Practical Issues for Automated Categorization of Web Sites

John M. Pierre
Metacode Technologies, Inc.
139 Townsend Street, Suite 100
San Francisco, CA 94107
jpierre@metacode.com

September 2000

Abstract:

In this paper we discuss several issues related to automated text classification of web sites. We analyze the nature of web content and metadata and requirements for text features. We present an approach for targeted spidering including metadata extraction and opportunistic crawling of specific semantic hyperlinks. We describe a system for automatically classifying web sites into industry categories and present performance results based on different combinations of text features and training data.

1. Introduction

There are an estimated 1 billion pages accessible on the web with 1.5 million pages being added daily. Describing and organizing this vast amount of content is essential for realizing its full potential as an information resource. Accomplishing this in a meaningful way will require consistent use of metadata and other descriptive data structures such as semantic linking[1]. Categorization is an important ingredient as is evident from the popularity of web directories such as Yahoo![2], Looksmart[3], and the Open Directory Project[4]. However these resources have been created by large teams of human editors and represent only one kind of classification task that, while widely useful, can never be suitable to all applications. Automated classification is needed for at least two important reasons. The first is the sheer scale of resources available the web and their ever-changing nature. It is simply not feasible to keep up with the pace of growth and change on the web through manual classification without expending immense time and effort. The second reason is that classification itself is a subjective activity. Different classification tasks are needed for different applications. No single classification scheme is suitable for all applications. In this paper we discuss some practical issues for applying methods of automated text categorization to web content. Rather than a take a one size fits all approach we advocate the use of targeted specific classification tasks, relevant to solving specific problems. In section 2 we discuss the nature of web content and its implications for extracting good text features. We describe a specialized system for classifying web sites into industry categories in section 3, and present the results in section 4. In section 5 we discuss related work. We state our conclusions and make suggestions for further study in section 6.

2. Web Sites

One the main challenges with classifying web pages is the wide variation in their content and quality. Most text categorization methods rely on the existence of good quality texts, especially for training[5]. Unlike many of the well known collections typically studied in automated text classification experiments (i.e. TREC, Reuters-22578, OSHUMED), in comparison the web lacks homogeneity and regularness. To make matters worse, much of the existing web page content is based in images, plug-in applications, or other non-text media. The usage of metadata is inconsistent or non-existent. In this section we survey the landscape of web content, and its relation to the requirements of text categorization systems.

2.1 Analysis of Web Content

In an attempt to characterize the nature of the content to be classified, we performed a rudimentary quantitative analysis. Our results were obtained by analyzing a collection of 29,998 web domains obtained from a random dump of the database of a well-known domain name registration company. Of course these results reflect the biases of our small samples and don't necessarily generalize to the web as a whole, however they should be reflective of the issues at hand. Since our classification method is text based, it is important to know the amount and quality of the text based features that typically appear in web sites. In Table 1 we show the percentage of web sites with a certain number of words for each type of metatag. We analyzed a sample of 19195 domains with live web sites and counted the number of words used in the content attribute of the <META name=``keywords''> and <META name=``description''> tags as well as <TITLE> tags. We also counted free text found within the <BODY> tag, excluding all other HTML tags.

**Table 1:** Percentage of Web Pages with Words in HTML Tags
Tag Type	0 words	1-10 words	11-50 words	51+ words
Title	4%	89%	6%	1%
Meta-Description	68%	8%	21%	3%
Meta-Keywords	66%	5%	19%	10%
Body Text	17%	5%	21%	57%

The most obvious source of text is within the body of the web page. We noticed that about 17% of top level web pages had no usable body text. These cases include pages that only contain frame sets, images, or plug-ins (our user agent followed redirects whenever possible). About a quarter of web pages contained 1-50 words, and the majority of web pages contained over 50 words. Other sources of text are the content in HTML tags including titles, metatags, and hyperlinks. One of the more promising sources of text features should be found in web page metadata. Though title tags are common the amount of text is relatively small with 89% of the titles containing only 1-10 words. Also, the titles often contain only names or terms such as ``home page'', which are not particularly helpful for subject classification. Metatags for keywords and descriptions are used by several major search engines, where they play an important role in the ranking and display of search results. Despite this, only about a third of web sites were found to contain these tags. As it turns out, metatags can be useful when they exist because they contain text specifically intended to aid in the identification of a web site's subject areas¹. Most of the time these metatags contained between 11 and 50 words, with a smaller percentage containing more than 50 words (in contrast to the number of words in the body text which tended to contain more than 50 words).

2.2 Good Text Features

In reference[5] it is argued that for the purposes of automated text classification text features should be:

Relatively few in number
Moderate in frequency of assignment
Low in redundancy
Low in noise
Related in semantic scope to the classes to be assigned
Relatively unambiguous in meaning

Due to the wide variety of purpose and scope of current web content, items 4 and 5 are difficult requirements to meet for most classification tasks. For subject classification, metatags seem to meet those requirements better than other sources of text such as titles and body text. However the lack of widespread usage of metatags is a problem if coverage of the majority of web content is desired. In the long term, automated categorization could really benefit if greater attention is paid to the creation and usage of rich metadata, especially if the above requirements are taken into consideration. In the short term, one must implement a strategy for obtaining good text features from the existing HTML and natural language cues that takes the above requirements as well as the goals of the classification task into consideration.

3. Experimental Setup

The goal of our project was to rapidly classify domain names (web sites) into broad industry categories. In this section we describe the main ingredients of our classification experiments including the data, architecture, and evaluation measures.

3.1 Classification Scheme

The categorization scheme used was the top level of the 1997 North American Industrial Classification Scheme (NAICS) [6], which consists of 21 broad industry categories shown in Table 2.

**Table 2:** Top level NAICS Categories
NAICS code	NAICS Description
11	Agriculture, Forestry, Fishing, and Hunting
21	Mining
22	Utilities
23	Construction
31-33	Manufacturing
42	Wholesale Trade
44-45	Retail Trade
48-49	Transportation and Warehousing
51	Information
52	Finance and Insurance
53	Real Estate and Rental and Leasing
54	Professional, Scientific and Technical Services
55	Management of Companies and Enterprises
56	Administrative and Support, Waste Management and Remediation Services
61	Educational Services
62	Health Care and Social Assistance
71	Arts, Entertainment and Recreation
72	Accommodation and Food Services
81	Other Services (except Public Administration)
92	Public Administration
99	Unclassified Establishments

Some of our resources had been previously classified using the older 1987 Standard Industrial Classification (SIC) system. In these cases we used the published mappings[6] to convert all assigned SIC categories to their NAICS equivalents. All lower level NAICS subcategories were generalized up to the appropriate top level category.

3.2 Targeted Spidering

Based on the results of section 2, it is obvious that selection of adequate text features is an important issue and certainly not to be taken for granted. To balance the needs of our text-based classifier against the speed and storage limitations of a large-scale crawling effort, we took an approach for spidering web sites and gathering text that was targeted to the classification task at hand. Our opportunistic spider begins at the top level web page and attempts to extract useful text from metatags and titles if they exist, and then follows links for frame sets if they exist. It also follows any links that contain key substrings such as prod, services, about, info, press, and news, and again looks for metatag content. These substrings were chosen based on an ad hoc frequency analysis and the assumption that they tend to point to content that is useful for deducing an industry classification. Only if no metatag content is found does the spider gather actual body text of the web page. For efficiency we ran several spiders in parallel, each working on different lists of individual domain names. What we were attempting here was to take advantage of the current web's implicit semantic structure. One the advantages of moving towards an explicit semantic structure for hypertext documents[1] is that an opportunistic spidering approach could really benefit from a formalized description of the semantic relationships between linked web pages. In some preliminary tests we found the best classifier accuracy was obtained by using only the contents of the keywords and description metatags as the source of text features. Adding body text decreased classification accuracy. However, due to the lack of widespread usage of metatags limiting ourselves to these features was not practical, and other sources of text such as titles and body text were needed to provide adequate coverage of web sites. Our targeted spidering approach attempts to gather the higher quality text features from metatags and only resorts to lower quality texts if needed.

3.3 Test Data

>From our initial list of 29,998 domain names we used our targeted spider to determine which sites were live and obtained extracted text using the approach outlined in section 3.2. Of those, 13,557 domain names had usable text content and were pre-classified according to industry category².

3.4 Training Data

We took two approaches to constructing training sets for our classifiers. In the first approach we used a combination of 426 NAICS category labels (including subcategories) and 1504 U.S. Securities and Exchange Commission (SEC) 10-K filings³ for public companies[7] as training examples. In the second approach we used a set of 3618 pre-classified domain names along with text for each domain obtained using our spider. The first approach can be considered as using ``prior knowledge'' obtained in a different domain. It is interesting to see how knowledge from a different domain generalizes to our problem. Furthermore it is often the case that training examples can be difficult to obtain (thus the need for an automated solution in the first place). The second approach is the more conventional classification by example. In our case it was made possible by the fact that our database of domain names was pre-classified according one or more industry categories.

3.5 Classifier Architecture

Our text classifier consisted of three modules: the targeted spider for extracting text features associated with a web site, an information retrieval engine for comparing queries to training examples, and a decision algorithm for assigning categories. Our spider was designed to quickly process a large database of top level web domain names (e.g. domain.com, domain.net, etc.). As described in section 3.2 we implemented an opportunistic spider targeted to finding high quality text from pages that described the business area, products, or services of a commercial web site. After accumulating text features, a query was submitted to the text classifier. The domain name and any automatically assigned categories were logged in a central database. Several spiders could be run in parallel for efficient use of system resources. Our information retrieval engine was based on Latent Semantic Indexing (LSI)[8]. LSI is a variation of the vector space model of information retrieval that uses the technique of singular value decomposition (SVD) to reduce the dimensionality of the vector space. In a previous work[7] it was shown that LSI provided better accuracy with fewer training set documents per category than standard TF-IDF weighting. Queries were compared to training set documents based on their cosine similarity, and a ranked list of matching documents and scores was forwarded to the decision module. In the decision module, we used a k-nearest neighbor algorithm for ranking categories and assigned the top ranking category to the web site. This type of classifier tends to perform well compared to other methods[11], is robust, and tolerant of noisy data (all are important qualities when dealing with web content).

3.6 Evaluation Measures

System evaluation was carried out using the standard precision, recall, and F1 measures[9,10]. The F1 measure combines precision and recall with equal importance into a single parameter for optimization and is defined as

		2 P R
F1	=
		P + R

where P is precision and R is recall. We computed global estimates of performance using both micro-averaging (results are computed based on global sums over all decisions) and macro-averaging (results are computed on a per-category basis, then averaged over categories). Micro-averaged scores tend to be dominated by the most commonly used categories, while macro-averaged scores tend to be dominated by the performance in rarely used categories. This distinction was relevant to our problem, because it turned out that the vast majority of commercial web sites are associated with the Manufacturing (31-33) category.

4. Results

In our first experiment we varied the sources of text features for 1125 pre-classified web domains. We constructed separate test sets based on text extracted from the body text, metatags (keywords and descriptions), and a combination of both. The training set consisted of SEC documents and NAICS category descriptions. Results are shown in Table 3.

**Table 3:** Performance vs. Text Features
Sources of Text	micro P	micro R	micro F1
Body	0.47	0.34	0.39
Body + Metatags	0.55	0.34	0.42
Metatags	0.64	0.39	0.48

Using metatags as the only source of text features resulted in the most accurate classifications. Precision decreases noticeably when only the body text is used. It is interesting that including the body text along with the metatags also results in less accurate classifications. The usefulness of metadata as a source of high quality text features should not be surprising since it meets most of the criteria listed in 2.2. In our second experiment we compared classifiers constructed from the two different training sets described in section 3.4. The results are shown in Table 4.

**Table 4:** Performance vs. Training Set
Classifier	micro P	micro R	micro F1	macro P	macro R	macro F1
SEC-NAICS	0.66	0.35	0.45	0.23	0.18	0.09
Web Pages	0.71	0.75	0.73	0.70	0.37	0.40

The SEC-NAICS training set achieved respectable micro-averaged scores, but the macro-averaged scores were low. One reason for this is that this classifier generalizes well in categories that are common to the business and web domains (31-33, 23, 51), but has trouble with recall in categories that are not well represented in the business domain (71, 92) and poor precision in categories that are not as common in the web domain (54, 52, 56). The training set constructed from web site text performed better overall. Macro-averaged recall was much lower than micro-averaged recall. This can be partially explained by the following example. The categories Wholesale Trade (42) and Retail Trade (44-45) have a subtle difference especially when it comes to web page text which tends to focus on products and services delivered rather than the Retail vs. Wholesale distinction. In our training set, category 42 was much more common than 44-45, and the former tended to be assigned in place of the latter, resulting in low recall for 44-45. Other rare categories also tended to have low recall (e.g. 23, 56, 81).

5. Related Work

Some automatically constructed, large-scale web directories have been deployed as commercial services such as Northern Light[12], Inktomi Directory Engine[13], Thunderstone Web Site Catalog[14]. Details about these systems are generally unavailable because of their proprietary nature. It is interesting that these directories tend not to be as popular as their manually constructed counterparts. A system for automated discovery and classification of domain specific web resources is described as part of the DESIRE II project[15][16]. Their classification algorithm weights terms from metatags higher than titles and headings, which are weighted higher than plain body text. They also describe the use of classification software as a topic filter for harvesting a subject specific web index. Another system, Pharos (part of the Alexandria Digital Library Project), is a scalable architecture for searching heterogeneous information sources that leverages the use of metadata[17] and automated classification[18]. The hyperlink structure of the web can be exploited for automated classification by using the anchor text and other context from linking documents as a source of text features[19]. Approaches to efficient web spidering[20][21] have been investigated and are especially important for very large-scale crawling efforts. A complete system for automatically building searchable databases of domain specific web resources using a combination of techniques such as automated classification, targeted spidering, and information extraction is described in reference[22].

6. Conclusions

Automated methods of knowledge discovery, including classification, will be important for establishing the semantic web. Classification is not objective. A single classification can never be adequate for all the possible applications. A specialized approach including pragmatic, targeted techniques can be applied to specific classification tasks. In this paper we described a practical system for classifying domain names into industry categories that gives good results. From the results in Table 3 we concluded that metatags were the best source of quality text features, at least compared to the body text. However by limiting ourselves to metatags we would not be able to classify the large majority web sites. Therefore we opted for a targeted spider that extracted metatag text first, looked for pages that described business activities, and then degraded to other text only if necessary. It seems clear that text contained in structured metadata fields results in better automated categorization. If the web moves toward a more formal semantic structure as outlined by Tim Berners-Lee[1], then automated methods can benefit. If more and different kinds of automated classification tasks can be accomplished more accurately, the web can be made to be more useful as well as more usable. We outline a basic approach for building a targeted automated web categorization solution:

Knowledge Gathering - It is important to have a clear understanding of the domain to be classified and the quality of the content involved. The web is a heterogeneous environment, but within given domains patterns and commonalties can emerge. Taking advantage of specialized knowledge can improve classification results.
Targeted Spidering - For each classification task different features will be important. However, due to the lack of homogeneity in web content, the existence of key features can be quite inconsistent. A targeted spidering approach tries to gather as many key features as possible with as little effort as possible. In the future this type of approach can benefit greatly from a web structure that encourages the use of metadata and semantically-typed links.
Training - The best training data comes from the domain to be classified, since that gives the best chance for identifying the key features. In cases where it's not feasible to assemble enough training data in the target domain, it may be possible to achieve acceptable results using training data gathered from a different domain. This can be true for web content which can be unstructured, uncontrolled, immense, and hence difficult to assemble quality training data. However, controlled collections of pre-classfied electronic documents can be obtained in many important domains (financial, legal, medical, etc.) and applied to automated categorization of web content.
Classification - In addition to being as accurate as possible, the classification method needs to be efficient, scalable, robust, and tolerant of noisy data. Classification algorithms that utilize the link structure of the web, including formalized semantic linking structures should be further investigated.

Better acceptance of metadata is one key to the future of the semantic web. However, creation of quality metadata is tedious and is itself a prime candidate for automated methods. A preliminary method such as the one outlined in the paper can serve as the basis for bootstrapping[23] a more sophisticated classifier that takes full advantage of the semantic web, and so on.

Acknowledgments

I would like to thank Roger Avedon, Mark Butler, and Ron Daniel for collaboration on the design of the system, and Bill Wohler for collaboration on system design and software implementation. Special thanks to Network Solutions for providing classified domain names.

Bibliography

1: T. Berners-Lee. Semantic Web Road Map.
2: Yahoo!, http://www.yahoo.com/
3: Looksmart, http://www.looksmart.com/
4: Open Directory Project, http://www.dmoz.org/
5: D. Lewis. Text Representation for Intelligent Text Retrieval: A Classification-Oriented View. In P. Jacobs, editor, Text-Based Intelligent Systems, Chapter 9. Lawrence Erlbaum, 1992.
6: North American Industrial Classification System (NAICS) - United States, 1997.
7: R. Dolin, J. Pierre, M. Butler, and R. Avedon. Practical Evaluation of IR within Automated Classification Systems. Eighth International Conference of Information and Knowledge Management, 1999.
8: S. Deerwester, S. Dumais, G. Furnas, T. Landauer, and R. Harshman. Indexing by latent semantic analysis. Journal of the American Society for Information Science, 41 (6):391-407, 1990.
9: C.J. van Rijsbergen. Information Retrieval. Butterworths, London, 1979.
10: D. Lewis. Evaluating Text Categorization. In Proceedings of the Speech and Natural Language Workshop, 312-318, Morgan Kaufmann 1991.
11: Y. Yang and X. Liu. A re-examination of text categorization methods. In Proceedings of the 22nd Annual ACM SIGIR Conference on Research and Development in Information Retrieval, 42-49, 1999.
12: Northern Light, http://www.northernlight.com/
13: Inktomi Directory Engine,
14: Thunderstone Web Site Catalog,
15: A. Ardo, T. Koch, and L. Nooden. The construction of a robot-generated subject index. EU Project DESIRE II D3.6a, Working Paper 1 1999.
16: T. Kock and A. Ardo. Automatic classification of full-text HTML-documents from one specific subject area. EU Project DESIRE II D3.6a, Working Paper 2 2000.
17: R. Dolin, D. Agrawal, L. Dillon, and A. El Abbadi. Pharos: A Scalable Distributed Architecture for Locating Heterogeneous Information Sources Version. In In Proceedings of the 6th International Conference on Information and Knowledge Management, 1997.
18: R. Dolin, D. Agrawal, A. El Abbadi, and J. Pearlman. Using Automated Classification for Summarizing and Selecting Heterogeneous Information Sources. In D-Lib Magazine, January, 1998.
19: G. Attardi, A. Gulli, and F. Sebastiani. Automatic Web Page Categorization by Link and Context Analysis. In Chris Hutchison and Gaetano Lanzarone (eds.), Proceedings of THAI'99, European Symposium on Telematics, Hypermedia and Artificial Intelligence, 105-119, 1999.
20: J. Cho, H. Garcia-Molina, and L. Page. Efficient crawling through URL ordering. In Computer Networks and ISDN Systems (WWW7), Vol. 30, 1998.
21: J. Rennie and A. McCallum. Using Reinforcement Learning to Spider the Web Efficiently. Proceedings of the Sixteenth International Conference on Machine Learning, 1999.
22: A. McCallum, K. Nigam, J. Rennie, and K. Seymore. A Machine Learning Approach to Building Domain-Specific Search Engines. The Sixteenth International Joint Conference on Artificial Intelligence, 1999.
23: R. Jones, A. McCallum, K. Nigam, and E. Riloff. Bootstrapping for Text Learning Tasks. In IJCAI-99 Workshop on Text Mining: Foundations, Techniques and Applications, 52-63, 1999.

Footnotes

... areas¹: The possibilities for misuse/abuse of these tags to improve search engine rankings are well known; however, we found these practices to be not very widespread in our sample and of little consequence.
... category²: Industry classifications were provided by InfoUSA and Dunn & Bradstreet.
... filings³: SEC 10-K filings are annual reports required of all U.S. public companies that describe business activities for the year. Each public company is also assigned an SIC category.

2000-08-30