IRJET- A Novel Technique for Inferring User Search using Feedback Sessions

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 06 Issue: 04 | Apr 2019 www.irjet.net p-ISSN: 2395-0072
© 2019, IRJET | Impact Factor value: 7.211 | ISO 9001:2008 Certified Journal | Page 1959
A Novel Technique for Inferring User Search using Feedback Sessions
Suraj Gothankar1, Pooja Kharat2, Karishma Panen3, Prof. Harish Barapatre4
1,2,3Final Year student, Department of Computer Engineering, Y.T.I.E.T., Karjat, Maharashtra, India
4Assistant Professor, Department of Computer Engineering, Y.T.I.E.T., Karjat, Maharashtra, India
------------------------------------------------------------------------***-------------------------------------------------------------------------
Abstract:- Web search engines have become a resource
pool to gain information. We submit queries to search
engines to retrieve search results. These queries
basically depict information needs of users. But there are
situations in which these queries do not represent user’s
specific information needs as varied users have different
needs with respect to same query. There are many
search engines like Google, Yahoo, Amazon that retrieve
search results for users based on their ranking
algorithm. It should also be noted that while inferring
user search goals, fulfillment of user’s desire is very
important. So a Novel Technique to Infer User Search
Goal using Feedback Session is proposed. The basic step
of Inferring user search goals is generation of feedback
session through user click-through logs. Feedback
session distinguishes between what user requires and
what the user does not require. Keywords in the query
will determine whether a document satisfies the user’s
needs. The initial search results will later be restructured
based on users’ needs and manually Average Precision
would be used to examine the performance of users. The
main objective of this technique is to improve
information retrieval by generating feedback sessions
and rank the results list based on user’s domain of
interest. These re ranked results will be categorization
based which will be dependent on the keywords
generated.
Keywords: User search goals, feedback session,
Reranked search results, Average Precision,
Categorization.
I. INTRODUCTION
Web Mining is a thriving & expanding research field. It
combines the domain of Internet and Data Mining. The
Web Mining research encapsulates areas of Computer
Science such as Database Technology, Artificial
Intelligence and Information Retrieval. Web Mining is
classified as:
a. Web Structure Mining
b. Web Content Mining
c. Web Usage Mining.
Web structure mining gives much stress on discovering
how to model the underlying link structures of the Web.
Web content mining gives much emphasis on
Information Retrieval and Discovery of relevant
information from the Web. Web usage mining is a class
that predicts user’s behaviour on the web based on their
usage pattern.
Web is a large Dynamic, ever-changing, innovative field
that provides huge amount of information. Data present
on the web is increasing at a large extent and most of
data on web is fetched with the help of web crawler. To
look up information in web, users submit their queries to
search engines. But some queries do not represent user’s
specific information needs. The query maybe very broad
and varied users have different needs with respect to
same query.
There are many search engines like Google, Yahoo,
Amazon that retrieve search results for users based on
their ranking algorithm. The results are identified by
web users by going through the list that are
accompanied by title and snippets. It becomes very
exhausting task since multiple subtopics get mixed
together. Therefore, it becomes necessary and vital to
capture different user search goals in information
retrieval. These goals are defined as clusters of
information needs for a query. It is essential that a
search engine satisfies users need to obtain proper
information.
User search goals can be divided into three main
categories:
a. Query classification
b. Search result reorganization
c. Session boundary detection
In Query Classification, user goals are inferred by
predefining some specific classes on which classification
is performed. It becomes difficult and nearly impossible
to find an appropriate predefined search goals class as
users needs vary with respect to different queries. . In
the second class, search results are reorganized by
analysing the clicked URLs directly from user click-
through logs. Now some URLs of query are small in size.
This is also one of the main drawbacks of this class. [1]
However, here user feedback is not considered. This
results in noisy search results being analysed even if not
clicked by any user. Therefore, this technique is
imprecise in inferring user search goals. In the third
class, session boundaries are detected. Session
boundaries establish a common context based on user
sessions and frequency of user activities. But this
method only identifies whether a pair of queries belongs

to the same category and does not try to find the
semantics of goal.
Taking into consideration all the drawbacks of existing
system a novel technique is proposed that will infer user
search goals with feedback sessions and generate re
ranked search results that are categorized according to
predefined keywords.
II. RELATED WORK
A lot of work has been investigated on user search goals.
It is as below:
In Zheng Lu et.al proposed two methods for inferring
user search goals. The first method deals with clustering
of search results. In this method, a program is created
that submits the queries to search engine. First 100
results are crawled up including the title and snippets.
Then each search result is mapped to feature vector. The
search results are clustered by means of K-means
clustering and the most optimal K is selected by means of
Classified Average Precision Criterion. But in this
technique, search results are clustered. However, this
technique resulted in noisy URLs which were randomly
clicked by users also getting selected, thus degrading the
performance.
In the second method different clicked URLs are
clustered to infer user search goals. User click-through
logs are created by combining different single sessions.
These URLs, which were very few were selected along
with their titles and snippets. A feature vector is
assigned to every clicked URL. Finally, results are
clustered. But in this method, instead of feedback
sessions the clicked URLs are clustered. So the number of
URLs clicked differently is reduced. Thus, precise
segmentation becomes difficult.
In user search logs were generated from feedback
session and accordingly a framework was proposed. This
resulted in Pseudo-document being created which
represented users search interest. An innovative
approach known as Pattern Matrix was used. It consisted
of documents and patterns were calculated as inputs.
And later the documents were clustered. In this
technique a new Clustering method called as semantic
clustering was used to cluster documents. And by using
TF*IDF matrix user search goal text were analyzed.
Though TF*IDF is a good weighting algorithm, it
clustered documents that had similar keywords so it
identified near identical documents but it failed
miserably for all those topics that shared one keyword
even if they shared the same topic. Also for effective
semantic clustering, the clusters should be correctly
specified by the programmer because in this approach
the clusters were created on demand and were not
specified by the user.
K.N. Vimal Shankar et.al [3] proposed a framework in
which user profile was created for each individual and it
got authenticated by client. The user interest was
registered in the database and according to the change in
interest of user the database got updated. Feedback
sessions were generated and they were clustered
dependent on the keywords. The original search results
were then restructured and they were based on user
search goals. In in this approach, keywords that are also
known as Goal texts determined whether a document
could satisfy the user’s needs. But goal texts were latent
and were not expressed explicitly. That is why Pseudo-
documents were created. Feedback session tells what the
user requires and what the user does not require. It
depends on the number of clicked and un-clicked URL.
Certain situations may arise where the user might not
click some URLs if they appeared similar to previously
clicked ones. The “unclicked “URLs reduced the weight of
some terms in pseudo-documents which resulted in
incorrect results.
Based on the study of related work, the following gaps
were identified:
a. Some URLs might get skipped by users as they
appeared similar to previous ones. Also the unclicked
URLs could wrongly reduce the weight of some terms in
–pseudo-documents.
b. TF*IDF, a good weighting algorithm clustered
documents that had similar keywords but it failed
miserably for topics that shared one keyword even if
they shared same topic.
c. There were many noisy URLs in the search results that
were seldom clicked by users. When search results were
clustered with noisy ones its performance degraded
greatly.
d. In situation where the number of clustered URLs that
were clicked by user appeared small, to segment them
precisely was a difficult task.
III. PROPOSED FRAMEWORK
Based on the loopholes identified in the existing systems,
a novel technique is proposed in this paper that will infer
the user search goals by clustering the feedback sessions
and then perform categorization based re ranking with
the help of Keywords predefined in database. The
proposed system represented its block diagram as
shown below in Fig.1:

Fig.1. Block Diagram of proposed system
The components of proposed solution are as follows:
1. INPUT QUERY
For feedback session to be generated it is essential to
have user-click patterns. In this the user enters any
vague search term in search engine.
2. GENERATE USER-CLICK LOG
The user searches for a term in search engine like Google
and Yahoo. The links that are generated while searching
for the term will be stored in database. The logs would
be created through the users click pattern. The user click
log for a query is represented below in Fig.2:
Fig.2. User Click Log
3. CREATE FEEDBACK SESSION
In this module, a feedback session is created which
consists of both clicked and unclicked URLs which are in
a way part of single session. All the URLs before the last
click will be scanned and evaluated. The clicked URLs are
those which are needed by users. The unclicked URLS
are on the other hand those which are not needed but
evaluated once by users. So, the unclicked ones will also
become a part of user feedback.
Consider the following example wherein a feedback
session inside a single session is displayed below in Fig.
3a:
Fig.3. A Single session consisting of Feedback session
4. PRE-PROCESSING DATA
For each distinct query, in each feedback session the
following pre-processing steps are carried out-
a. Stop words Removal: Removes frequently occurring
keywords like a, an, the, on, over etc. from each pseudo-
document.
b. Non-words Removal: Removes non-words i.e. symbols
like @, #, %, etc. from each pseudo- document.
c. The extraction and removal of stem words like “ing”,
“ly” at the end of each word was done using Porter’s
stemming algorithm.
The Pseudo-document generated for one of the clicked
URL is shown below in Fig.4:
Fig.4. List of Pseudo-document
5. GENERATE KEYWORDS FROM RESTRUCTURED
WEB SEARCH RESULTS
In the above table, a “0” in the click sequence indicates
“unclicked URL” All the 10 URLs are used to construct
one session. The URLs in the box construct a feedback
session. As the URLs are scanned from top to bottom, not
only the four clicked URLs but also the six unclicked
ones are browsed and evaluated. The clicked URLs give

information about user’s needs and un -clicked ones tell
us which links to ignore. Thus it effectively analyzes
feedback session and infers search goals.
After the data is preprocessed the keywords would be
generated from restructured web search results. Top 10
queries would be selected from the application and the
links would be shown to the user. Finally based on the
user-click pattern the results will be stored in database.
And Feedback session would be represented in more
meaningful manner.
6. RERANKING SEARCH RESULTS
Every time the user performs click-through it will keep
on adding in database. If the user again performs similar
click- through there will be reranking performed. There
are many algorithms used in past for reranking like
Graph algorithm, Pattern Similarity Matching algorithm,
Relevancy algorithm.
The search results will be recorded. And the reranking
will also be performed based on users past history. The
restructured results is shown below in Fig. 5a :
Fig.5. Restructured results
7. CATEGORIZE RERANKED SEARCH RESULTS
For categorization, certain categories will be pre-defined
in database. Depending upon the Keywords generated
the reranked search results will be categorized. Again
there are many categorization algorithms used in past
like Deep Classifier, Search Result Record (SRR)
grouping, Support Vector Machine [8][9]. One out of
these algorithms can be used for categorization
whichever is found feasible and easy to understand by
the specified end –user.
IV. PERFORMANCE ANALYSIS
As user search goals are not defined, inference of user
goals becomes an important problem in Information
Retrieval. Thus, we need to use a metric that evaluates
the performance of user search goal inference
objectively and effectively. The novel criterion “Average
Precision” (AP) evaluates the restructured results. The
formulae for AP is displayed as equation (1),
∑
‒ AP is the Average of precisions computed in the
point of each relevant document in the ranked sequence
‒ N + is the number of relevant (or clicked)
documents in the retrieved ones, r is the rank
‒ N is the total number of retrieved documents
‒ rel () is a binary function on the relevance of a
given rank
‒ Rr is the number of relevant retrieved
documents of rank r or less.
This is the evaluation metric that will be compared
individually against the existing techniques like Bayesian
Classifier, Naïve Bayes Algorithm, and Bisecting K-means
that clusters both the clicked URLs and search results for
inferring user search goals.
V. CONCLUSION
In this work, we have presented a technique to infer user
search goals with clustering of feedback session. Because
previous works focused on user click feedback only for a
single session, as user may search different concepts at
different sessions [11]. Instead of relying on search
results or clicked URLs, feedback sessions are used to
infer user search goals. Feedback session are the URLs
list obtained after query submission [10]. The search
results are restructured and represented with some
keywords. A new criterion AP i.e. Average Precision is
used to evaluate the performance of user goals. The
reranked search results are categorized dependent upon
keywords. This further improves the performance of
underlying search engine and proves itself to be very
effective technique with better results to infer user
search goal.
VI. REFERENCES
[1] Zheng Lu , Hongyuan Zha , Xiaokang Yang , Weiyao
Lin ,Zhaohui Zheng, “A New Algorithm for Inferring User
Search Goals with Feedback Sessions”, IEEE
TRANSACTIONS ON KNOWLEDGE AND DATA
ENGINEERING, Vol. 25, No. 3, pp.502-512 ,March 2013.
[2] Shine Sonali Bhaskar, Prof. Bharat Tides, “A New
Approach and Compressive Survey on Restructuring
User Search Results by using Feedback Session”, IEEE
International Conference on Computing Communication

Control and Automation (ICCUBEA), pp.479-484,
February 2015.
[3] S .Preetha , K.N. Vimal Shankar, “Personalized Search
Engines On Mining User Preferences Using Click through
Data”,IEEE International Conferences on Information
Communication & Embedded systems (ICICES), pp. 01-
06 , February 2014.
[4] Ms. D. Indumathi ,Dr. A.Chitra, “A Collaborative
Search with Query Expansion and Result Re-ranking”,
IEEE World Congress on Information and
Communication Technologies, pp.986-989, December
2011.
[5] Dasari Amarendra, Kaveti Kiran Kumar, ”Inferring
User Search Goals with Feedback Sessions using K-
means Clustering Algorithm”, International Journal of
Computer Engineering In Research Trends (IJCERT)
Volume 2, Issue 11, pp. 780-784, November 2015.
[6] Jiafeng Guo, Lin Ding, Gang Zhang, Yue Liu, Xueqi
Cheng, “PSM: A new Re-ranking algorithm for Named-
page”, In Proc. of the Fifteenth Text Retrieval Conference,
TREC 2006, Gaithersburg, Maryland, pp.1-5 November
14-17, 2006.
[7] Antonis Sidiropoulos, Yannis Manolopoulos,
“Generalized comparison of graph-based ranking
algorithms for publications and authors”, The Journal of
Systems and Software 79, Department of Informatics,
Aristotle University, pp. 1679-1700, 16 January 2006.
[8] Sukanya S. Gawade , Gyankamal J.Chhajed, “Using
Feedback Sessions for Inferring User Search Goals“,
International Journal of Advanced Research in Computer
and Communication Engineering (IJARCCE),Vol. 3, Issue
6 , pp.7015-7019, June 2016.
[9] Dikan Xing, Gui-Rong Xue, Qiang Yang, Yong Yu1,
“Deep Classifier: Automatically Categorizing Search
Results”, In Proc. First ACM International Conference on
Web Search and Data Mining (ACM WSDM 08), Stanford,
California, USA, pp. 139-148, February 11-12, 2008.
[10] Febna V, Anish Abraham, “User Search Goals
Evaluation withFeedback Sessions“, International
Conference on Emerging Trends in Engineering , Science
and Technology (ICETEST), Procedia Technology (24),
pp.1256-1262, Available online at
www.sciencedirect.com.

IRJET- A Novel Technique for Inferring User Search using Feedback Sessions

More Related Content

What's hot (18)

Similar to IRJET- A Novel Technique for Inferring User Search using Feedback Sessions (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET- A Novel Technique for Inferring User Search using Feedback Sessions