CLASSIFICATION OF WEB LOG DATA TO IDENTIFY

INTERESTED USERS USING DECISION TREES

K. R. Suneetha, R. Krishnamoorthi

Bharathidasan Institute of Technology, Anna University

[email protected]
[email protected]

ABSTRACT
Focus of many industries change towards customer orientation to retain regular frequent
accessors for the improvement of customer relationship management. Study of interested
web users, provides valuable information for web designer to quickly respond to their
individual needs. The aim of this paper is, instead of tracking the behavior of overall
users to redesign the web site, our model extracts only focused group of interested users.
The proposed model consists of two phases. In first phase, the web server log data is
preprocessed. The purpose of data preprocessing is to extract useful data from raw web
log. In the second phase data is classified using enhanced version of Decision tree
algorithm C4.5. NASA web server data is used for experimental purpose, which results
in less execution time and reduced memory utilization with high accuracy.

Keywords: Web Usage Mining, Preprocessing, Decision Tree, Frequent accessors.

1 INTRODUCTION users activities in e-service systems. It includes i) Web

Content Mining- extracting knowledge from the
The World Wide Web (WWW) is rapidly emerging content of documents, ii) Web Structure Mining-
as an important communication means of information inferring knowledge from Internet links iii) Web Usage
related to a wide range of topics (e.g., education, Mining-extracting interesting patterns from web
business, government). It has created an environment access logs. Our work focuses on Web Usage Mining.
of abundant consumer choices, where organizations Web Usage Mining (WUM) [2],[3] discovers usage
must give importance to improve customer loyalty. To patterns from web data, to better serve the needs of
reorganize a website in terms of structuring links and web-based applications. Three different tasks of usage
attractive design of web pages, organizations must mining are Data Preprocessing, Pattern Discovery and
understand their customers behavior, preferences, and Pattern Analysis. Data preprocessing consists of data
future needs. This imperative leads many vendors to cleaning, data integration, data transformation and path
develop more e-service systems for data collection and completion. Pattern discovery techniques used to
analysis. The navigation patterns of users generally extract knowledge from preprocessed data. Some of
gathered by the web servers and stored in server access the techniques used in pattern discovery are
logs. Analysis of server access log data provide Association rules, Classification, Clustering,
information to restructure a web site to increase Sequential Patterns, etc. Pattern Analysis filters out
effectiveness, better management of work group uninteresting rules or patterns from the set found in the
communication, and to target ads to specific users. pattern discovery phase.
Web mining [1] is a popular technique for analyzing This paper focuses on extracting interested user
group from web log data. Due to large amount of applied on this group instead of considering overall
irrelevant information in the web log, the original log entries, which intern improves the performance.
cannot be directly used in the web log mining
procedure. Thus in the first phase the web server log 3 PROPOSED MODEL
data is preprocessed, to extract useful data and then to
map these data in to the abstract data necessary for Finding information relevant to user needs has
pattern discovery. In this step, the original size of the become increasingly important now a days, and how
database will be reduced. In second phase, data is efficiently and effectively retrieve the `required web
categorized using decision tree depending on splitting pages are becoming more challenge. This issue is
attributes. addressed in our approach and is focused on
The rest of the paper is organized as follows. In discovering frequent accessors behavior which helps in
section 2 we discuss the related work. The proposed efficient organization of the web site. The Fig. 1 shows
approach and its details are presented in section 3. the building blocks of the entire process.
Section 4 discusses decision tree algorithm C4.5 and
generation of decision rules. Section 5 presents the
results. Finally, conclusions are drawn in section 6.

2 RELATED WORK

In recent years abundant work has been carried out

in the area of web mining, specifically on analysis of
web log data of Web Usage Mining.
The paper [4] describes improved statistical-based Figure 1: User Grouping Model
time oriented heuristics for the reconstruction of user
The raw data should be cleaned to eliminate
sessions from a server log. The authors Bamshad
irrelevant information from the original log file and to
Mobasher et.al., [5] presents transaction data models
form the Web log database convenient for user
for various web mining tasks such as the discovery of
identification, session identification. Then the
association rules and sequential patterns from the Web
Classification model classifies preprocessed data by
data. In paper [6] Cooley et.al presented methods for
considering number of factors like time stamp, request
user identification, sessionizing (i.e. constructing or
type, number of pages referred and depth wise analysis
reconstructing sessions), and page view identification.
as the attributes and predicates.
In another work [7], the authors highlights the
importance of analyzing users web log data and
extracting their interests of web-watching behaviors 3.1 Preprocessing Module
The raw data should be cleaned to eliminate
and describes a method for clarifying users interests
irrelevant information from the original log file and to
based on an analysis of the site-keyword graph. The
form the Web log database convenient for user
papers [8],[9] discusses importance of decision trees in
identification, session identification This module
web log mining, the authors suggested several
removes useless requests from the log files. Since all
hypotheses to improve web sites retention and showed
the log entries are not valid, we need to eliminate the
that it is possible to predict future user actions with
irrelevant entries. Usually, this process removes
reasonable misclassification error as well as to find
requests concerning non-analyzed resources such as
combinations of sequential pages resulting in browsing
images, multimedia files, and page style files. Data
termination. Swamy Kutti et.al., [10] discuss the
cleaning includes the elimination of irrelevant entries
importance of navigation patterns of users. It provides
like:
information about the major aspects and problems
• Entries with unsuccessful HTTP status codes;
related to the task of modeling the user behavior. It
also gives information about recent developments HTTP status codes are used to indicate the success
related to automatic web navigation, implicit capturing or failure of a requested event, and we only consider
of user interests and future directions on web log successful entries with codes between 200 and 299.
analysation. • Entries with request methods except GET and
A number of research work concentrated on POST.
applying data mining techniques on to preprocessed Once the data is preprocessed, user identification is
web log data to identify frequent patterns. But our done based on the attributes of IP addresses, time spent
model tries to form well focused data of interested on that particular session and also the agent or
users and then frequent pattern mining algorithm is operating system which is used as one of the parameter
to identify unique users. These users are classified as The information gain measure C4.5 algorithm is
interested and not interested users using Decision tree used to select the test attribute at each node in the tree.
algorithm. Such a measure is referred to as an attribute selection
measure of the goodness of split. The attribute with the
3.2 Decision Trees highest gain is chosen as the test attribute for the
Decision trees are used in classification and current node. This attribute minimizes the information
prediction. It is simple yet a powerful way of needed to classify the samples in the resulting
knowledge representation. The models produced by partitions. Such an information-theoretic approach
decision trees are represented in the form of tree minimizes the expected number of tests needed to
structure. A leaf node indicates the class of the classify an object and guarantees that a simple (but not
examples. The instances are classified by sorting them necessary the simplest) tree is found. The information
down the tree from the root node to leaf node. gain measure and gain ratio criterion is provided
In our work we have used enhanced version of further.
decision tree C4.5 algorithm. The reason for selecting
C4.5 decision tree algorithm is ability to handle data 4.1 Information Gain
with missing values. It also avoids over fitting the data Let S be a set of training set samples with their
and reduce error pruning. The C4.5 algorithm is corresponding labels. If there are m classes and the
described in the following section. training set contains si samples of class I and s is the
total number of samples in the training set. Expected
4 ENHANCED C 4.5 ALGORITHM information needed to classify a given samples is
calculated by:
Input: training samples, represented by discrete m
attributes; the set of candidate Attributes, attribute-list. I (S1, S2…….Sm) = ∑ (Si/S) log2 (Si/S) (1)
Output: set of classes i =1
Method: A feature F with values { f1,f2,…fv} can divide the
1. Create a node N; training set into v subsets {s1,s2,…Sv} where sj is the
2. If samples are all of the same class C, then subset which has the value fj for feature F.
Return N as a leaf node labeled with the class Furthermore let Sj contain Sij samples of class i.
C; Entropy of the feature F is
v
3. If attribute list is empty then Return N as a E(F)= ∑ (S1j + ……..+Smj)/ S x I(S1j……Smj) (2)
leaf node labeled with the most common
class in samples (majority voting) J=1
Information gain for F can be calculated as:
4. Select test attribute, the attribute among
attribute-list with the highest information gain
ratio; Gain (F) = I (S1, S2…., Sm) – E(F) (3)
In our experiment, information gain is calculated
5. Label node N with test-attribute; for class labels by employing a binary discrimination
6. For each known value ai of test-attribute for each class. That is, for each class, a dataset instance
is considered in-class, if it has the same label: out-
7. Grow a branch from node N for the condition class, if it has a different label. Consequently, as
test-attribute= ai; opposed to calculating one information gain as a
general measure on the relevance of the feature for all
8. Let si be the set of samples in samples for classes, we calculate an information gain for each
which test-attribute = ai; class. Thus this signifies how well the feature can
discriminate the given class from other classes.
9. If si is empty then
10. Attach a leaf labeled with the most common 4.2 Gain ratio Criterion
The notion of information gain introduced earlier
class in samples; tends to favor attributes that have a large number of
11. Else attach the node returned by generate- values. For example if we have an attribute D that has
decision-tree a distinct value for each record, then Info (D, T) is 0,
thus Gain (D, T) is maximal. To compensate for this, it purchase interest and those who simply visit the site.
was suggested to use the following ratio instead of Generally interested users exhibit certain access
gain. patterns; they access certain web pages for a rather
SplitInfo is the information due to the split of T on long time because they need time to spend on its
the basis of the value of the categorical attributes D, contents. The ratio between the time spent on content
which is defined by reading and the amount of time they navigate is large.
The interested users often use the HTTP POST mode,
because they are interested in registering with web
n │Ti│ │Ti│ (4)
sites and are willing to fill out forms with their own
SplitInfo(X) = -∑ xlog2 information. The user who does not have interest
simply accesses many pages quickly to browse
i=1 │T│ │T│
contents. These users do not often use POST method
and Gain Ratio is then calculated by because they are not interested in registering at web
sites. The web log files are classified as Interested
Gain (D, T) (5)
users (IU) and Not Interested Users(NIU) based on the
GainRatio (D, T) = parameters Time Stamp (30 min), method used(GET/
POST ), number of pages referred(min 5pages.) and
SplitInfo (D,T)
depth wise reference by the user.
The gain ratio, expresses the propotion of useful Fig.2 shows the construction of decision tree by
information generation split, i.e., that appears helpful considering the above mentioned attributes and
for classification. If the split is near trivial, split decision rules are generated according to the
information will be small and this ratio will be conditions.
unusable. To avoid this, the gain ratio criterion selects
a test to maximize the ratio above, subject to the
constraint that the information gain must be large, at
least as great as the average gain over all tests
examined.
Using the decision tree algorithm of C4.5 and
measures used to select test attribute, the generation of
decision rules generation is discussed in the next
section

4.3 Decision Tree Generation

The set of attributes (A1…A5) used to recognize
users interest, their behavior (Table 1).The basic idea
of our approach is to identify a small training data set
in the beginning itself, so that we can achieve high
accuracy for the most important class and user with
purchase interest.

Table 1: Attributes Used For Classification.

Attri- Description
butes
A1 total session time:15-30 min.
Figure 2: Decision Tree Generation
A2 total time a user stays at the site: > 30secs
A3 total number of accessed pages during the IU: Interested Users
whole session: >5pages NIU: Not Interested Users
A4 access methods used to interact with the
site: GET, POST Decision Rules:
A5 Depth Wise Access from particular Rule1: If (Time spent <30 and No. of pages referred
page(DWA) <5and Method used GET and
DepthwiseReference=’NO’) = “NIU”
We subjectively identify users who have some
Rule2: If (Time spent <30 and No. of pages referred <5 users and interested users with Data Base Size. Fig. 5
and Method used GET and indicates variation between the entries of unique and
DepthwiseReference=’YES’) = “IU” interested users with sessions. Percentage of reduction
Rule3: If (Time spent <30 and No. of pages referred <5 in Data Base size of interested users with unique users
and Method used POST) = “IU” is shown in Fig. 6. This shows a small training dataset
Rule4: If (Time spent <30 and No. of pages referred >5 (interested user group) achieve high accuracy for the
and Method used GET and most important class of users with purchase interest.
DepthwiseReference=’NO’) = “NIU”
Rule5: If (Time spent <30 and No. of pages referred >5
Entrie s and IP addre ss Ratio
and Method used GET and
DepthwiseReference=’YES’) = “IU” 10 0 0 0 0
Rule6: If (Time spent <30 and No. of pages referred >5 90000

and Method used POST) = “IU” 80000

Entries Vs IP Address
Rule7: If (Time spent >30 and No. of pages referred <5 70 0 0 0

and Method used GET and 60000

To t al
DepthwiseReference=’NO’) = “NIU” 50 0 0 0 Ent ries
40000 To t al IP
Rule8: If (Time spent >30 and No. of pages referred <5 30000
Ad d res s

and Method used GET and 20000

DepthwiseReference=’YES’) = “IU” 10 0 0 0
Rule9: If (Time spent >30 and No. of pages referred <5 0
and Method used POST) = “IU” 1 2 3 4 5 6 7

Rule10: If (Time spent >30 and No. of pages referred Days

>5 and Method used GET and

DepthwiseReference=’NO’) = “NIU”
Rule11: If (Time spent >30 and No. of pages referred Figure 3: Total number of unique users after
>5 and Method used GET and preprocessing
DepthwiseReference=’YES’) = “IU”
Rule12: If (Time spent >30 and No. of pages C omparati ve Study
referred >5 and Method used POST) = “IU”
3500

We have grouped all these classes as “Interested 3000

Users” and “Not Interested Users”. This grouping 2500

helps to choose only focused set of data for 2000

identification of user behavior, user interest etc. Thus DB Siz e Us ers
DB Size
1500
instead of analyzing whole web log file content, which Int. Us ers

in terms results in consumption of more memory and 1000 DB s ize

unnecessary processing of unwanted data, our method 500

gives preference to the group of interested users for 0

1 2 3 4 5 6
further behavior analysis which reflects in faster Days
execution and less space utilization. The details of
experimentations are presented in the next section.

5 EXPERIMENTAL RESULTS Figure 4: Comparative Study of users with interested

users with Data Base Size
NASA server web log file of 695MB data is
considered in our work. For the purpose of testing 10
days of data is collected (68MB). After preprocessing,
the data base size is reduced to 18 MB that is 60 to
70% by eliminating unnecessary data sets (see Fig. 3).
From the preprocessed data 1,34,937 records were
considered and from which, we obtained 10,119 unique
users of 12,857 sessions as a training data set. C4.5
algorithm is applied on its data set and 2618 entries are
classified as entries of interested users. The graph
shown in Fig. 4 gives comparative study of unique
 3000
in time and memory utilization. More effective
association rules can be generated from the well
2500
focused group formed by our work to find frequently
accessed patterns, which results in less time
2000 consumption and less memory utilization with high
Unique users accuracy.
Sessions
Ent ries 1500
Interested Users
Sessions
7 REFERENCES
1000

[1] H. Blockeel and R. Kosala: Web Mining

500
Research-a survey, SIGKDD Explorations.,
Volume 2, Issue 1, pages 1-15 (2000).
[2] Hui-You Chang and Feng Zhang: Research and
0

Development in Web Usage Mining System – key

Figure 5: Observation of variation between sessions issues and proposed solutions, Proceedings of the
of unique and interested users First International Conference on Machine
Learning and Cybernetics., pp.986-990 ( 2002).
2500
[3] Jaideep Srivastava, Prasanna Desikan, and Vipin
Comparative Study Kumar: Web Mining – Accomplishments and
Future Directions, Technical Report Computer
2000
Science Department, University of Minnesota,
Minneapolis, USA. pp. 51-6 (2002).
1500 [4] Jie Zhang Ali and A. Ghorbani: The
Unique users Entries Reconstruction of User Sessions from a Server
Interested Users Log Using Improved Time-oriented Heuristics,
Reduced % 1000
t Proceedings of the Second Annual Conference on
Communication Networks and Services Research,
500 IEEE (2004).
[5] Namit Jain, Bamshad Mobasher, Eui_Hong
0
_Sam_ Han and Jaideep Srivastava: Web Mining
5 4 3 2 1
Pattern Discovery from World Wide Web
6
9 8 7
Days Transactions, Technical Report, Department of
Computer Science, University of Minnesota
(1996).
Figure 6: Percentage of reduction in data base of [6] Robert Cooley, B. Mobasher, and Srivastava: Data
interested with unique users. Preparation for Mining World Wide Web
Our approach is well suited for application like e- Browsing Patterns, Knowledge and Information
Commerce and e-Business in identifying actionable System., Volume 1, No 1, pp. 5-32 (1997).
information to acquire new customers and to retain [7] Tsuyoshi Murata and Kota Saito: Extracting Users
old customers. Interests from Web Log Data, Proceedings of the
IEEE/WIC/ACM International Conference on
6 CONCLUSIONS Web Intelligence, Hong Kong, China. pp.343-346
(2006).
The work focused on group of the frequently [8] V Singh, Alsabti K and Ranka S: CLOUDS: A
accessed patterns of interested users. It assists the web Decision Tree Classifier for Large Datasets, In
site designers to improve the performance of the web Proceedings of KDD 98, 4th International
by giving preference to the patterns navigated by the Conference on Knowledge discovery and Data
regular interested users. Step1 removes unnecessary Mining, New York City, pp. 2-8 (1998).
and junk requests from the log file. The enhanced [9] Zidrina, Pabarskait: Decision trees for web log
version of decision tree C4.5 algorithm is applied for mining, Intelligent Data Analysis., Volume 7,
identification of interested users from web server log Issue2, pp. 141-154 (2003).
file. Instead of tracking the behavior of overall users [10] Zuhoor A1-Khajri, Swamy Kutti, Muna Hatem
(interested or not interested) to redesign the web site, and Haider AL Khajri: A Classification Technique
our approach generates the well focused set of for Web Usage Analysis, Journal of Computer
interested users. The results showed the improvement Science, ISSN Publication, pp. 413-418, (2005).