data Mining , pp. 335{338. AAAI Press.
 Salton, G. {1989} Automatic text processing . Addison-Wesley , Reading,
Massachusetts.
 Teahan, W.J. {1998} \\Modelling English text." Ph.D. Thesis, University
of Waikato, New Zealand. Witten, I.H., Bray, Z., Mahoui, M. and Teahan, W.J.
{1999} \\Text mining: a new frontier for lossless com- pression." Proc Data
Compression Conference , pp. 198{207. IEEE Press, Los Alamitos, CA.
 Witten, I.H. and Frank, E. {2000} Data mining: Practical machine learning
tools and techniques with Java
 implementations . Morgan Kaufmann, San Francisco, CA.
 Yang, Y. {1994} \\Expert network: Effective and efficient learning from
human decisions in text categoriza-
 tion and retrieval." Proc ACM SIGIR, pp. 13{22.
 Yang, Y. {1999} \\An evaluation of statistical approaches to text categorization."
Information Retrieval ,
 Vol. 1, No. 1, pp. 69{90.
 Yang, Y. and Pederson, J.O. {1997} \\A comparative study on feature selection
in text categorization." Proc Int Conf on Machine Learning , pp. 412-420.
 10