Content uploaded by Michel Klein
Author content
All content in this area was uploaded by Michel Klein on Nov 23, 2013
Content may be subject to copyright.
GoldenBullet: Automated Classification of Product Data in E-commerce
Y. Ding, M. Korotkiy, B. Omelayenko, V. Kartseva, V. Zykov, M. Klein, E. Schulten,
and D. Fensel
Vrije Universiteit Amsterdam, De Boelelaan 1081a, 1081 HV Amsterdam, NL, ying@cs.vu.nl
Withold Abramowicz (ed.), Business Information Systems, Proceedings of BIS 2002, Poznan, Poland
Abstract
Internet and Web technology starts to penetrate many
aspects of our daily life. Its importance as a medium for
business transactions will grow exponentially during the
next years. In terms of the involved market volume the
B2B area will hereby be the most interesting area. Also it
will be the place, where the new technology will lead to
drastic changes in established customer relationships and
business models. B2B market places provide new kinds of
services to their clients. Simple 1-1 connections are
getting replaced by n-m relationships between customers
and vendors. However, this new flexibility in electronic
trading also generates serious challenges for the parties
that want to realize it. The main problem here is caused by
the heterogeneity of information descriptions used by
vendors and customers. Intelligent solutions that help to
mechanize the process of structuring, classifying,
aligning, and personalizing are a key requisite for
successfully overcoming the current bottlenecks of B2B
electronic commerce. In this paper, we describe a system
called GoldenBullet that applies techniques from
information retrieval and machine learning to the
problem of product data classification. The system helps
to mechanize an important and labor-intensive task of
content management for B2B Ecommerce.
1. Introduction
The World Wide Web (WWW) has drastically changed
the on-line availability of information and the amount of
electronically exchanged information. Meanwhile the
computer has mutated from a device for computation into
a entrance portal of large information volumes,
communication, and business transactions (cf. [Fensel,
2001]). It starts to change the commercial relationships
between suppliers and customers. Currently, a large
fraction of the B2B transactions are still realized by
traditional non-Internet networks, such as those conducted
over EDI systems. In this traditional paradigm, direct 1-1
connections and mappings are programmed based on
standards like EDIFACT (cf. [EDIFACT, 1999]).
However, this traditional paradigm does not at all employ
the full power of electronic commerce and it is quite likely
that it will soon be out-ranged by more timely, Internet
and web-based transaction types. Internet-based electronic
commerce provides a much higher level of flexibility and
openness that will help to optimize business relationships.
Instead of implementing one link to each supplier, a
supplier is linked to a large number of potential customers
when linked to the market place.
However, preventing their customers from the
bottleneck of facing exponential growth in the number of
implemented business connections faces B2B market
places with a serious problem. The have to deal with the
problem of heterogeneity in product, catalogue, and
document description standards of their customers.
Effective and efficient management of different
description styles become a key task for these market
places.
Successful content management for B2B electronic
commerce has to deal with various aspects: information
extraction from rough sources, information classification
to make product data maintainable and accessible,
reclassification of product data, information
personalization, and mappings between different
information presentations [Fensel et al., 2001]. All of
these sub-tasks are hampered by the lack of proper
standards (or in other words by the inflation and non-
consistency of arising pseudo-standards). The paper will
focus on these challenges for content management and
will discuss some potential solution paths.
The contents of the paper is organized as follows. In
Section 2 we describe the overall content management
problem that needs to be solved for effective E-commerce.
Section 3 introduces our system GoldenBullet that applies
information retrieval and machine learning techniques to
one of the important sub-tasks of content management.
GoldenBullet helps to mechanize the process of product
classification. Section 4 provides an evaluation of our
approach based on real-world data provided by B2B
market places. Finally Section 5 provides conclusions and
Withold Abramowicz (ed.), Business Information Systems, Proceedings of BIS 2002, Poznan, Poland
2 BUSINESS INFORMATION SYSTEMS - BIS 2002
discusses future directions.
2. Content Management in E-Commerce
B2B market places are an intermediate layer for
business communications providing one serious
advantages to their clients. They can communicate with a
large number of customers based on one communication
channel to the market place. The market places reduce the
number of mappings to their user community from n*m to
n+m. However, in order to provide this service, they have
to solve the significant mapping and normalization
problem for their clients. A successful market place has to
deal with various aspects. It has to integrate with various
hardware and software platforms and has to provide a
common protocol for information exchange. However, the
real problem is the heterogeneity and openness of the
exchanged content. Therefore, content management is one
of the real challenges in successful B2B electronic
commerce. It tackles with a number of serious problems
[Fensel et al., 2001]:
1 Product descriptions are unstructured.
2 Product descriptions are unclassified.
3 Product descriptions must be classified and described
in various dimensions because no standard product
classifications exist.
Product descriptions must be structured. Suppliers
have product catalogues that describe their products to
their potential clients. This information should be made
on-line available by a B2B market place. One could think
that this may be a simple task because most product
catalogues already exist electronically. However, these
product catalogues are designed for the human reader.
Extracting the actual product information and storing it in
a structured format is therefore mainly a manual task. A
content management solution provider like Content
Europe
1
has several hundred employees working in
content factories to manually structure the product
information. In the worst case, they take printed copies of
the product catalogues as input.
Product descriptions must be classified. At this stage
in the content management process we can assume that
our product information is structured in a tabular way.
Each product corresponds to an entry in a table where the
columns reflect the different attributes of a product.
Similar products are group together in the same table.
1.
http://www.contenteurope.com
Each supplier uses different structures and vocabularies to
describe its products. This may not cause a problem for a
1-1 relationship where the buyer may get used to the
private terminology of his supplier. B2B market places
that enable n-m commerce cannot rely on such an
assumption. They must classify all products according to
a standard classification schema that help buyers and
suppliers in communicating their product information. A
widely used classification schema in the US is UNSPSC
2
(for details about UNSPSC, please see next section).
Again it is a difficult and mainly manual task to classify
the products according to a classification schema like
UNSPSC. It requires domain expertise and knowledge
about the product domain.
Product descriptions must be re-classified.
Bottlenecks in exchanging information have led to a
plethora of different standards that should improve the
situation. However, usually there are two problems. First,
there are too many “standards”, i.e., none of them is an
actual standard. Second, mostly, standards lack important
features for various application problems. Not
surprisingly, both problems appear also in B2B electronic
commerce. UNSPSC is a typical example for a horizontal
standard that covers all possible product domain,
however, is not very detailed in any domain. Another
example for such a standard is the Universal Content
Extended Classification (UCEC)
3
. It takes UNSPSC as a
starting point and refines it by attributes. Rosetta Net
4
is
an example for a vertical standard describing products of
the hardware and software industry in detail. Vertical
standards describe a certain product domain in more
detail than common horizontal ones. More examples for
such “standards” can be found in [Fensel, 2001].
In the reminder of the paper we focus on one of these
sub-tasks. We will describe our solution we developed for
product classification. However, we would also like to
mention that we are currently evaluating similar
techniques for product data structuring and re-
classification.
3. GoldenBullet
Finding the right place for a product description in a
standard classification system such as UNSPSC is not at
all a trivial task. Each product must be mapped to the
2.
http://www.un-spsc.net and http://www.unspsc.org.
3.
http://www.ucec.org
4.
http://www.rosettanet.org/
Withold Abramowicz (ed.), Business Information Systems, Proceedings of BIS 2002, Poznan, Poland
GoldenBullet: Automated Classification of Product Data in E-commerce 3
corresponding product category in UNSPSC to create the
product catalog. Product classification schemes contain
huge number of categories with far from sufficient
definitions (e.g. over 12,000 classes for UNSPSC) and
millions of products must be classified according to them.
This requires tremendous labor effort and the product
classification stage takes altogether up to 25% of the time
spent for content management. Because product
classification is that expensive, complicated, time-
consuming and error-prone. Content Management needs
support in automation of the product classification process
and automatic creation of product classification rules.
GoldenBullet is a software environment targeted to
support product classification according to certain content
standards. It is currently designed to automatically
classify the products, based on their original descriptions
and existent classifications standards (such as UNSPSC).
It integrates different classification algorithms from the
information retrieval and machine learning areas and
some natural language processing techniques to pre-
process data and index UNSPSC.
3.1. UNSPSC
The Universal Standard Products and Services
Classification (UNSPSC) is an open global coding system
that classifies products and services. It was first developed
by Dun & Bradstreet and the United Nations
Development Program. It is now maintained by the
Electronic Commerce Code Management Association
(ECCMA) which is a not-profit membership organization.
The UNSPSC code covers almost any product or service
that can be bought or sold, which includes 12,000 codes
covering 54 industry segments from electronics to
chemical, to medical, to educational services, to
automotive to fabrications, etc. The UNSPSC is heavily
Figure 1. The layered structure of UNSPSC.
XX
Segment
The logical aggregation of families for analytical purpose
XX Family
A commonly recognized group of inter-related commodity categories
XX
Class
A group of commodities sharing a common use or function
XX Commodity
A group of substitutable products or services
Business Type
The function performed in support of the commodity. This value is
vendor specific and does not affect the base code for each commodity.
deployed around the world in the electronic catalogs,
search engines, procurement application systems and
accounting systems. It is a 10 digit hierarchical code that
consists of 5 levels (see Figure 1).
3.2. Overall GoldenBullet Functionality
GoldenBullet as a software environment provides the
following functions to fully achieve semi-automatic or
automatic product classification: Data input and export
facilities; text processing techniques; classification of
product data; and learning and enrichment of product
classification information (see Figure 2).
3.2.1 Data Input, Output, and Validation
A wrapper factory gathers various wrappers to convert
raw data description from external formats to internal
format, and final results to preferable output format or
Figure 2. Overview on GoldenBullet.
GoldenBullet
Data
Wrapper
- input/output
- manual or
automatic
- data
v
alidation
Text pre-
p
rocessing
- stemming
- tagging
- indexing
clas
sifier
L
ear
ning
- Online
- off-line
Enriching
- UNSPSC
-
Classifica
tion
rule base
Input
data
(raw
product
description
)
UNSP
SC
UNSP
SC
UNSPSC
Input d
(raw
product
description
)
Input
data
(raw
product
description )
Withold Abramowicz (ed.), Business Information Systems, Proceedings of BIS 2002, Poznan, Poland
4 BUSINESS INFORMATION SYSTEMS - BIS 2002
user-designed formats. Besides the automatic importing
and exporting data, GoldenBullet also provides the editor
for manually inputting data, which suits well for small and
medium vendors.
3.2.2 Text pre-processing
The validated product data will be pre-processed before
the classification has been performed. Some of the Natural
Language Processing algorithms have been implemented
into GoldenBullet. The product data will be stemmed
(grouping different words with the same stems) and
tagged (extracting noun-phrases). Furthermore, UNSPSC
is also being pre-processed (stemmed and tagged) to make
sure that noisy words or information have been screened.
A stop word list has been generated, updated and extended
during the whole process. Currently, GoldenBullet can
handle English and French product data.
3.2.3 Product Classification
Figure 3 shows the user interface of the classifier. The
imported UNSPSC is browsable from the screen, which
directs the end user to the right location of UNSPSC. The
classifier classifies the pre-processed product data and
proposes the ranked solutions based on various weighting
algorithms. The end user can pull down the proposed list
Figure 3. A screenshot of GoldenBullet.
and make the final choice. But when he highlights one of
the proposed solutions, the above UNSPSC browse
window will show the exact location of it in UNSPSC
with the details of each level.
Performing the classification task is viewed as an
information retrieval problem (see [Ribiero-Neto &
Baeza-Yates, 1999] for an introduction to the field). The
problem of finding the right class is viewed as the
problem to find the right document as an answer to a
query:
• A product description is viewed as a query and
UNSPSC is viewed as a document collection.
• Each of the commodities in UNSPSC is treated as a
document, where each commodity description forms
the text of the document.
• Assigning a proper category for a product is achieved
via retrieving a correspondent UNSPSC commodity
description.
The performance of such an approach is rather low
(see the next sub-section for more details). Directly using
UNSPSC as document collection fails in this respect
because the class descriptions are very short (i.e., we deal
Withold Abramowicz (ed.), Business Information Systems, Proceedings of BIS 2002, Poznan, Poland
GoldenBullet: Automated Classification of Product Data in E-commerce 5
with very short documents) and the product descriptions
are often very short too and use very specific vocabulary
that cannot directly be matched with more generic terms
in UNSPSC. Evaluation with real world data showed that
less than 1% of actual product data could classified
correctly which such a naïve approach. Therefore, we
employed various strategies to achieve a more reasonable
and workable result that are described in the next
subsection. Basically we employed different retrieval
strategies and we made use of large volumes of manually
classified data to improve the performance.
3.3. The heart of GoldenBullet: Intelligence and
Knowledge
The essence of GoldenBullet is its ability to
automatically classify product descriptions. This sub-
section will present our classification approaches and
show how they can make use of pre-classified data as
background knowledge. An evaluation of the approaches
in given in section 4.
3.3.1 The Vector Space Model (VSM)
A standard method in Information Retrieval is the well-
known Vector space model (VSM). Salton's Vector Space
Model (cf. [Salton et al., 1975]). It uses the word vector to
represent document and user query, then applies the
cosine similarity formula to calculate the similarity
between the document and query so as to retrieve the most
relevant document to user's query. The same model has
been applied in text categorization. [Gomez-Hidalgo &
Rodriguez, 1997] used Salton's vector space model to
represent document (in our case product description) and
existing categories (e.g. in our case UNSPSC). Then the
category (UNSPSC) can be assigned to a document
(product) when the cosine similarity between them
exceeds a certain threshold. The basic idea is to represent
each document as a vector of certain weighted word
frequencies.
VSM is adopted by us to find the match between
UNSPSC commodities and product descriptions. In the
following, we will describe two strategies how VSM can
make use of pre-classified examples in two ways. Both
strategies treat an unclassified product description as a
query, however, differ in what they us as a document
collection.
1 The first version takes each commodity as a
document. The examples are used to enrich the
commodity description. Basically we extract words
from pre-classified product data and add them to the
word list describing the commodity.
2 The second version takes each pre-classified product
description as a document. We use VSM to retrieve
the instance fitting best to a newly product
description and infer the UNSPSC code of the latter
from the known UNSPSC code of the former.
We will describe pros and cons of both approaches in
section 4.
3.3.2 K-nearest neighbor
Another instance-based classifier we implemented is
based on the k-Nearest Neighbor method KNN. Again, the
algorithm uses the set of pre-classified examples directly
to classify an example. The algorithm passes the whole
set of training examples and searches for the most similar
one, and then assigns the class to the new example, equal
to the class of the most similar one. KNN is
computationally expensive and requires lots of memory to
operate depending on the number of pre-classified
examples. Again we can distinguish two modes in regard
to whether the algorithm works directly on the pre-
classified product data or on enriched class descriptions.
3.3.3 Naïve-Bayes classifier
The final paradigm we employed is the machine
learning paradigm (see [Mitchell, 1997] for a good
introduction and analysis of the field). This paradigm
assumes existence of a set of (manually) pre-classified
products, which is called a training set, and a set of
product descriptions to be classified by the systems,
which is called a test set. The Naïve-Bayes classifier NB
[Burges, 1998] uses Bayes theorem to predict the
probability of each possible class, given a product
description and the set of training pre-classified examples
as input.
The classifier assigns the commodity, which has the
highest probability of being correct. Naïve-Bayes is a
standard text classification algorithm, with a long
successful application history.
5
3.3.4 Hierarchical Classification
UNSPSC provides a hierarchy of four levels for
classifying products: Segment, Family, Class, and
Commodity. Therefore, it is quite natural to employ a
hierarchical classification approach for our task. In
addition, we made the experience that lots of pre-
classified products we received for our evaluations we
report in the next section are not classified up to the
5.
[Koller & Sahami, 1997]
Withold Abramowicz (ed.), Business Information Systems, Proceedings of BIS 2002, Poznan, Poland
6 BUSINESS INFORMATION SYSTEMS - BIS 2002
Commodity level, but only up to Class or Family levels.
Therefore, we build a hierarchical classifier, which
actually consists of four classifiers, each of which is
working on a correspondent level (see also [Chakrabarti et
al., 1997], [Koller & Sahami, 1997], [Dumais & Chen,
2000], and [Agrawal & Srikant, 2001]).
3.4. Implementation
GoldenBullet is designed to provide widest access to
product description classification service. So, we intended
to make some kind of web service out of our tool. Our
current version of the prototype is oriented on an “html
like” user interface. Our main goal, concerning user
interface design, was to provide fully functional and
convenient for a user interaction environment and, at the
same time, not to put too many requirements on the user–
side software. Currently all what a user needs to use
GoldenBullet prototype is an html browser that supports
JavaScript 1.2.
The web service is provided by means of a client-server
approach. So, the core of our tool is a server-side set of
Java packages that implements the functionality and
generates all interaction pages. The server side module
was implemented as a web application. We use Java
Servlets technology and its extension Java Server Pages to
generate all client side user interface pages. All intelligent
content of our prototype was implemented by means of
Java 1.3. Due to the separation of the training-
classification process in an “off-line” training step
(computationally highly expensive) and an “on-line”
classification step we managed to achieve acceptable
performance.
4. Evaluation
The evaluation we report is based on around 40,000
real product descriptions that were already classified
manually. They come from various vendors and cover a
large number of categories in UNSPSC.
6
During the
following we compare the performance of a number of
algorithmic variants of our classifier.
4.1. Test Data Set
The data provide 41913 manually classified data
described in France. Table 1. summarizes the population
of the UNSPSC categories in the data set according to the
6.
The data were collected based on a cooperation with Hubwoo which is
a MRO market place in France.
UNSPSCv7.2.
The table can be understood as follows: The number of
UNSPSC categories reports the number of Commodities,
Classes, Families, and Segments of UNSPSCv7.2 that are
populated by some instances found in the data set. Second
we analyzed the distribution of the data. We took around
5% of the covered Commodities, Classes, Families, and
Segments and selected the ones with the highest
coverage. It shows that 28247 data of the data set are in
21 commodities, i.e., around 60% of the data are in 21 of
the around 15000 commodities. Similar only in one
Segment are around 25% of all data. In total, we worked
with 41913 product descriptions.
4.2. Accuracy Results for a Naïve Classifier
Up to a large number of the product descriptions in this
test data set are represented by the name of the product
models, such as “proliant”, “pentium”, “presario”,
“carepaq”, “RA3000”, but do not use any of the
functional terms of the product itself. In this case, our
Naïve Classifiers are not capable to secure high accuracy.
In fact, the accuracy is extremely low and reaches
maximally 0,2%. Clearly such a classifier is without any
use and we will describe in the following how training
could make a different story.
4.3. Accuracy Results of the trained Algorithms
For training the algorithms we have chosen the
following approach. A 60% random sample from product
descriptions data set was used as training set, and the rest
40% data – as test set. We repeated the test based on
several random splits of the data set. The results are
reported in Table 2. We applied two quality
measurements:
• The total accuracy asks whether the commodity
recommendation of GoldenBullet with the highest
Table 1. Population of UNSPSC
Category Number of
UNSPSC
categories
5% Number of
instances in the
most frequent
categories
Segments 28 1 15233
Families 56 3 27407
Classes 150 7 24800
Commodities 421 21 28247
Withold Abramowicz (ed.), Business Information Systems, Proceedings of BIS 2002, Poznan, Poland
GoldenBullet: Automated Classification of Product Data in E-commerce 7
rating based on the product description matches the
actual commodity of a product.
•The “First 10 Accuracy” asks whether one of the ten
commodity recommendations of GoldenBullet with
highest ratings based on the product description
matches the actual commodity of a product.
In addition, we distinguished two modes for all
algorithms. Either we treated the pre-classified product
data or the enriched class descriptions (based on the pre-
classified data) as documents that should be retrieved. In
general, the bayesian classifier outperforms all other
classifiers significantly.
7
Working directly with the pre-
classified data works best for all algorithms. Only in
regard to the “First 10 Accuracy” there is no difference
for the bayesian classifier in this respect. In general, an
accuracy between 78% to 88% looks rather convincing
and easily outperforms and qualify equal with the quality
of human classification.
8
We repeated the experiments for hierarchical versions
of the algorithms, i.e., first a classification decision is
taken at the segment level and then recursively on the
families, classes, and commodity level. Against our initial
intuition this lead to significant lowering of the accuracy
of the classification algorithms. Obviously, too many
wrong decision in the early steps of the classification
7.
Compare similar results of [Agrawal & Srikant, 2001].
Table 2. Accuracy of trained (non-hierarchical)
algorithms
Algorithm Total Accuracy First 10 Accuracy
VSM
I
60% 78%
VSM
C
28% 69%
KNN
I
45% 84%
KNN
C
29% 56%
NB
I
78% 88%
NB
C
59% 88%
8.
That is, higher accuracy would just mean over-fitting to human
classification decisions that also have a significant error rate which we
encountered in labour intensive manual checks.
process happens. The results are presented in Table 3.
4.4. Summary
Untrained algorithms fail completely to provide any
support in semi-automatic or automatic product
classification. Trained algorithms can provide significant
support. Up to 90% accuracy can be achieved based on
learning from pre-classified example. As always the
significant assumption of such an approach is the
availability of representative pre-classified examples.
5. Conclusions and Future Works
Market places for B2B electronic commerce have a
large economic potential. They provide openness and
flexibility in establishing commercial relationships for
their clients. In order to provide this service they have to
tackle with serious obstacles. The most prominent one is
concerned with integrating various styles to describe the
content and the structure of the exchanged information.
Product catalogues corresponds to large and complex
domain ontologies and in the case of horizontal standards
to upper-layer ontologies. Large modeling and mapping
effort is required to handle these descriptions. Content
manager have to structure, classify, re-classify, and
personalize large volumes of data to make product
descriptions automatically accessible via B2B market
places.
GoldenBullet aims on mechanizing the classification
process of product data. Accuracy rates between 70% and
98% indicate that this process can be mechanized to a
degree where severe cost reduction can be achieved
which is a pre-requisite for scalable E-commerce. The
success of GoldenBullet is based on the combination of
natural language processing, information retrieval,
machine learning and the use large volumes of manually
classified data. Future versions of the GoldenBullet will
provide more advanced features as explained in the
Table 3. Accuracy of trained (hierarchical) algorithms
Algorithm Total Accuracy First 10 Accuracy
VSM
I
22% 41%
VSM
C
14% 27%
KNN
I
25% 22%
KNN
C
13% 15%
NB
I
38% 42%
NB
C
29% 42%
Withold Abramowicz (ed.), Business Information Systems, Proceedings of BIS 2002, Poznan, Poland
8 BUSINESS INFORMATION SYSTEMS - BIS 2002
following.
Recently, Support Vector Machines have been shown
to be a very useful tool for text categorization (cf.
[Burges, 1998], [Cortes & Vapnik, 1995], and [Joachims,
1998], and [Dumais & Chen, 2000]). We want to explore
the applicability of this algorithmic paradigm for our
specific product classification task.
Multi-linguality (i.e. the product catalog and the
product classification standard are described in different
languages) is a severe requirement for E-commerce in
Europe. Currently, GoldenBullet supports English and
French. An extension to further languages is a pre-
requisite for open and flexible E-commerce.
Multi-standard classification is an important issue for
open and flexible E-commerce (cf. [Agrawal & Srikant,
2001], [Madhaven et al., 2001], and [Schulten et al.,
2001]). Currently, market places establish one “standard”
to make data of their clients accessible. However,
openness and flexibility of internet-enabled trading
demands multi-standard classification. Besides UNSPSC
there are other “standards” like UCEC
9
, ecl@ss
10
, and
RosettaNet
11
that are widely used in certain domain-
specific or geographical areas of E-commerce. Providing
mechanized support in reclassifying product data in an
additional classification schema is a demanding feature
for realizing the full potential of E-commerce.
GoldenBullet applies successfully information
retrieval and machine learning techniques to the problem
of automatically classifying product description.
GoldenBullet will also challenge other existing severe
problems for B2B market places, such as mapping and
reclassifying product descriptions according to different
product code systems
12
and to personalize views on
product data for divergent customers.
References
[Agrawal & Srikant, 2001]
R. Agrawal and R. Srikant: On Integrating Catalogs. In
Proceedings of the Tenth International World Wide Web
Conference (WWW2001), Hong Kong, May 2001.
9.
http://www.ucec.org
10.
http://www.eclass.de
11.
http://www.rossettanet.org
12.
See [Agrawal & Srikant, 2001] and [Navathe et al., 2001].
[Burges, 1998]
C. J. C. Burges: A Tutorial on Support Vector Machines for
Pattern Recognition, Data Mining and Knowledge
Discovery, 2(2):955-974, 1998.
[Chakrabarti et al., 1997]
S. Chakrabarti et al.: Using taxonomy, discriminants, and
signatures for navigating in text databases. In Proceedings
of the 23rd VLDB Conference, Greece, 1997.
[Cheeseman and Stutz, 1995]
P. Cheeseman and J. Stutz: Bayesian classification
(AutoClass): Theory and results. In U. M. Fayyad, G.
Piatetsky-Shapiro, P. Smyth, and R. Uthurusamy (eds.),
Advances in knowledge discovery and data mining, The
AAAI press, Menlo Park, 1995.
[Cortes & Vapnik, 1995]
C. Cortes and V. Vapnik: Support-Vector Networks,
Machine Learning, 20:273-297, 1995.
[Dumais & Chen, 2000]
S. Dumais and H. Chen: Hierarchical Classification of Web
Content. In Proceedings of SIGIR-00, 23rd ACM
International Conference on Research and Development in
Information Retrieval, pp. 256-263, Athens, Greece,
August 2000.
[EDIFACT, 1999]
United Nation: UN/EDIFACT-Directory. http://
www.unece.org/trade/untdid, 1999.
[Fensel, 2001]
D. Fensel: Ontologies: Silver Bullet for Knowledge
Management and Electronic Commerce, Springer-Verlag,
Berlin, 2001.
[Fensel et al., 2001]
D. Fensel, Y. Ding, E. Schulten, B. Omelayenko, G.
Botquin, M. Brown, and A. Flett: Product Data Integration
in B2B E-commerce, IEEE Intelligent System, 16(3), 2001.
[Gomez-Hidalgo & Rodriguez, 1997]
J. M. Gomez-Hidalgo and M. B. Rodriguez: Integrating a
lexical database and a training collection for text
categorization. In the Proceedings of ACL/EACL (the
Association for Computational Linguistics/European
Association for Computational Linguistics: Automatic
Information Extraction and Building of Lexical Semantic
Resources for NLP Applications), Madrid, Spain, July,
1997.
[Joachims, 1998]
T. Joachims: Text Categorization with Support Vector
Machines: Learning with Many Relevant Features. In C.
Nedellec and C. Rouveirol (eds.), Machine Learning:
ECML-98, Proceedings of the 10th European Conference
on Machine Learning, Lecture Notes in Artificial
Intelligence, LNAI 1398, pages 137-142, 1998.
[Koller & Sahami, 1997]
D. Koller and M. Sahami: Hierarchically classifying
documents using very few words. In Proceedings of the
International Conference on Machine Learning, vol 14,
Morgan-Kaufmann, July 1997.
[Madhaven et al., 2001]
Withold Abramowicz (ed.), Business Information Systems, Proceedings of BIS 2002, Poznan, Poland
GoldenBullet: Automated Classification of Product Data in E-commerce 9
J. Madhaven, P. A. Bernstein, and E. Rahm: Generic
Schema Matching with Cupid. In Proceedings of the 27th
VLDB Conference, Roma, Italy, 2001.
[Mitchell, 1997]
T. Mitchell, Machine Learning, McGraw Hill, 1997.
[Navathe et al., 2001]
S. Navathe, H. Thomas, M. Satitsamitpong, and A. Datta: A
Model to Support E-Catalog Integration. In Proceedings of
the 9th IFIP 2.6 Working Conference on Database (DS-9),
Semantic Issues in e-commerce Systems, Hong Kong, April
2001.
[Ribiero-Neto & Baeza-Yates, 1999]
B. Ribiero-Neto and R. Baeza-Yates, Modern Information
Retrieval, Addison Wesley, 1999.
[Salton et al., 1975]
G. Salton, A. Wong, and C. S. Yang: (1975): A vector
space model for automatic indexing, Communications of the
ACM, 18(7):613-620, 1975.
[Schulten et al., 2001]
E. Schulten, H. Akkermans, G. Botquin, M. Dörr, N.
Guarino, N. Lopes, and N. Sadeh: The ecommerce product
classification challenge, IEEE Intelligent systems, 16(4),
2001.
