Dm4

World Wide Web Usage Mining Systems and Technologies
Wen-Chen Hu
Department of Computer Science
University of North Dakota, Grand Forks, ND 58202
wenchen@cs.und.edu

Xuli Zong
GE Global Research
General Electric Company, Niskayuna, NY 12309
zong@crd.ge.com

Chung-wei Lee
Department of Computer Science and Software Engineering
Auburn University, Auburn, AL 36849
chwlee@eng.auburn.edu

and

Jyh-haw Yeh
Department of Computer Science
Boise State University, Boise, ID 83725
jhyeh@cs.boisestate.edu

ABSTRACT mining technologies and systems. A Web usage mining system
must be able to perform five major functions: i) data gathering,
Web usage mining is used to discover interesting user naviga- ii) data preparation, iii) navigation pattern discovery, iv) pattern
tion patterns and can be applied to many real-world problems, analysis and visualization, and v) pattern applications.
such as improving Web sites/pages, making additional topic or
product recommendations, user/customer behavior studies, etc. Requirements of Web Usage Mining
This article provides a survey and analysis of current Web usage It is necessary to examine what kind of features a Web usage
mining systems and technologies. A Web usage mining system mining system is expected to have in order to conduct effective
performs five major tasks: i) data gathering, ii) data preparation, and efficient Web usage mining, and what kind of challenges
iii) navigation pattern discovery, iv) pattern analysis and visu- may be faced in the process of developing new Web usage min-
alization, and v) pattern applications. Each task is explained in ing techniques. A Web usage mining system should be able to:
detail and its related technologies are introduced. A list of ma-
jor research systems and projects concerning Web usage mining • Gather useful usage data thoroughly,
is also presented, and a summary of Web usage mining is given • Filter out irrelevant usage data,
in the last section.
• Establish the actual usage data,
Keywords: World Wide Web, Usage Mining, Navigation Pat- • Discover interesting navigation patterns,
terns, Usage Data, and Data Mining. • Display the navigation patterns clearly,
• Analyze and interpret the navigation patterns correctly, and
• Apply the mining results effectively.
1. INTRODUCTION

World Wide Web Data Mining includes content mining, hyper- Paper Organization
link structure mining, and usage mining. All three approaches Many Web usage mining technologies have been proposed and
attempt to extract knowledge from the Web, produce some use- each technology employs a different approach. This article first
ful results from the knowledge extracted, and apply the results describes a generalized Web usage mining system, which in-
to certain real-world problems. The first two apply the data cludes five individual functions. Each system function is then
mining techniques to Web page contents and hyperlink struc- explained and analyzed in detail. It is organized as follows:
tures, respectively. The third approach, Web usage mining, the Section 2 gives a generalized structure of a Web usage mining
theme of this article, is the application of data mining tech- system and Sections 3 to 7 introduce each of the five system
niques to the usage logs of large Web data repositories in order functions and list its related technologies in turn. Major re-
to produce results that can be applied to many practical subjects, search systems and projects concerning Web usage mining are
such as improving Web sites/pages, making additional topic or listed in Section 8 and the final section summarizes the material
product recommendations, user/customer behavior studies, etc. covered in the earlier sections. Related surveys of Web usage
This paper provides a survey and analysis of current Web usage mining techniques can also be found in [18,20,32].

SYSTEMICS, CYBERNETICS AND INFORMATICS VOLUME 1 - NUMBER 4 53

2. A SYSTEM STRUCTURE • Impersonal: The user is observed as a unit of unknown
identity, although some properties may be accessible from
A variety of implementations and realizations are employed by demographic data. In this case, a usage mining system
Web usage mining systems. This section gives a generalized works for a general population, for example, the most
structure of the systems, each of which carries out five major popular products are listed for all customers.
tasks:
This paper concentrates on the impersonal systems. Personal
• Usage data gathering: Web logs, which record systems are actually a special case of impersonal systems, so
user activities on Web sites, provide the most comprehen- readers can easily infer the corresponding personal systems,
sive, detailed Web usage data. given the information for impersonal systems.
• Usage data preparation: Log data are normally
too raw to be used by mining algorithms. This task re-
stores the users' activities that are recorded in the Web 3. DATA GATHERING
server logs in a reliable and consistent way.
• Navigation pattern discovery: This part of a Web usage data are usually supplied by two sources: trial runs
by humans and Web logs. The first approach is impractical and
usage mining system looks for interesting usage patterns
rarely used because of the nature of its high time and expense
contained in the log data. Most algorithms use the method
costs and its bias. Most usage mining systems use log data as
of sequential pattern generation, while the remaining meth-
their data source. This section looks at how and what usage
ods tend to be rather ad hoc.
data can be collected.
• Pattern analysis and visualization: Navi-
gation patterns show the facts of Web usage, but these re-
quire further interpretation and analysis before they can be Client Web P roxy Web
Browser Server Server
applied to obtain useful results.
• Pattern applications: The navigation patterns Requests Requests

discovered can be applied to the following major areas,
among others: i) improving the page/site design, ii) making Results Results
additional product or topic recommendations, iii) Web per-
sonalization, and iv) learning the user or customer behav-
ior. Log Log Log

WWW
Figure 2: Three Web log file locations.

Instructions
Usage Data Gathering Web Logs
Raw Data A Web log file records activity information when a Web user
Instructions submits a request to a Web server. A log file can be located in
Usage Data P reparation
three different places: i) Web servers, ii) Web proxy servers,
Prepared Dat a and iii) client browsers, as shown in Figure 2, and each suffers
Queri es from two major drawbacks:
Navigation P attern Discovery

System Result Navigation Patterns • Server-side logs: These logs generally supply the
Administrator Patterns
P attern Analysis & Visualization
most complete and accurate usage data, but their two draw-
backs are:
Result Patterns
Instructions
P attern Applications
o These logs contain sensitive, personal information,
therefore the server owners usually keep them closed.
o The logs do not record cached pages visited. The
Figure 1: A Web usage mining system structure. cached pages are summoned from local storage of
browsers or proxy servers, not from Web servers.

Figure 1 shows a generalized structure of a Web usage mining • Proxy-side logs: A proxy server takes the HTTP
system; the five components will be detailed in the next five
requests from users and passes them to a Web server; the
sections. A usage mining system can also be divided into the
proxy server then returns to users the results passed to
following two types:
them by the Web server. The two disadvantages are:
• Personal: A user is observed as a physical person, for o Proxy-server construction is a difficult task. Ad-
whom identifying information and personal data/properties vanced network programming, such as TCP/IP, is re-
are known. Here, a usage mining system optimizes the in- quired for this construction.
teraction for this specific individual user, for example, by o The request interception is limited, rather than cover-
making product recommendations specifically designed to ing most requests.
appeal to this customer.

54 SYSTEMICS, CYBERNETICS AND INFORMATICS VOLUME 1 - NUMBER 4

The proxy logger implementation in WebQuilt [7], a Web • Status: The HTTP status code returned to the client,
logging system, can be used to solve these two problems, e.g., 200 is “ok” and 404 is “not found.”
but the system performance declines if it is employed be-
cause each page request needs to be processed by the proxy The CGI environment variables [8] supply values for many of
simulator. the above items.
• Client-side logs: Participants remotely test a Web
site by downloading special software that records Web us-
age or by modifying the source code of an existing 4. DATA PREPARATION
browser. HTTP cookies could also be used for this pur-
pose. These are pieces of information generated by a Web The information contained in a raw Web server log does not
server and stored in the users’ computers, ready for future reliably represent a user session file. The Web usage data
access. The drawbacks of this approach are: preparation phase is used to restore users' activities in the Web
server log in a reliable and consistent way. This phase should at
o The design team must deploy the special software and a minimum achieve the following four major tasks: i) removing
have the end-users install it. undesirable entries, ii) distinguishing among users, iii) building
o This technique makes it hard to achieve compatibility sessions, and iv) restoring the contents of a session [11].
with a range of operating systems and Web browsers.
Removing Undesirable Entries
Web logs contain user activity information, of which some is
Web Log Information not closely relevant to usage mining and can be removed with-
A Web log is a file to which the Web server writes information out noticeably affecting the mining, for example:
each time a user requests a resource from that particular site.
Examples of the types of information the server preserves in- • All log image entries. The HTTP protocol
clude the user' s omain, subdomain, and hostname; the re-
d requires a separate connection for every file that is re-
sources the user requested (for example, a page or an image quested from the Web server. Images are automatically
map); the time of the request; and any errors returned by the downloaded based on the HTML page requested and the
server. Each log provides different and various information downloads are recorded in the logs. In the future, images
about the Web server and its usage data. Most logs use the may provide valuable usage information, but the research
format of a common log file [10] or extended log file [14]. For on image understanding is still in the early stages. Thus,
example, the following is an example of a file recorded in the log image entries do not help the usage mining and can be
extended log format. removed.
• Robot assesses. A robot, also known as spider or
#Version: 1.0 #Date: 12-Jan-1996
crawler, is a program that automatically fetches Web
00:00:00 #Fields: time cs-method
cs-uri 00:34:23 GET /foo/bar.html pages. Robots are used to feed pages to search engines or
12:21:16 GET /foo/bar.html other software. Large search engines, like Alta Vista, have
12:45:52 GET /foo/bar.html many robots working in parallel. As robot-access patterns
12:57:34 GET /foo/bar.html are usually different from human-access patterns, many of
the robot accesses can be detected and removed from the
The following list shows the information may be stored in a logs.
Web log:
As much irrelevant information as possible should be removed
• Authuser: Username and password if the server re- before applying data mining algorithms to the log data.
quires user authentication.
• Bytes: The content-length of the document transferred. A
• Entering and exiting date and time.
• Remote IP address or domain name: An IP B C D
address is a 32-bit host address defined by the Internet Pro-
tocol; a domain name is used to determine a unique Inter-
net address for any host on the Internet such as,
cs.und.nodak.edu. One IP address is usually defined E F H I
for one domain name, e.g., cs.und.nodak.edu points
to 134.129.216.100.
• Modus of request: GET, POST, or HEAD method of Figure 3: A sample Web site.
CGI (Common Gateway Interface).
• Number of hits on the page.
• Remote log and agent log. Distinguishing among Users
• Remote URL. A user is defined as a single individual that accesses files from
one or more Web servers through a browser. A Web log se-
• “request:” The request line exactly as it came from the quentially records users’ activities according to the time each
client. occurred. In order to study the actual user behavior, users in the
• Requested URL. log must be distinguished. Figure 3 is a sample Web site where
• rfc931: The remote logname of the user. nodes are pages, edges are hyperlinks, and node A is the entry
page of this site. The edges are bi-directional because users can


easily use the back button on the browser to return to the pre- where the number inside the parentheses is the number of visi-
vious page. Assume the access data from an IP address re- tors per trail. An aggregate tree constructed from the list is
corded on the log are those given in Table 1. Two user paths shown in Figure 4, where the number after the page is the sup-
are identified from the access data: i) A-D-I-H-A-B-F and ii) port, the number of visitors having reached the page. A Web
C-H-B. These two paths are found by heuristics; other possi-
bilities may also exist.
(F, 3) (H, 3)
(B, 5)
Table 1: Sample access data from an IP address on the Web site
(A, 9) (E, 2)
in Figure 3.
(D, 4) (I, 4)
No. Time Requested URL Remote URL ( , 50)
1 12:05 A – (E, 2) (F, 2) (H, 2)
2 12:11 D A
(B, 5)
3 12:22 C –
4 12:37 I D (F, 3) (C, 3) (H, 3)
5 12:45 H C
6 12:58 B A
Figure 4: An aggregate tree constructed from the list
7 01:11 H D of visitor trails.
8 02:45 A –
9 03:16 B A
10 03:22 F B usage mining system then looks for “interesting” navigation
patterns from this aggregate tree. Some of the interesting navi-
gation patterns are related to the following three topics:
Building Sessions
For logs that span long periods of time, it is very likely that • Statistics: for example, which are the most popular
individual users will visit the Web site more than once or their paths?
browsing may be interrupted. The goal of session identification • Structure: for example, what pages are usually ac-
is to divide the page accesses of each user into individual ses- cessed after users visit page A?
sions. A time threshold is usually used to identify sessions. For
• Content: for example, thirty percent of sports page
example, the previous two paths can be further assigned to three
viewers will enter the baseball pages.
sessions: i) A-D-I-H, ii) A-B-F, and iii) C-H-B if a threshold
value of thirty minutes is used.
Figure 5 shows an example of navigation patterns from page B
to page H in Figure 4.
Restoring the Contents of a Session
This task determines if there are important accesses that are not
recorded in the access logs. For example, Web caching or using (E, 2) (F, 2) (H, 2)
the back button of a browser will cause information discon-
(B, 5)
tinuance in logs. The three user sessions previously identified
can be restored to obtain the complete sessions: i) A-D-I-D-H, (F, 3) (C, 3) (H, 3)
ii) A-B-F, and iii) C-H-A-B because there are no direct links
between I and H and between H and B in Figure 3. (B, 5) (F, 3) (H, 3)

Figure 5: The navigation patterns from page B to page H in
5. NAVIGATION PATTERN DISCOVERY
Figure 4.
Many data mining algorithms are dedicated to finding naviga-
tion patterns. Among them, most algorithms use the method of
sequential pattern generation, while the remaining methods tend Sequential Pattern Generation
to be rather ad hoc. The problem of discovering sequential patterns consists of find-
ing intertransaction patterns such that the presence of a set of
A Navigation Pattern Example items is followed by another item in the time-stamp ordered
Before giving the details of various mining algorithms, the fol- transaction set [4]. The following three systems each use a
lowing example illustrates one procedure that may be used to variant of sequential pattern generation to find navigation pat-
find a typical navigation pattern. Assume the following list terns:
contains the visitor trails of the Web site in Figure 3.
• WUM (Web Utilization Miner) [31] discovers navigation
1. A-D-I (4) patterns using an aggregated materialized view of the Web
2. B-E-F-H (2) log. This technique offers a mining language that experts
3. A-B-F-H (3) can use to specify the types of patterns they are interested
4. A-B-E (2) in. Using this language, only patterns having the specified
5. B-F-C-H (3) characteristics are saved, while uninteresting patterns are
removed early in the process. For example, the following
query generates the navigation patterns shown in Figure 5.


select glue(t) • Displays the discovered navigation patterns clearly.
from node as B, H • Provides essential functions for manipulating navigation
template B×H as t patterns, e.g., zooming, rotation, scaling, etc.
where B='B' and H='H';
WebQuilt [17] allows captured usage traces to be aggregated
• MiDAS [6] extends traditional sequence discovery by add- and visualized in a zooming interface. The visualization also
ing a wide range of Web-specific features. New domain shows the most common paths taken through the Web site for a
knowledge types in the form of navigational templates and given task, as well as the optimal path for that task as desig-
Web topologies have been incorporated, as well as syntac- nated by the designers of the site.
tic constraints and concept hierarchies.
• Chen et al. [9] propose a method to convert the original
sequence of log data into a set of maximal forward refer- 7. PATTERN APPLICATIONS
ences. Algorithms are then applied to determine the fre-
quent traversal patterns, i.e., large reference sequences, The results of navigation pattern discovery can be applied to the
from the maximal forward references obtained. following major areas, among others: i) improving site/page
design, ii) making additional topic or product recommendations,
iii) Web personalization, and iv) learning user/customer behav-
Ad Hoc Methods ior. Web caching, a less important application for navigation
Apart from the above techniques of sequential pattern genera- patterns, is also discussed.
tion, some ad hoc methods worth mentioning are as follows:
Web Site/Page Improvements
• Association rule discovery can be used to find unordered The most important application of discovered navigation pat-
correlations between items found in a set of database trans- terns is to improve the Web sites/pages by (re)organizing them.
actions [3]. In the context of Web usage mining, associa- Other than manually (re)organizing the Web sites/pages [19],
tion rules refer to sets of pages that are accessed together there are some other automatic ways to achieve this. Adaptive
with a support value exceeding some specified threshold. Web sites [26] automatically improve their organization and
• OLAP (On-Line Analytical Processing) is a category of presentation by learning from visitor access patterns. They
software tools that can be used to analyze of data stored in mine the data buried in Web server logs to produce easily navi-
a database. It allows users to analyze different dimensions gable Web sites. Clustering mining and conceptual clustering
of multidimensional data. For example, it provides time mining techniques are applied to synthesize the index pages,
series and trend analysis views. WebLogMiner [33] uses which are central to site organization.
the OLAP method to analyze the Web log data cube, which
is constructed from a database containing the log data. Additional Topic or Product Recommendations
Data mining methods such as association or classification Electronic commerce sites use recommender systems or col-
are then applied to the data cube to predict, classify, and laborative filtering to suggest products to their customers or to
discover interesting patterns and trends. Büchner and provide consumers with information to help them decide which
Mulvenna [7] also make use of a generic Web log data products to purchase. For example, each account owner at
hypercube. Various online analytical Web usage data Amazon.com is presented with a section of Your Recom-
mining techniques are then applied to the hypercube to mendations, which suggests additional products based on the
reveal marketing intelligence. owner’s previous purchases and browsing behavior. Various
• Borges and Levene’s [5] model views navigation records technologies have been proposed for recommender systems [27]
in terms of a hypertext probabilistic grammar, which is a and many electronic commerce sites have employed recom-
probabilistic regular grammar. For this grammar, each mender systems in their sites [28]. For further studies, the
non-terminal symbol corresponds to a Web page and a GroupLens research group [16] at the University of Minnesota
production rule corresponds to a link between pages. The is known for its successful projects on various recommender
higher probability generated strings of the grammar corre- systems.
spond to the user' s pr
eferred trails.
Web Personalization
• Pei et al. [25] propose a data structure WAP-tree to store
Web personalization (re)organizes Web sites/pages based on the
highly compressed, critical information contained in Web
Web experience to fit individual users’ needs [22,30]. It is a
logs, together with an algorithm WAP-mine that is used to
broad area that includes adaptive Web sites and recommender
discover access patterns from the WAP-tree. systems as special cases. The WebPersonalizer system [23]
uses a subset of Web log and session clustering techniques to
derive usage profiles, which are then used to generate recom-
6. PATTERN ANALYSIS AND VISUALIZATION mendations. An overview of approaches for incorporating se-
mantic knowledge into the Web personalization process is given
Navigation patterns, which show the facts of Web usage, need in the article by Dai and Mobasher [12].
further analysis and interpretation before application. The
analysis is not discussed here because it usually requires human
User Behavior Studies
intervention or is distributed to the two other tasks: navigation Knowing the users' purchasing or brows behavior is a critical
ing
pattern discovery and pattern applications. Navigation patterns factor for the success of E-commerce. The 1:1Pro system [2]
are normally two-dimensional paths that are difficult to perceive constructs personal profiles based on customers’ transactional
if a proper visualization tool is not supported. A useful visuali- histories. The system uses data mining techniques to discover a
zation tool may provide the following functions:


Table 2: Major research systems and projects concerning Web usage mining.

No. Title URL Major Method/Application
1 Adaptive Web Sites http://www.cs.washington.edu/research/adaptive/ Pattern application
2 GroupLens http://www.cs.umn.edu/Research/GroupLens/ Recommender systems
3 MiDAS

Sequence discovery
4 WebQuilt http://guir.berkeley.edu/projects/webquilt/ Proxy logging
5 WebLogMiner http://www.dbminer.com/ OLAP application
6 WebSift http://www.cs.umn.edu/Research/webshift/ Data mining
7 WUM http://wum.wiwi.hu-berlin.de/ Sequence discovery

set of rules describing customers’ behavior and supports human A Web usage mining system performs five major functions: i)
experts in validating the rules. Fu et al. [15] propose an algo- data gathering, ii) data preparation, iii) navigation pattern dis-
rithm to cluster Web users based on their access patterns, which covery, iv) pattern analysis and visualization, and v) pattern
are organized into sessions representing episodes of interaction applications. Each function requires substantial effort to fulfill
between Web users and the Web server. Using attributed- its objectives, but the most crucial and complex part of this
oriented induction, the sessions are then generalized according system is its navigation pattern discovery function. Many usage
to the page hierarchy, which organizes pages according to their mining algorithms use the method of sequential pattern genera-
generalities. The generalized sessions are finally clustered us- tion, while the rest tend to use ad hoc methods. Sequential pat-
ing a hierarchical clustering method. tern generation does not dominate the algorithms, since naviga-
tion patterns are defined differently from one application to
Web Caching another and each definition may require a unique method.
Another application worth mentioning is Web caching, which is
the temporary storage of Web objects (such as HTML docu-
ments) for later retrieval. There are significant advantages to 10. REFERENCES
Web caching, e.g., reduced bandwidth consumption, reduced
server load, and reduced latency. Together, they make the Web [1] Access log analyzers. Retrieved June 02, 2003 from
less expensive and improve its performance [13]. Web caching http://www.uu.se/Software/Analyzers/Access-
may in turn be enhanced by navigation patterns. Lan et al. [21] analyzers.html
propose an algorithm to make Web servers “pushier.” Which [2] Gediminas Adomavicius and Alexander Tuzhilin. Us-
document is to be prefetched is determined by a set of associa- ing data mining methods to build customer profiles.
tion rules mined from a sample of the access log of the Web IEEE Computer, 34(2):74-82, February 2001.
server. Once a rule of the form “Document1 > Docu- [3] Rakesh Agrawal and Ramakrishnan Srikant. Fast algo-
ment2” has been identified and selected, the Web server de- rithms for mining association rules. In Proceeding of
cides to prefetch “Document2” if “Document1” is requested. the 20th Very Large DataBases Conference (VLDB),
Two other Web page prefetching methods based on access log pages 487-499, Santiago, Chile, 1994.
information are described in other papers [24,29]. [4] Rakesh Agrawal and Ramakrishnan Srikant. Mining
sequential patterns. In Proceedings of the 11th Interna-
tional Conference on Data Engineering, pages 3-14,
8. MAJOR SYSTEMS AND PROJECTS Taipei, Taiwan, March 1995.
[5] José Borges and Mark Levene. Data mining of user
Though Web usage mining is a fairly new research topic, many navigation patterns. In Proceedings of the Workshop on
systems and tools are already on the market [1]. Most provide Web Usage Analysis and User Profiling (WEBKDD),
only limited knowledge or information, such as the number of pages 31-36, San Diego, California, August 1999.
hits, the popular paths/products, etc., but Table 2 shows the [6] Alex G. Büchner, Matthias Baumgarten, Sarabjot S.
latest major research systems and projects in the field. They Anand, Maurice D. Mulvenna, and John G. Hughes.
make it possible to extract hidden knowledge from log data and Navigation pattern discovery from Internet data. In
apply the knowledge to certain real-world problems. Proceedings of the Workshop on Web Usage Analysis
and User Profiling (WEBKDD), San Diego, California,
August 1999.
9. SUMMARY [7] Alex G. Büchner and Maurice D. Mulvenna. Discover-
ing Internet marketing intelligence through online ana-
In less than a decade, the World Wide Web has become one of lytical Web usage mining. ACM SIGMOD Record,
the world’s three major media, with the other two being print 27(4):54-61, December 1998.
and television. Electronic commerce is one of the major forces [8] CGI environment variables. Retrieved May 15, 2003
that allow the Web to flourish, but the success of electronic from http://hoohoo.ncsa.uiuc.edu/cgi/env.html
commerce depends on how well the site owners understand [9] Ming-Syan Chen, Jong Soo Park, and Philip S. Yu.
users' ehavior and needs. Web usage mining can be used to
b Efficient data mining for path traversal patterns. IEEE
discover interesting user navigation patterns, which can then be Transactions on Knowledge and Data Engineering,
applied to real-world problems such as Web site/page improve- 8(6):866-883, 1996.
ment, additional product/topic recommendations, user/customer [10] Common log file format. Retrieved June 02, 2003 from
behavior studies, etc. This paper has provided a survey and http://www.w3.org/Daemon/User/Config/Logging.html
analysis of current Web usage mining systems and technologies.


[11] Robert Cooley, Bamshad Mobasher, and Jaideep Knowledge Discovery, Kluwer Publishing, 6(1):61-82,
Srivastava. Data preparation for mining World Wide January 2002.
Web browsing patterns. Knowledge and Information [24] Alexandros Nanopoulos, Dimitris Katsaros, and Yannis
Systems, 1(1):5-32, February 1999. Manolopoulos. Effective prediction of Web-user ac-
[12] Honghua Dai and Bamshad Mobasher. A road map to cesses: A data mining approach. In Proceedings of the
more effective Web personalization: Integrating domain Workshop on Mining Log Data Across All Customer
knowledge with Web usage mining. In Proceedings of Touchpoints (WEBKDD), San Francisco, California,
the International Conference on Internet Computing 2001.
(IC), Las Vegas, Nevada, June 2003. [25] Jian Pei, Jiawei Han, Behzad Mortazavi-asl, and Hua
[13] Brian D. Davison. A Web caching primer. IEEE Inter- Zhu. Mining access patterns efficiently from Web logs.
net Computing, 5(4):38-45, July/August 2001. In Proceedings of the Pacific-Asia Conference on
[14] Extended log file format. Retrieved June 03, 2003 from Knowledge Discovery and Data Mining (PAKDD),
http://www.w3.org/TR/WD-logfile.html 2000.
[15] Yongjian Fu, Kanwalpreet Sandhu, and Ming-Yi Shih. [26] Mike Perkowitz and Oren Etzioni. Towards adaptive
A generalization-based approach to clustering of Web Web sites: Conceptual framework and case study. Arti-
usage sessions. In Brij M. Masand and Myra Spiliopou- ficial Intelligence, 118:245-275, 2000.
lou, editors, Web Usage Analysis and User Profiling, [27] Badrul Sarwar, George Karypis, Joseph Konstan, and
Lecture Notes in Artificial Intelligence, 1836:21-38, John Riedl. Analysis of recommender algorithms for e-
Springer, 2000. commerce. In Proceedings of the ACM Electronic
[16] GroupLens Research. Retrieved May 12, 2003 from Commerce Conference, pages 158-167, October 2000.
http://www.cs.umn.edu/Research/GroupLens/ [28] J. Ben Schafer, Joseph Konstan, and John Riedl. Elec-
[17] Jason I. Hong and James A. Landay. WebQuilt: A tronic commerce recommender applications. Journal of
framework for capturing and visualizing the Web ex- Data Mining and Knowledge Discovery, 5(1/2):115-
perience. In Proceedings of the 10th International 152, 2000.
World Wide Web Conference, pages 717-724, Hong [29] Stuart E. Schechter, Murali Krishnan, and Michael D.
Kong, 2001. Smith. Using path profiles to predict HTTP requests.
[18] Wen-Chen Hu, Xuli Zong, Hung-Ju Chu, and Jui-Fa In Proceedings of the 7th International World Wide Web
Chen. Usage mining for the World Wide Web. In Pro- Conference, Brisbane, Australia, April 1998.
ceedings of the 6th World Multi-Conference on Sys- [30] Myra Spiliopoulou. Web usage mining for site evalua-
temics, Cybernetics and Informatics (SCI), pages 75-80, tion: Making a site better fit its users. Communications
Orlando, Florida, July 14-18, 2002. of ACM, 43(8):127-134, August 2000.
[19] Melody Y. Ivory and Marti A. Hearst. Improving Web [31] Myra Spiliopoulou and Lukas C. Faulstich. WUM: A
site design. IEEE Internet Computing, 6(2):56-63, tool for Web utilization analysis. In Proceedings of the
March/April 2002. Workshop on the Web and Databases (WEBDB),
[20] Raymond Kosala and Hendrik Blockeel. Web mining pages 184-203, Valencia, Spain, March 1998.
research: A survey. SIGKDD Explorations, 2(1):1-15, [32] Jaideep Srivastava, Robert Cooley, Mukund Deshpande,
2000. and Pang-Ning Tan. Web usage mining: Discovery and
[21] Bin Lan, Stephane Bressan, and Beng Chin Ooi. Mak- applications of usage patterns from Web data. ACM
ing Web servers pushier. In Proceedings of the Work- Special Interest Group on Knowledge Discovery and
shop on Web Usage Analysis and User Profiling, pages Data Mining (SIGKDD) Explorations, 1(2):12-23,
112-125, San Diego, California, August 1999. 2000.
[22] Bamshad Mobasher, Robert Cooley, and Jaideep [33] Osmar R. Zaiane, Man Xin, and Jiawei Han. Discover-
Srivastava. Automatic personalization based on Web ing Web access patterns and trends by applying OLAP
usage mining. Communications of the ACM, 43(8):142- and data mining technology on Web logs. In Proceed-
151, 2000. ings of Advances in Digital Libraries (ADL), pages 19-
[23] Bamshad Mobasher, Honghua Dai, Tao Luo, and Miki 29, Santa Barbara, California, April 1998.
Nakagawa. Discovery and evaluation of aggregate us-
age profiles for Web personalization. Data Mining and


Dm4

More Related Content

Viewers also liked

Similar to Dm4

Recently uploaded

Dm4