The rapid growth of the Web has made it impossible to learn various properties of the entire Web directly. Thus, we need to use statistical methods to estimate the properties of the Web. S. Lawrence and Giles proposed two different methods for estimating the number of Web pages [1,2]. The first method used search engines as a random sampling method. The second method used random sampling of Internet Protocol (IP) addresses and Web servers.
We have applied these two methods to Japanese Web pages in the JP domain to see if we can apply these two methods to a subset of Web. The first method gave an estimate of 88 million pages as a lower bound on the size of the Japanese indexable Web; the second method gave an estimate of 17 million pages. These results mean that these methods do not measure exactly the same Web pages. They also suggest that Japanese search engines cover only a part of Japanese indexable Web pages, and that the Web pages in the JP domain are less connected compared with the whole Web.
The rapid growth of the Web has made it impossible to learn various properties of the entire Web directly. Thus we need to use statistical methods to estimate the properties of Web, such as the size of Web (the number of Web pages), the amount of Web (the number of bytes of Web pages), the number of links, the coverage of search engines, the proportion of various data types, and so on.
S. Lawrence and Giles [1,2] proposed two different methods for estimating the number of Web pages. The first method used search engines as a random sampling method. They gave an estimate of 320 million pages as a lower bound on the size of the indexable Web in December 1997 . The second method used random sampling of IP addresses and Web servers and gave an estimate of 800 million pages as the size of the Web in February 1999 .
The purpose of this study is to see if these two methods can be applied to a subset of the Web. As the diversity of Web users grows, we need new approaches to measure the properties of a subset of the Web written in various languages and in many countries. We have applied the two methods to Japanese Web pages in the JP domain. The first method gave an estimate of 88 million pages as a lower bound on the size of the Japanese indexable Web; the second method gave about 28 million pages.
These results suggest the followings.
In section 2, we explain the first method proposed by Lawrence, which uses search engines' coverage to estimate the size of the indexable Web. We then describe how we adapted these methods for Japanese Web pages in the JP domain. After giving an estimate of the number of Japanese indexable Web pages, we describe several statistics, such as the distribution of Japanese Web pages in and outside of the JP domain. In section 3, we explain the second method also proposed by Lawrence, which uses randomly selected IP addresses, and gives an estimate of the number of Web servers and Web pages. In section 4, we discuss the possible reasons for the difference between the two estimates.
Lawrence's first method is to estimate Web pages that contain text indexable by search engines. These pages need to be accessible and not resticted by
Let U denote a set of indexable Web pages, and let the number of Web pages be the size of U. Let A be a set of Web pages that a search engine, SA, has collected and B be a set of Web pages that a search engine, SB,has collected. Let Pr(A) and Pr(B) be the probabilities that a page is collected by search engine, SA and SB, |A| and |B| be the numbers of Web pages collected by search engine, SA and SB. Then we can define the probability Pr(A) as follows.
Pr(A) = |A|/|U|
|U| = 1/Pr(A) |A|
If A and B are independent, then Pr(A) can be obtained from Pr(A B), the probability of the intersection of A and B, i.e., the probability that a page is collected by both SA and SB as follows.
Pr(A)Pr(B) = Pr(A B)
Pr(A) = |A B|/|B|
Thus, |U|, the number of indexable Web pages, can be calculated as follows, if we know |A|, |B|, and |A B|, the size of overlap between A and B (figure 1).
|U| = (|B|/ |A B|) |A|
Figure 1: Computing overall size from overlap size
Figure 2: Estimating overlap size by random sampling
However, it is not feasible to know |A B| unless two search engines have all the Web pages available to the public. Instead, we obtain a random sample A' from A, and B' from B by obtaining search results for the same query set from the search services SA and SB, then we approximate the Pr(A) as follows.
Pr(A) |A' B'| / |B'|
Then the size of U can be estimated from |B'|, |A' B'|, and |A| as follows.
|U| (|B'| / |A' B'| ) |A|
The size of U can also be estimated by
|U| (|A'| / |A' B'| ) |B|
We obtain an estimate of |U| by averaging the results of the above two equations.
Note this method's accuracy depends on the following assumptions:
Lawrence and Giles  analyzed the search results of six major search engines by using the queries obtained form the Web access log of NEC researchers. The 302 queries were chosen that meet the following conditions:
Lawrence and Giles retrieved all the documents in the search result and checked the existence of query terms. They then computed the size of the overlap between the results of the two largest search engines, AltaVista and HotBot, and obtained estimates of 320 million with a 95 percent confidence interval of 34 million.
There are several difficulties in duplicating Lawrence's approach to estimate the number of Web pages that contain Japanese text. One of the difficulties is caused by the fact that we need Japanese terms as queries to obtain Web pages that contain Japanese text. Another is that the search engines based in the United States do not to cover as many Japanese Web pages as the engines based in Japan. We used Japanese query terms selected from a keyword index from newspaper articles published in the Mainichi Shinbun between 1997 and 1998.
We used the following four major search engines in Japan.
We found 597 query terms that meet the conditions described in section 2.2. Then we retrieved all the documents in the search result, and checked the existence of query terms from 27-29 December 1999. Table 1 shows the number of pages found and the overlap size among four services. To estimate the number of Japanese indexable Web pages, we need the numbers of Web pages indexed by search engines. We used the following numbers: 35 million pages for Goo and 30 million pages for Lycos Japan, which were publicly known at the time of our experiment [3,9]. By computing the size of the overlap between the results of these two largest search engines, we have obtained an estimate of 88 million pages with a 95 percent confidence level interval of 1 million.
|Search Engines||Probabilities||Indexable Web
(millions of pages)
This estimate, 88 million pages, is larger than other estimates currently known in Japan. A white paper published by the Ministry of Post and Telecommunications, Japan, lists an estimate of 29 million pages in 1999. The size of the largest search engine in Japan is about 40 percent of our estimated size of the Japanese Web. Figure 3 shows the relationship between our estimate and other statistics.
Figure 3: Estimated size of Japanese Web pages
Figure 4 shows the relative coverage of search engines in Japan, and the ratio between the pages in and outside of the JP domain. It shows that less than 10 percent of Japanese Web pages are outside of the JP domain.
Figure 4: Relative coverage of search engines in Japan
Figure 5 shows the percentages of invalid URLs and invalid pages in the search results. Invalid URLs are those for which we could not retrieve the corresponding pages. Invalid pages are those we retrieved, but that did not contain the query term used in the search request. This figure suggests that two search engines, Infoseek Japan and Goo, crawl more often than the other two search engines.
Figure 5: Invalid URLs and invalid pages in the search results
There are currently 2564, about 4 billion, possible IP addresses. By obtaining a random sample of IP addresses and testing for a Web server at a standard port, we can estimate the number of Web servers at a standard port. Furthermore, we can estimate the number of Web pages if we know the distribution of the number of Web pages among Web servers.
There are several reasons that this method's estimate is different from the estimate in the previous section.
Lawrence and Giles chose 3.6 million IP addresses at random and tested for a Web server at a standard port. They found a Web server for one in every 269 addresses, and estimated the total number of Web servers as 16 million. After excluding Web servers with empty contents, they have estimated 2.8 million as the total number of public Web servers. Then they have observed the number of indexable Web pages of 2,500 Web servers, chosen from the Web servers found in the above samples. They found the mean number of Web pages to be 289 and produced an estimate of 800 million Web pages.
Among 2564 IPv4 addresses, about 2.8 million IP addresses are currently managed by JPNIC . From these IP addresses, we chose at random 28,000 IP addresses. We tested for a Web server at a standard port and got 335 responses. Among 335 responses, there were 175 successful responses. Examining these 175 Web servers, we found that 85 servers hold indexable Web pages. We obtain 85,000 as an estimate of the total number of public Web servers obtained by random IP sampling of JPNIC address space. We then retrieved indexable web pages from these 85 servers, found the mean number of Web pages to be about 200, and obtained an estimate of 17 million Web pages in the IP address space managed by JPNIC.
Although the estimate for the number of Web pages, 17 million, seems too small, we believe the estimate for the number of Web servers in the IP address space managed by JPNIC, 85,000, is reasonable. The Netcraft Web Server Survey reports 70,851 Web servers in the JP domain in May 1999 . The WWW-in-JP Server Survey by Hitachi Seibu Software, Ltd., reports 78,015 Web servers with the name in the form of www.*.*.jp in November 1999 . We are aware of the fact that all IP addresses allocated by JPNIC are not necessarily in the JP domain, and that some names in the JP domain are assigned to IP addresses not managed by JPNIC. However, by studying IP address of the host names found in the search results in the previous section, we found that less than 10 percent of the names in the IP address space managed by JPNIC are not in the JP domain. We also found that less than 5 percent of the names in the JP domain are outside the IP address space managed by JPNIC. Thus, we believe that the number of Web servers in the JP domain is close to the number of Web servers in the IP address space managed by JPNIC.
In the experiments by Lawrence and Giles, the value estimated by the first method, 320 million, is smaller than the value by the second, 800 million. This result seemingly suggests that the first method might only give an estimate of a subset of Web pages that the second method can estimate. After all, the first method is based on the sampling of Web pages that can be retrieved by English query terms, while the second method is based on the sampling of Web pages regardless of the relevance of their contents. However, when our first methods was applied to the JP domain, it gave the estimate of 88 million, which is much larger than the 17 million estimated by the second method. This differnce suggests that the two methods do not measure the same set of Web space.
One reason for the difference between the results by Lawrence and by us might be found in an explanation that the Web pages in the JP domain are less linked to each other than Web pages in the entire Web observed by Lawrence and Giles. More precisely, the pages in the JP domain are more likely to be isolated from the root document of a Web server. This situation can be understood by an example. Assume a user, X, has placed his Web pages at a provider, Y. The user X's home page, http://www.Y.ne.jp/~X/, is usually not linked from the root document http://www.Y.ne.jp/, either directly or indirectly. If the user X registers his home page at some directory services, these pages will be crawled eventually by various search engines' crawlers and will become searchable.
At the moment, we made an observation that about one-third of the URLs that were collected in the experiment in section 2.2 cannot be reached by following the links from the root document. To understand the cause of the difference between Lawrence's experiments and ours, we still need to collect more data on back links, i.e., which pages have links to a given page. Eventually we need models for describing the connectivity of the Web, such as the small-world networks proposed by Watts .
We have applied Lawrence's two methods to estimate the number of Japanese Web pages in the JP domain. The first method gave an estimate of 88 million pages, as a lower bound on the size of Japanese indexable Web pages. This method showed that the number of Japanese Web pages is much larger than existing statistics and Japanese search engines cover only a part of the Japanese Web. The second method gave an estimate of about 17 million pages, suggesting that Web pages in the JP domain are less connected compared with Web pages in the overall Web, although we need more research to explain the difference between these two methods.