Web mining: Pattern discovery from world wide web transactions

Tags: transaction, association rules, data mining, WEBMINER, transactions, sequential patterns, application, Web server, access log, Web mining, association rule, techniques, World Wide Web, data mining techniques, client, products, access data, association, Web analysis tools, unstructured data, data cleaning, log entries, data integration, Web servers, knowledge discovery techniques
Content: Web Mining: Pattern Discovery from World Wide Web Transactions Bamshad Mobasher, Namit Jain, Eui-Hong (Sam) Han, Jaideep Srivastava fmobasher,njain,han,[email protected] Department of Computer Science University of Minnesota 4-192 EECS Bldg., 200 Union St. SE Minneapolis, MN 55455, USA March 8, 1997 Abstract Web-based organizations often generate and collect large volumes of data in their daily operations. Analyzing such data can help these organizations to determine the life time value of clients, design cross marketing strategies across products and services, evaluate the e ectiveness of promotional campaigns, and nd the most e ective logical structure for their Web space. This type of analysis involves the discovery of meaningful relationships from a large collection of primarily unstructured data, often stored in Web server access logs. We propose a framework for Web mining, the applications of data mining and knowledge discovery techniques to data collected in World Wide Web transactions. We present data and transaction models for various Web mining tasks such as the discovery of association rules and sequential patterns from the Web data. We also present a Web mining system, WEBMINER, which has been implemented based upon the proposed framework, and discuss our experimental results on real-world Web data using the WEBMINER. Keywords: data mining, world wide web, association rules, sequential patterns, web mining. This work was supported in part by NSF grant ASC-9634719, Army Research O ce contract DA/DAAH04-951-0538, Cray Research Inc. Fellowship, and IBM partnership award, the content of which does not necessarily re ect the policy of the government, and no o cial endorsement should be inferred. Access to computing facilities was provided by AHPCRC, Minnesota Supercomputer Institute, Cray Research Inc., and NSF grant CDA-9414015. See http://www.cs.umn.edu/ han/papers.html#DataMiningPapers for other related papers.
1 Introduction and Background As more organizations rely on the World Wide Web to conduct business, traditional strategies and techniques for market analysis need to be revisited. Organizations collect large volumes of data and analyze it to determine the life time value of customers, cross marketing strategies across products, and e ectiveness of promotional campaigns. In the Web, such information is generally gathered automatically by Web servers and collected in server or access logs. Analysis of server access data can provide information on how to restructure a Web site for increased e ectiveness, better management of workgroup communication, and analyzing user access patterns to target ads to speci c groups of users. Most existing Web analysis tools Inc96, eSI95, net96] provide very primitive mechanisms for reporting user activity, i.e. it is possible to determine the number of accesses to individual les, the times of visits, and URLs of users. However, these tools usually provide little analysis of data relationships among the accessed les, which is essential to fully utilizing the data gathered in daily transactions. A comprehensive analysis tool must automatically discover such relationships among users accesses. In this paper we describe a framework for the application of two data mining techniques, i.e. discovery of association rules and sequential patterns to extract relationships from data collected by Web servers. Recently, several researchers have proposed the application of data mining techniques to facilitate information discovery on global information systems such as the Internet ZH95, KKS96]. The focus of these proposals is knowledge discovery across the Internet, based on content, and not the analysis of user access patterns on various Web servers. Web server access logs, however, have been used as a testbed for the application of certain data mining tasks such as the discovery of frequent episodes MTV95]. Recently, maximal forward references have been proposed CPY96] as a way to extract meaningful user access sequences. Web mining is the application of data mining techniques to large Web data repositories, examples of which are provided below. Discovering Association Rules: In of Web mining, an example of an association rule is the correlation among accesses to various les on a server by a given client. For example, using association rule discovery techniques we can nd the following correlations: (i) 60% of clients who accessed the page with URL , /company/products/ also accessed the page /company/products/product1.html (ii) 40% of clients who accessed the Web page with URL , /company/products/product1.html also accessed /company/products/product2.html and (iii) 30% of clients who accessed , /company/special-offer.html placed an online order in . /company/products/product1 In Web mining additional properties of data can be used to prune the search space, since information about a site's structural hierarchy can be used. For example, if the support for /company/products/ is low, one may conclude that the search for association between the two secondary pages with URLs /company/products/product1 and /company1/products/product2 should be pruned since neither are likely to have adequate support. Discovery of Sequential Patterns: Given a database of time-stamped transactions, the problem of discovering sequential patterns MTV95, SA96] is to nd inter-transaction patterns, i.e. the presence of a set of items followed by another item, in the time-stamp ordered transaction set. In Web server transaction logs, a visit by a client is recorded over a period of time. By analyzing this information, we can determine temporal relationships among data items such as: (i) 30% of clients who visited , /company/products/product1.html had done a search in Yahoo, within the past week on keywords w1 and w2 and (ii) 60% of clients who placed an online order in , /company/products/product1.html also placed an online order in /company1/products/product4 within 15 days. Another important kind of information we may be interested in is the common characteristics of all clients that visited a particular le within the time period t1 t2]. Alternatively, we may be interested in a time interval, i.e. an hour, day or week, during which a particular {1{
le is most accessed. As the examples above show, mining for knowledge from web log data has the potential of revealing information of great value. While this certainly is an application of existing data mining algorithms, e.g. discovery of association rules or temporal sequences, the overall task is not one of simply adapting existing algorithms to new data. Because of many unique characteristics of the client-server model in the World Wide Web, including radical di erences between the physical and logical data organizations of web repositories, it is necessary to develop a new framework to enable the mining process. Speci cally, there are a number of issues in pre-processing data for mining that must be addressed before the mining algorithms can be run. These include developing a model of web log data, developing techniques to clean/ lter the raw data to eliminate outliers and/or irrelevant items, grouping individual page accesses into semantic units (i.e. transactions), and specializing generic data mining algorithms to take advantage of the speci c nature of web log data. In this paper we present an architecture for a web mining system called the WEBMINER, and our initial experiences with it. Our speci c contributions include (i) development of a exible architecture for web mining, (ii) developing a model for a user transaction which consists of multiple log entries, (iii) Clustering algorithms for grouping log entries into transactions, (iv) adaptation of association rule and temporal sequence discovery algorithms to web mining, and (v) experimental evaluation of the system. The rest of this paper is organized as follows: Section 2 discusses the architecture of the Web mining process. Section 3 presents data and transaction models for Web mining. Section 4 we present the implementation of the WEBMINER, and also present our experimental results with the WEBMINER. Finally, in Section 5, we look at future and continuing Research and Development issues. 2 An Architecture for Web Mining A Web server access log contains a complete history of le accesses by clients. Most WWW access logs follow the Common Log Format speci ed as part of the HTTP protocol by CERN and NCSA Luo95]. A log entry, following this standard, contains the client IP address, user id, access time, request method, and the URL of the page accessed, the protocol used for data transmission, an error code, and the number of bytes transmitted. Table 1 shows a snapshot of a portion of the WWW access log for the main Web server in the Computer Science Department at the University of Minnesota. The primary objective of Web mining is to discover interesting patterns in accesses to various Web pages within the Web space associated with a particular server. In order to successfully apply generic data mining techniques to web data, one must rst transform this data to a suitable form. In particular, unlike market basket analysis, where a single transaction is de ned naturally according to customer purchase activity, in Web data the notion of a transaction must be de ned based on the properties of the application domain. Our proposed architecture, thus divides the web mining process into two main parts. The rst part includes the domain dependent processes of transforming the Web data into suitable \transaction" form, and the second part includes the, largely domain independent, application of generic data mining techniques (such as the discovery of association rule and sequential patterns) as part of the system's data mining engine. The overall architecture for the Web mining process is depicted in Figure 1. Generally, there are a variety of les accessed as a result of a request by a client to view a {2{
Figure 1: General Architecture for WEBMINER ... looney.cs.umn.edu han - 09/Aug/1996:09:53:52 -0500] "GET / mobasher/courses/cs5106/cs5106l1.html HTTP/1.0" 200 9370 mega.cs.umn.edu njain - 09/Aug/1996:09:53:52 -0500] "GET / HTTP/1.0" 200 3291 mega.cs.umn.edu njain - 09/Aug/1996:09:53:53 -0500] "GET /images/backgnds/paper.gif HTTP/1.0" 200 3014 mega.cs.umn.edu njain - 09/Aug/1996:09:53:53 -0500] "GET /images/misc/footer.jpg HTTP/1.0" 200 13355 mega.cs.umn.edu njain - 09/Aug/1996:09:54:12 -0500] "GET /cgi-bin/Count.cgi?df=CS-home.dat&dd=C&ft=1 HTTP/1.0" 200 646 mega.cs.umn.edu njain - 09/Aug/1996:09:54:18 -0500] "GET / advisor HTTP/1.0" 302 mega.cs.umn.edu njain - 09/Aug/1996:09:54:19 -0500] "GET / advisor/ HTTP/1.0" 200 487 looney.cs.umn.edu han - 09/Aug/1996:09:54:28 -0500] "GET / mobasher/courses/cs5106/cs5106l2.html HTTP/1.0" 200 14072 mega.cs.umn.edu njain - 09/Aug/1996:09:54:31 -0500] "GET / advisor/csci-faq.html HTTP/1.0" 200 13786 looney.cs.umn.edu han - 09/Aug/1996:09:54:47 -0500] "GET / mobasher/courses/cs5106/princip.html HTTP/1.0" 200 6965 moose.cs.umn.edu mobasher - 09/Aug/1996:09:55:50 -0500] "GET / suharyon/lisa.html HTTP/1.0" 200 654 moose.cs.umn.edu mobasher - 09/Aug/1996:09:55:53 -0500] "GET / suharyon/line/line16.gif HTTP/1.0" 200 1423 moose.cs.umn.edu mobasher - 09/Aug/1996:09:55:57 -0500] "GET / suharyon/joko1.jpg HTTP/1.0" 200 30890 ... Table 1: Sample Entries from a Web Server Access Log particular Web page. These include image, sound, and video les executable cgi les coordinates of clickable regions in image map les and HTML les. Thus, the server logs contain many entries that are redundant or irrelevant for the data mining tasks. For example, all the image le entries are irrelevant or redundant, since as a URL with several image les is selected, the images are transferred to the client machine and these les are recorded in the log le as independent entries. We call the process of removing redundant or irrelevant entries from the Web server log les data cleaning. Data cleaning is performed by checking the su x of the URL name. For instance, all the log entries with lename su xes such as, gif, jpeg, GIF, JPEG, jpg, JPG and map are removed from the log. After the data cleaning, the log entries must be partitioned into logical clusters that represent a single user transaction. One of the signi cant factors which distinguish Web mining from other data mining activities is the method used for identifying user transactions. This process, which we call transaction derivation is particularly important in the context of web data, because it allows the discovery phase to focus on relevant access points of a particular user rather than the intermediate {3{
pages accessed for navigational reasons. Recently, maximal forward references have been proposed CPY96] as a way to extract mean- ingful user access sequences. However, this method does not consider that some of the \backward" references may provide useful information or that not all forward references are meaningful for the discovery process. We prefer to cluster several user references (entries in the clean log le) into what would be considered a single user transaction. These transaction clusters are formed either statically based on a user speci ed time gap, or dynamically, according to time distance between entries, time spent on a particular page relative to its size, logical structure of the Web space, or other possibly application dependent criteria. The transaction model for WEBMINER is presented in the Section 3. As depicted in Figure 1, access log data may not be the only source of data for the Web mining process. User registration data, for example, is playing an increasingly important role, particularly as more security and privacy conscious client-side applications restrict server access to a variety of information, such as the client IP addresses or user IDs. The data collected through user registration must then be integrated with the access log data. While WEBMINER currently does not incorporate user registration data, we are exploring various data integration issues in the context of Web mining. For a study of data integration in databases see LHS+95]. Once the domain-dependent data transformation phase is completed, the resulting transaction data must be formatted to conform to the data model of the appropriate data mining task. For instance, the format of the data for the association rule discovery task may be di erent than the format necessary for mining sequential patterns. The Webmining data model for association rules and sequential patterns are described in the next section. This allows for the application of generic data mining algorithms to the Web transaction data. Finally, a query mechanism will allow the user (analyst) to provide more control over the discovery process by specifying various constraints. The emerging data mining tools and systems lead naturally to the demand for a powerful data mining query language, on top of which many interactive and exible graphical user interfaces can be developed HFW+96]. Most of the conventional query languages are constrained by a schema, but in our case, the data model does not fall in This category. Recently, several query languages have been proposed which are not constrained by a schema. DMQL HFW+96], UnQL BDS95, BDHS96] and Lorel QRY+95] fall into this category, and can be easily extended to query the World Wide Web. Some guidelines for a good data mining language were proposed in HFW+96], which among other things, highlighted the need for specifying the exact data set and various thresholds in a query. Such a query mechanism can provide user control over the data mining process and allow the user to extract only relevant and useful rules. In WEBMINER, we have implemented a simple Query mechanism by adding some primitives to an SQL-like language. This allows the user to provide guidance to the mining engine by specifying the patterns of interest. As an example, consider a situation where the user is interested in the patterns which start with URL A, and contain B and C in that order, this pattern can be expressed as a regular expression A B C . To see how this expression is used within a SQL-like query, suppose further that the analyst is interested in nding all such rules with a minimum support of 1 % and a minimum con dence of 90 %. Moreover, let us assume that the analyst is interested only in clients from the domain .edu, and only wants to consider data later than Jan 1, 1996. The query based on these parameters can be expressed as follows: SELECT association-rules(A*B*C*) FROM log.data WHERE date 960101 AND domain = edu AND support = 1.0 AND con dence = 90.0 {4{
This information from the query is used to reduce the scope, and thus the cost of the mining process. The development of a more general query mechanism along with appropriate Web-based user interfaces and visualization techniques are issues relevant to the future development of the WEBMINER. 3 A Transaction Model for Web Mining Unlike market basket analysis, where a market basket transaction is de ned as the set of items bought by a customer in a single purchase, there is no natural de nition of a user transaction in a web browsing scenario. For example, at one extreme we could consider each log entry to be a separate transaction, while at the other we could consider all the entries in the log made by a particular user as a single transaction. Since a transaction is used to model a useful unit of work, either of these choices is probably too extreme. We believe that any approach to group web log entries into transactions must use information about both the nature of data and the type of analysis to be done on it. We propose to use such information in a 2-step process. In the rst step we use clustering as a general approach to grouping web log entries into transactions. The clustering is based on comparing pairs of log entries and determining the similarity between them by means of some kind of distance measure(s). Entries that are su ciently close are grouped together. In the second step, we use information about the type of analysis and specialize the groups formed in step 1 into transactions suited to the speci c analysis. In the following, we rst describe our general clustering approach and its present implementation. Next, we show how information about the type of analysis, viz. association rule discovery and temporal sequence discovery, can be used to develop the respective transactions. 3.1 Clustering Log Entries into Groups Let L be a set of server access log entries. A log entry l 2 L includes the client IP address l:ip, the client user id l:uid, the URL of the accessed page l:url, and the time of access l:time. There are other elds in web log entries, such as the request method used (e.g., POST or GET) and the size of the le transmitted. For our present illustration however, we only focus on the elds listed above. In using clustering to determine the similarity of two log entries, i.e. whether they belong to the same group, distance metrics on many di erent attributes can be de ned. Determining an appropriate set of attributes to cluster on, and de ning appropriate distance metrics for them is an important problem, and is being addressed in our ongoing research. As an initial cut, we have used only the time dimension for clustering of log entries. This is because time is the principal dimension that captures the nature of web browsing, since this process is essentially sequential in nature. Further, as illustrated in de nition 1, the log entries are laid out on the temporal dimension, and a time window is used for the grouping, where successive entries in a group can be at most a maximum time gap t apart. In addition, we consider partition the log entries based on IP address and user ID, since we are interested in browsing patterns of individual users. De nition 1 A Log Entry Group (LEG) g is a triple: g =< ipg uidg f(l1g:url l1g:time) : : : (lmg :url lmg :time)g > where, for 1 k m, lkg 2 L, lkg:ip = ipg, lkg:uid = uidg, and lkg+1:time ; lkg:time t {5{
As an example consider the log entries of Table 1. If the user-speci ed maximum time gap is 1 minute, then the transaction belonging to the client njain (starting from time ), 09/Aug/1996:09:53:52 will be the URL set f= = = adviser = : . adviser csci ; faq htmlg
3.2 Extracting Association Transactions From LEGs
Determining association rules does not need any temporal information. Hence, in extracting association transactions from the log entry groups, we lter out the time information.
De nition 2 t =< ipg uidg
Afln1g:uasrslo:c:ia:tilomgn:utrral ngsa>ct(ionnotet
itshaattlrkgi+pl1e:teimxtera;ctlekgd:tifmroem
a
log t).
entry
group
g,
so
that
aabc(cUoevL)sese)=t,loTjwgfhtabjeUsereWthUSeURsR=LeLtt St=ogtfj2.TfalIlU1ltn:uaRosrtLslhote.:cr:ia:Wwtiloeomtrndd:uestr,rlantgnhe.seatWhscuteeipospnduoseprptoonfcreottuhtchneoetufnofwortermbfoUrsttiem(dWaesswS(diwtehtiotnthbheidnee the access log) that the URLs in U have been accessed by clients in one transaction. We can now fTiso(hrXametAaslnuellapYyasp)tsdo=sreaotc(gnsiXiaevot)eitf.onhtneThthhenrrueoreltusetihaloeisonsklXdoaofnaf=asn)dendxisapcYsrosevioissescsraiidioatnentglieonoaanfesntdrtuahaaslneessoofitconh(iraXetmtrhieogXnicvYoer=n)uns=t)lejetTxhitjYrs,eot,safohnwWodhlndeetbd.hreeFamoXlclironrneiunxlgade.WmsenXpSclee=,as)nthdiseYYdr,uewlenheeWdreSass.
f g =) =company=products= =company=products=product1:html 0:01 0:75 =company=products=product2:html
indicates that 75% of clients who accessed the \products" section of the Web site and connected to the page for \product1" in that section, also visited the page for \product2", and that this combination of events occurred in 1% of all transactions.
3.3 Extracting Temporal Transactions From LEGs
For the task of mining sequential patterns, we need to keep track of the access time for each of the URLs accessed within that transaction. Furthermore, we are interested in the activity of a particular client spanning the whole access log. We de ne a temporal transaction to be the set of all the URL names and their access times for the same client where successive log entries are within a user-speci ed time gap t. We also associate a unique time stamp with each such transaction. In other words, a transaction is simply an LEG with a computed time stamp.
De nition 3 A temporal transaction t is a 4-tuple t =< ipt uidt UTt time(t) > where
UTt =< (l1t :url l1t :time) : : : (lmt :url lmt :time) >
such that for 1 time(t) = max1
i
k m, m lit:time.
lkt
2
L,
lkt :ip
=
ipt,
lkt :uid
=
uidt,
lkt +1:time ; lkt :time
t, and
Let T be the set of all temporal transactions. For each transaction t =< ipt uidt UTt >2 T , we
call UTt the URL{Time Set (UT{Set) for t. A UT{Sequence is a list of UT{Sets ordered according
to transaction times. In other words, given a set T 0 = fti 2 T j1 i kg of transactions, a
UT{Sequence S for Given a client c
T 0 is: S with ip
a=d:a:n: dUuTsetkr
>, id
where time(ti) u, let Tc be the
< time(ti+1) for 1 set of all temporal
i k ; 1. transactions
involving that client. So, Tc = ft 2 T jipt = ip and uidt = ug: The UT{Sequence for Tc is a special
{6{
sequence, a client c.
Sc, In
called other
the client{sequence words,
for
c,
composed
of
all
the
UT{Sets
of
transactions
involving
Sc =< UTt1 UTt2 : : : UTtn >
where for 1 i n, ti 2 Tc.
De nition 4 A UT{Sequence A =< a1 a2 : : : an > is a subsequence of another UT{Sequence B =< b1 b2 : : : bm >, denoted A v B, if there exists integers i1 < i2 < : : : < in, such that a1 bi1 a2 bi2 : : : an bin.
Let ID be the set of all client ids (ip and uid pair) in the Web server access log. Then the support count for a UT{Sequence S, denoted (S), is:
(S) = jfScjc 2 ID and S v Scgj:
Intuitively, the support count for the UT{sequence S is the number of client sequences that support
S (i.e., include S as a subsequence). Now, given UT{sequences S = X Y (where denotes
cXon=sc)ateYna, twiohneroef
two sequences), the support s of
taheserquuleenXtia=ls)patYterisndceannebde
de as
ned (X
as an expression of the form Y )=jIDj, and the con dence
is de ned as (X Y )= (X). The task of Web mining for sequential patterns is to
nd all rules X =s) Y (among all UT{
sequences S = X Y ), where s is at least a given threshold and is at least another given threshold.
4 Experimental Evaluation In this section we describe the implementation of WEBMINER, a Web mining system based on the framework presented in the previous sections and our experience with it. We give some of the algorithms we have used in the implementation of techniques for discovering association rules and sequential patterns from Web data. We have run these algorithms on the access log of Cray Research Home Page located at http://www.cray.com for the experiment. The log contains about 520K entries corresponding to the requests made during May of 1996 and its size is about 56M Bytes.
4.1 Mining Association Rules Given the access log of Cray Research Home Page, we used max time gap of 10 minutes to have 44K transactions of size 670K Bytes. There were 3686 distinct URLs referenced in the transaction. Given the transactions, the problem of mining association rules is to generate all association rules that have support and con dence greater than the user-speci ed minimum support (called minsup) and minimum con dence (called minconf) respectively. We have used apriori algorithm presented in AS94] for mining association rules. We experimented with support between 0.05% and 1.0% and con dence of 80%. The experiment was performed on a SUN SPARC Station 5 with 70 MHz CPU, 32M Bytes main memory, 510M Bytes disk drive and SunOS Release 5.4. Table 2 shows some examples of the rules discovered. The rst two are singleton rules without an antecedent. Rule 1 shows that 1.23% of transactions contain the Liquid-cooled Cray T3E1 System home page, while rule 2 shows that 0.68% of transactions contain the Air-cooled Cray T3E Systems. This result is di erent from the results 1Trade Mark of Cray Research Inc.
{7{
Rule No. 1 2 3 4 5
Con dence(%) 100.00 100.00 82.83 90.00 97.18
Support(%) 1.23 0.68 3.17 0.14 0.15
Association Rules /PUBLIC/product-info/T3E/LC T3E.html /PUBLIC/product-info/T3E/AC T3E.html /PUBLIC/product-info/T3E =) /PUBLIC/product-info/T3E/CRAY T3E.html /PUBLIC/product-info/J90/J90.html /PUBLIC/product-info/T3E =) /PUBLIC/product-info/T3E/CRAY T3E.html / /PUBLIC/product-info/J90 /PUBLIC/product-info/T3E/CRAY T3E.html /PUBLIC/product-info/T90 =) /PUBLIC/product-info/T3E /PUBLIC/sc.html
Table 2: Some Examples of the Association Rules Discovered
from other tools that provide statistical summary on the access log. Other tools do not have the concept of transactions as de ned here. As a result, if one client accesses one site several times in a short span of time, simple statistics collection tools would report the hit several times. On the other hand, in our system, the hit would be recorded only once because all of the accesses belong to one transaction and the association transaction is de ned as a set of items. Rule 3 says that 82.83% of clients that accessed the URL , /PUBLIC/product-info/T3E also visited /PUBLIC/product-info/T3E/CRAY T3E.html which is under . /PUBLIC/product-info/T3E Rule 4, on the other hand, shows that 90% of clients that accessed and , /PUBLIC/product-info/T3E /PUBLIC/product-info/J90/J90.html also visited /PUBLIC/product-info/T3E/CRAY T3E.html. These two rules demonstrate that clients who access J902 home page and T3E top page visit T3E main home page (CRAY T3E.html) about 7% more than other clients who just accessed the T3E main home page. Based on the discussion of Section 1, we can already observe how these discovered rules can be useful. For example, the combinations of rules 3 and 4 in Table 2 might suggest that there is a portion of the content of the J90 page that encourages clients to go back and access the T3E page. By, moving or copying this portion to a higher level in the hierarchy (e.g., ) /PUBLIC/product-info/T3 we might be able to increase the overall support for rule 2 in the table. 4.2 Mining Sequential Patterns In the implementation of WEBMINER, the time gap de ning transactions for sequential patterns, de ned in section 3, is taken to be zero. Thus, UT-sequences are made up of singleton sets of items and each item is a URL accessed by a client in a transaction. We have modi ed the apriori algorithm to mine frequent sequential patterns. Note that the support for a pattern now depends on the ordering of the items, which was not true for association rules. For example, a transaction consisting of URLs ABCD in that order contains BC as an subsequence, but does not contain CB. So, all permutations have to be generated. This is even more computationally intensive than generating all combinations which has already been reported 2Trade Mark of Cray Research Inc.
{8{
Pattern No. 1 2 3 4
Support(%) 5.63 2.69 2.89 0.93
Sequential Patterns ! /PUBLIC/sc.html /PUBLIC/product-info/T3E/AC T3E.html ! /PUBLIC/sc.html ! /PUBLIC/product-info/T3E/AC T3E.html /PUBLIC/product-info/T3E/quotes.html ! /PUBLIC/sc.html ! /PUBLIC/product-info/T90/BAYER.html /PUBLIC/product-info/T90/T90apps.html ! /PUBLIC/product-info/RM/convergence9.html ! /PUBLIC/product-info/RM/images ! /PUBLIC/product-info/RM/hw7.html /PUBLIC/product-info/RM/unifying6.html
Table 3: Some Examples of the Sequential Patterns Discovered
to be di cult AS94]. For getting around this problem, we have changed the joining operation for candidate generation itself. In generating candidates of size k+1, instead of matching the rst k;1 elements of two frequent itemsets, we match the last k ; 1 elements of the rst frequent sequential pattern of size k with the rst k ; 1 elements of the second frequent sequential pattern of size k. For nding the support of candidate sequential patterns, the question is: Given a UT-sequence < A1 : : : An >, how can we determine if there exists a subsequence < B1 : : : Bk > within a certain time span T S? The important factor to consider is that the number of occurrences of that subsequence within the time span is insigni cant. The only thing that matters is whether the subsequence is present or not. Figure 2 gives the desired algorithm. The basic idea behind the algorithm is to keep the latest starting times of all the pre xes of the subsequence. The algorithm takes a transaction along with its time stamp, a subsequence, and a time span (T S) as input.
1. 2. 3.
Ifoniritfi(a(allilzresittipetmreems xiotifinmsuterbatsnoesq-au1cetniocne)mdaotcbheegs)inthen
4. 5. 6.
forpifr(e(iatxleltmoimtjheeo1rf ]sit=uebmtsresaqjnusientnimcseuebmis]eaqtucheensc)e t)hdeon
7. 8. 9. 10. 11.
ereinftdu(rperrptneureFrxnAtxitmLTimSeREelUajs/]Et/=] />s/upbr0sesue&bqx&suteiemqtnurecaeennjsnct-eoimt1fo]feuouni]nd-d. p. re
xtime last]
TS) then
Figure 2: Algorithm gen count We have performed an experiment to nd frequent sequential patterns from the access log of Cray Research Home Page using the above algorithm. For this experiment, we did not consider the con dence of the sequential patterns. Table 3 shows some examples of the sequential patterns discovered. Pattern 1 shows that 5.63% of the clients accessed the SuperComputing Systems home page (sc.html) followed by Air-cooled Cray T3E systems home page. Pattern 2 shows that 2.69% of the clients went on to check out the customer quotes on the T3E system after accessing the two URLs of
{9{
Pattern 1. Pattern 3 demonstrates the sequential pattern of clients accessing the Supercomputing Systems home page, the page containing a story on T903 technical solution for Bayer group and the page containing application vendor quotes, in that order. Discovered sequential patterns can be used to predict user actions or provide suggestions for restructuring a site. For example, if one of the patterns in Table 3 is not desirable, then direct links can be placed in corresponding pages in order to redirect client tra c patterns. Or, if a pattern has a high con dence, it may be useful to put certain kinds of information (e.g., advertisements) in the predicted path of clients corresponding to that pattern. 5 Conclusions and Future Directions In this paper we have presented a framework for Web mining, the application of data mining and knowledge discovery techniques to WWW server access logs. This framework includes a exible architecture distinguishing the domain speci c data transformation tasks from the generic data mining engine. We have also described WEBMINER, a system based on the proposed framework, and presented experimental results on real-world industrial data to illustrate some of its applications (Section 4). Currently, we are extending the implementation of WEBMINER to incorporate mechanisms for clustering analysis and discovery of classi cation rules. The query mechanism in WEBMINER will also be extended to include clustering and classi cation constraints. Also an important area of ongoing research is to nd better methods of clustering log entries into user transactions, including using criteria such as time di erential among entries, time spent on a page relative to the page size, and user pro le information collected during user registration. Another interesting area of future work involves the development of autonomous agents that analyze the discovered rules to provide meaningful courses of action or suggestions to users (for instance to make suggestions about modifying the organization of content within the Web space, to automatically prefetch Web les based on discovered rules and user pro les, or to present various users with dynamically generated content based on user patterns). Other areas of future work include developing a more exible query mechanism which can handle both pre-discovery queries (on the data) and post-discovery queries (on the rules) integrating user registration data with the access log data in the discovery process and extending the WEBMINER to process Web data distributed across several servers each collecting their own access log and user registration data. References AS94] R. Agrawal and R. Srikant. Fast algorithms for mining association rules. In Proc. of the 20th VLDB Conference, pages 487{499, Santiago, Chile, 1994. BDHS96] P. Buneman, S. Davidson, G. Hillebrand, and D. Suciu. A query language and optimization techniques for unstructured data. In Proc. of 1996 ACM-SIGMOD Int. Conf. on Management of Data, 1996. BDS95] P. Buneman, S. Davidson, and D. Suciu. Programming constrcuts for unstructured data. In Proceedings of ICDT'95, Gubbio, Italy, 1995. 3Trade Mark of Cray Research Inc. { 10 {
CPY96] M.S. Chen, J.S. Park, and P.S. Yu. Data mining for path traversal patterns in a web environment. In Proceedings of the 16th International Conference on Distrib uted Computing Systems, pages 385{392, 1996. eSI95] e.g. Software Inc. Webtrends. http://www.webtrends.com, 1995. HFW+96] J. Han, Y. Fu, W. Wang, K. Koperski, and O. Zaiane. Dmql: A data mining query language for relational databases. In SIGMOD'96 Workshop on Research Issues in Data Mining and Knowledge Discovery (DMKD'96), Montreal, Canada, 1996. Inc96] Open Market Inc. Open market web reporter. http://www.openmarket.com, 1996. KKS96] I. Khosla, B. Kuhn, and N. Soparkar. Database search using informatiuon mining. In Proc. of 1996 ACM-SIGMOD Int. Conf. on Management of Data, Montreal, Quebec, 1996. LHS+95] E. Lim, S.Y. Hwang, J. Srivastava, D. Clements, and M. Ganesh. Myriad: Design and implementaion of federated database prototype. Software { Practive & Experience, 25(5):533{562, 1995. Luo95] A. Luotonen. The common log le format. http://www.w3.org/pub/WWW/, 1995. MTV95] H. Mannila, H. Toivonen, and A. I. Verkamo. Discovering frequent episodes in sequences. In Proc. of the First Int'l Conference on Knowledge Discovery and Data Mining, pages 210{215, Montreal, Quebec, 1995. net96] net.Genesis. net.analysis desktop. http://www.netgen.com, 1996. QRY+95] D. Quass, A. Rajaraman, Y.Sagiv, J. Ullman, and J. Widom. Querying semistructured heterogeneous information. In International Conference on Deductive and Object Oriented Databases, 1995. Qui93] J. Ross Quinlan. C4.5: Programs for Machine Learning. Morgan Kaufmann, San Mateo, CA, 1993. SA96] R. Srikant and R. Agrawal. Mining sequential patterns: Generalizations and performance improvements. In Proc. of the Fifth Int'l Conference on Extending Database Technology, Avignon, France, 1996. ZH95] O. R. Zaane and J. Han. Resource and knowledge discovery in global information systems: A preliminary design and experiment. In Proc. of the First Int'l Conference on Knowledge Discovery and Data Mining, pages 331{336, Montreal, Quebec, 1995. { 11 {

File: web-mining-pattern-discovery-from-world-wide-web-transactions.pdf
Title: webmining.dvi
Published: Tue Mar 16 14:14:53 2004
Pages: 12
File size: 0.77 Mb


, pages, 0 Mb

Digital communications, 6 pages, 0.04 Mb
Copyright © 2018 doc.uments.com