Structural Properties of the African Web

Paolo Boldi
Dipartimento di Scienze dell'Informazione
Università degli Studi di Milano
Via Comelico 39/41
I-20135 Milano, Italy
boldi@dsi.unimi.it
Bruno Codenotti
Istituto di Informatica e Telematica
Consiglio Nazionale delle Ricerche
Via Moruzzi 1
I-56010 Pisa, Italy
codenotti@imc.pi.cnr.it
Massimo Santini
Università di Modena e Reggio Emilia
Via Fratelli Manfredi
I-42100 Reggio Emilia, Italy
msantini@unimo.it
Sebastiano Vigna
Dipartimento di Scienze dell'Informazione
Università degli Studi di Milano
Via Comelico 39/41
I-20135 Milano, Italy
vigna@acm.org

ABSTRACT

In this poster we illustrate some data about the African web. These data have been collected using UbiCrawler, a distributed Web crawler designed and developed by the authors.

Keywords

African web, web graph, structure analysis, content analysis

1. INTRODUCTION

The purpose of this poster is to present some structural information on the African Web, obtained by means of a distributed Web crawler [1]. Since classifying the nationality of a .com or .net site is a debatable process, we have been filtering sites by using suffixes of African Internet addresses. Therefore, with the term "African web" we mean the set of web sites whose address ends with the suffix of an African country.

The results we get suffer from the widespread use of .com and .net domains [2]. However, our choice has also some advantages. First of all, we divide URLs by country, and thus measure the degree of interconnection. Second, we believe that nationally based servers best reflect the status of awareness of web technology in Africa. Sites with other suffixes are often outsourced or externally hosted, and thus do not really reflect the degree of technology of the customer.

The already mentioned report [2] by Jensen, dated May 2001, puts into evidence some important characteristic of the African Internet. First of all, the Internet has grown rapidly in Africa, especially over the last 2-3 years, although it has been largely confined to major cities; recent estimates of the number of African Internet users give figures around 4 million in total, with about 1.5 million outside of South Africa. Another interesting issue is that of connectivity: it turns out that "Aside from local Internet links between South Africa, Lesotho and Swaziland network and a link between Mauritius and Madagascar, there are no other regional backbones or links between neighbouring countries." [2]. Our results complement and give further evidence to some of the above observations. Whenever relevant, we give comparisons with companion data we have gathered for the .com and .it domains.

1.1. TOOLS

In this section we briefly discuss the tools used to perform the data collection and analysis.

Data Collection. We collected about 2,000,000 pages using UbiCrawler (formerly named Trovatore) [1], a distributed web crawler we developed to gather data about the web, starting from a seed of about 2,500 sites (chosen from popular directories and search engines sites). The downloading of the pages started on 9 February 2002 and required one day to be completed, due to the very high latency and low network bandwidth of the African internet [2].

Data Extraction and Manipulation. Data have been extracted using tools integrated with UbiCrawler; in particular, HTML parsing was performed using standard Java 1.4 API [11]. Hence, data was processed both with ad-hoc statistical tools and R [6], a sophisticated language for data analysis and plotting.

Language Recognition. We used text_cat, a tool for n-gram based text categorization [4].

2. DATA OBTAINED FROM PARSING

2.1. HTML LEVEL

The large majority of pages in the African domain does not have a document type (the DOCTYPE declaration is mandatory in SGML documents, and very important for validation). A report about the English web in 1997 [3] showed that 25% of the documents online had a DOCTYPE declaration. Most documents declaring a document type stating that they conform to HTML 4, but there are still pages using a HTML level of 3 or lower.

DOCTYPENumber%
HTML 41509657.71%
HTML 3815054.16%
HTML 2405282.07%
HTML 16930.03%
none154219178.81%
other1396427.11%
Table 1. Number and percentage of pages classified by HTML level (as declared in the DOCTYPE)

2.2. HTTP HEADER

The distribution of headers in pages does not seem to differ significantly from data which are known for more general investigations on the Web. The only relevant difference is the almost complete absence of the p3p header, specifying the privacy policy, which ranks eleventh (13.68%) in the .com domain.

Header %
content-type99.88%
server99.71%
date99.16%
connection96.00%
content-length64.80%
last-modified43.70%
accept-ranges42.57%
etag41.05%
set-cookie34.00%
cache-control31.98%
Table 2. Frequenct of header types

2.3. SERVER

The majority of sites in the African domain use Microsoft® technology. The figures about IIS and Apache are almost exactly exchanged with respect to the trend of the world-wide web as reported by the NetCraft survey [4] (Apache 63.69%, IIS 26.97%); they are also different from the .it domain, where their figures are about the same (~45%).

Server %
Microsoft-IIS56.10%
Apache37.95%
Netscape-Enterprise1.50%
Lotus-Domino1.04%
Apache-AdvancedExtranetServer0.92%
WebSitePro0.30%
WebSTAR0.29%
Oracle0.21%
Netscape-FastTrack0.19%
IBM0.15%
Table 3. Percentage of server types

2.4. LAST MODIFICATION DATE

The distribution of last modification dates, as emerging from HTTP headers, is concentrated around the month preceding our visit of the African Web.

Last-modification date histogram
Figure 1. Histogram representing the last-modification date (as declared in the Last-Modified header); "Crawl" denotes the date of visit (9 Feb 2002).

2.5. PAGE SIZE

Figure 2 shows the histogram of page sizes (in logarithmic scale); the data obtained agree with those presented in other works analyzing local portions of the web (e.g., [14] and [15]). Note that only the textual content is considered (embedded images are ignored).

Histogram of page size frequencies
Figure 2. Histogram of page sizes
Size
Min 0
Max 524300
1st Qu. 2307
Median 6935
3rd Qu. 15990
Mean 12920
Std. dev. 24061.55
Table 4. Page size

2.6. NATURAL LANGUAGE

The language distribution was studied using n-gram based text categorization [4], as implemented by text_cat. The results shown concern the top 7 languages. It may be worth noticing that there is not even the faintest relation between this distribution and the African reality: for example, according to [12], the number of English native speakers is slightly more than 5,500,000, no more than 0.007% of the whole African population; French and Spanish are spoken by only 950,000 and 31,000 people, respectively.

Language%
English74.68%
French7.33%
Spanish5.57%
Afrikaans2.97%
German1.21%
Danish1.05%
Portuguese0.92%
Table 5. Frequency of languages

2.7. TAG AND TAG/ATTRIBUTE PAIR

The high percentage of table-related elements, and in particular of TD elements of fixed width, suggests that tables are being used for layout purposes. The IMG element, which ranks just after the ubiquitous A element both in .it and .com, is much rarer in the African web, probably due to the low bandwitdh, which makes textual information preferrable. The percentage of IMG elements having an ALT attribute for accessibility is about 34.52% (for .com is about 44.32%).

Tag# / page
TR22.51
A22.49
FONT20.47
BR17.58
IMG14.07
P13.56
B9.53
SPAN8.40
TABLE6.56
OPTION4.53
DIV2.62
O2.54
INPUT2.44
LI2.36
STRONG1.95
Tag% pages
HTML99.41%
HEAD99.39%
BODY99.38%
TITLE93.37%
A85.47%
TABLE79.03%
TR79.01%
TD78.94%
IMG77.97%
P75.46%
BR74.84%
FONT73.70%
B67.70%
META63.07%
DIV43.09%
Tag + Attribute# / page
A HREF21.62
TD WIDTH19.25
FONT SIZE19.10
FONT FACE18.11
IMG SRC14.00
TD VALIGN13.11
TD ALIGN11.96
IMG WIDTH10.95
IMG HEIGHT10.92
FONT COLOR10.40
TD BGCOLOR10.16
IMG BORDER9.37
TD HEIGHT8.39
TD CLASS6.67
TABLE BORDER5.88
Tag + Attribute% pages
A HREF85.01%
IMG SRC77.96%
TABLE BORDER74.98%
TABLE WIDTH72.59%
IMG HEIGHT69.15%
IMG WIDTH69.11%
TABLE CELLPADDING68.55%
IMG BORDER67.81%
TABLE CELLSPACING67.67%
TD WIDTH65.80%
FONT SIZE65.71%
META CONTENT62.77%
TD VALIGN59.91%
FONT COLOR57.81%
FONT FACE57.67%
Table 6. Average number of tags, and tag/attribute paris, per page and percentage of pages where tags, or pairs, occurs

2.8. SCRIPTING LANGUAGE

Scripting is comparatively less common in the African web (39.45% of the pages). The domains .com and .it sport a percentage of scripted pages of about 62.43% and 48.01%, respectively. The type of language is deduced from the deprecated LANGUAGE attribute, except for text/javascript, which is deduced from TYPE.

Script# / page
javascript0.7946
text/javascript0.1233
javascript1.20.0426
javascript1.10.0365
vbscript0.0043
jscript0.0035
Script %
javascript32.37%
text/javascript6.66%
javascript1.23.16%
javascript1.12.70%
vbscript0.39%
jscript0.18%
Table 7. Average number of occurrence of a scripting language, and percentage of pages where a scripting language occurs. The data are taken only from pages with SCRIPT elements (39.45%), but the percentage sum is slightly larger because some pages contain scripts with different TYPE (or LANGUAGE) attributes.

2.9. FILE EXTENSION IN CHILD URLs

Coherently with the predominance of IIS among servers, the most common extension among child URLs is .asp. Static pages follow, and then PHP dynamic pages. Some data about image types in child URLs: JPG 68.06%, GIF 31.30%, PNG 0.55%.

Extension# / page
.asp6.12
.htm3.94
.html3.19
.php1.64
.cfm0.42
.cgi0.41
.php30.37
.pl0.36
.length0.19
.exe0.18
Extension% page
.htm36.48%
.asp35.96%
.html30.55%
.php7.22%
.cgi4.23%
.pl3.22%
.cfm2.64%
.php32.25%
.jpg1.63%
.shtml1.42%
Table 8. Average number of file extensions in child URLs per page and percentage of pages where file extensions occur (note that .htm and .html are synonymous).

2.10. PROTOCOL IN CHILD URLs

The distribution of protocols is very similar to the one typically observed for other domains, except for the low occurrence of the https protocol.

Protocol# / page
http21.32
mailto0.62
javascript0.54
https0.05
ftp0.04
file0.01
news< 0.01
gopher< 0.01
Protocol%
http86.02
mailto29.97
javascript14.94
https2.30
ftp0.38
file0.36
news0.07
gopher0.02
Table 9. Average number of protocol type per page and percentage of pages where a protocol type occurs

3. THE AFRICAN WEB GRAPH

3.1. DEGREE DISTRIBUTION

To study the indegree and outdegree distribution of the African Web graph, we have performed some statistical evaluations, computing, for instance, extremals, quantiles and mean. One immediately notices that the max in-degree is characterized by very high figures, associated with a large standard deviation. A human inspection of the collected URLs revealed that this comes from the presence of many "portal sites" and "directories", which link to some popular URLs from different points of their hierarchy. As expected, such a phenomenon is less relevant for the outdegree distribution.

In-degree Out-degree
Min 0 0
Max 23660 1975
1st Qu. 1 1
Median 2 6
3rd Qu. 4 17
Mean 13.32 13.32
Std. dev. 117.53 24.12
Table 10. In- and out-degree distributions

3.2. POWER LAW

It is observed that in-degrees of web pages typically follow a Power Law distribution [8, 9, 10]. This means that the number of URLs with i in-links is proportional to iα for some constant α < 0. The analysis of the data collected for the African web gives futher evidence to such observation. After discarding all the pages whose in-degree exceeded 1,000, we have computed some figures which led to an estimate of -1.92 for α.

The power-law distribution
Figure 3. Blue circles represent the logarithm of in-degree plotted against the logaritm of the number of pages with such in-degree; the red line is a linear fit of such data

3.3. THE STRUCTURE OF STRONGLY CONNECTED COMPONENTS

The graph structure of the African web presents some differences from the "bowtie theory" [7]. Half of the pages we have downloaded are condensed into a single giant strongly connected component, pointing to several smaller components. We have no evidence regarding components pointing to the giant component, possibly because of our seed choice. Nonetheless, the size of the main component is much larger than usually observed for the Web, possibly because most sites are from South Africa (suffix .za), and regional web sites tend to be more connected.

w/ singletons w/o sinlgetons
Max 977300
1st Qu. 1 3
Median 1 6
3rd Qu. 1 13
Mean 3.33 85.8
Std. dev. 1274.85 7696.56
Table 11. Distribution of sizes
The graph of strongly connected components
Figure 4. The graph of strongly connected components, where we have only represented components with at least 800 pages. The length of the diameter of each nodes is logarithmic in the size of the corresponding component.

3.4. INTERCONNECTION

The poor connectivity properties observed in [2] are confirmed by our experiments; indeed, with the exception of a large number of links from Namibia to South Africa, and some interconnection between Morocco and Senegal, the African web graph is largely disconnected, and presents a high degree of internal connection within each state (represented by a specific suffix).

No. of pages.za.ma.tn.eg.na.zw.sn.mz.ly.com
.za (South Africa)16097222567061772111217825854013333661355982
.ma (Morocco)649422768136632132  18633417299
.tn (Tunisia)47904377383543511   7001
.eg (Egypt)417801441437665     8169
.na (Namibia)251225595  735556511 5 5406
.zw (Zimbabwe)24631185  24534892 3 25983
.sn (Senegal)2318470275112272103011 5797
.mz (Mozambique)11564445   2111177031 30528
.ly (Libya)1087618       1386167263
Table 12. Pages per country and country interconnection (number of links)

4. CONCLUSIONS

Putting Africa on the Web was a goal of the early 90's, with several organizations involved in the process. As already mentioned, a status report on the growth of the African Web come out in 2001 [2], which indicated trends and properties of the growth of Internet usage in Africa. This poster has provided a complementary view of the African Web, in terms of both structure of the pages and most used technologies, and structural properties of the African web graph.

The main evidences emerging from our analysis are significant departures from known properties of the web graph in terms of connectivity and the widespread use of mature technologies.

We plan to periodically refresh our information in order to keep track of the evolution of the African web graph, and in particulat to analyze how its connectivity changes over time.

5. REFERENCES

  1. Paolo Boldi, Bruno Codenotti, Massimo Santini, and Sebastiano Vigna. Trovatore: Towards a highly scalable distributed web crawler. In Proc. of Tenth International World Wide Web Conference, Hong Kong, China, 2001.
  2. Mike Jensen. The African Internet—A Status Report. http://demiurge.wn.apc.org/africa/afstat.htm
  3. Dave Beckett. 30% Accessible—A Survey of the UK Wide Web. In Proceedings of Sixth International World Wide Web Conference, Santa Clara, California, USA, 1997.
  4. William B. Cavnar and John M. Trenkle. N-Gram-Based Text Categorization. In Proceedings of Third Annual Symposium on Document Analysis and Information Retrieval, Las Vegas, NV, UNLV Publications/Reprographics, pages 161−175.
  5. The Netcraft Web Server Survey, January 2002 http://www.netcraft.com/survey/.
  6. Ross Ihaka and Robert Gentleman. R: A Language for Data Analysis and Graphics. Journal of Computational and Graphical Statistics, 5(3):299−314, 1996.
  7. Andrei Broder, Ravi Kumar, Farzin Maghoul, Prabhakar Raghavan, Sridhar Rajagopalan, Raymie Stata, Andrew Tomkins, and Janet Wiener. Graph structure in the web: experiments and models. In Proceedings of the Ninth International World--Wide Web Conference, Amsterdam, The Netherlands, 2000.
  8. Jon M. Kleinberg, Ravi Kumar, Prabhakar Raghavan, Sridhar Rajagopalan, and Andrew S. Tomkins. The Web as a graph: Measurements, models, and methods. In T. Asano, H. Imai, D. T. Lee, S. Nakano, and T. Tokuyama, editors, Proc. 5th Annual Int. Conf. Computing and Combinatorics, COCOON, number 1627 in Lecture Notes in Computer Science, LNCS. Springer--Verlag, July 1999.
  9. Reka Albert, Albert Laszlo Barabasi, and Hawoong Jeong. Diameter of the World Wide Web. Nature, 401(6749), September 1999.
  10. A. L. Barabasi, R. Albert, H. Jeong, and G. Bianconi. Power-law distribution of the World Wide Web. Science, 287:2115a, 2000.
  11. Java ™, http://java.sun.com/j2se/1.4/.
  12. Ethnologue: Volume 1 Languages of the World, 14th Edition. Edited by Barbara F. Grimes, 2000 http://www.ethnologue.com/home.asp.
  13. The Platform for Privacy Preferences 1.0 (P3P1.0) Specification. W3C Proposed Recommendation, 28 January 2002. http://www.w3.org/TR/P3P/.
  14. Surasak Sanguanpong, Punpiti Piamsa-nga, Yuen Poovarawan and Suthiphol Warangrit. Measuring Thai Web Using NontriSpider. Proceedings of the International Forum cum Conference on Information Technology and Communication, pages 123-132, Bangkok, June 2000.
  15. Altigran S. da Silva, Eveline A. Veloso, Paulo B. Golghe, Berthier Ribeiro-Neto, Alberto H. F. Laender and Nivio Ziviani. CoBWeb - A Crawler for the Brazilian Web. Proceedings of the String Processing and Information Retrieval Symposium & International Workshop on Groupware, Cancun, Mexico, 1998.