ISSN (Print) : 0974-6846
Indian Journal of Science and Technology, Vol 9(29), DOI: 10.17485/ijst/2016/v9i29/91060, August 2016 ISSN (Online) : 0974-5645
Performance Comparison of Different Similarity
Measures for Collaborative Filtering Technique
K. G. Saranya*, G. Sudha Sadasivam and M. Chandralekha
Department of Computer Science and Engineering, P. S. G College of Technology, Coimbatore - 641004,
Tamil Nadu; India; saranyaa87@gmail.com, sudhasadhasivam@yahoo.com, rdbmchandralekha@gmail.com
Abstract
Objectives: As the plenty of Web services on the Internet increases, developing efficient techniques for Web service
recommendation has become more significant. The main objective of this paper is to compare and study the drawbacks of
the performance of different existing similarity measures against the proposed similarity measure that use the concept of
collaborative filtering technique. Methods/Analysis: Collaborative filtering has turned into one of the most used technique
to give personalized services for users. The key of this technique is to find alike users or items using user-item rating matrix
such that the system can show recommendations for users. Experiments on Web Service (WSDL) data sets are conducted
and compared with many traditional similarity measures namely Pearson correlation coefficient, JacUOD, Bhattacharyya
coefficient. The result shows the superiority of the proposed similarity model in recommendation performance. Findings:
However, existing approaches related to these techniques are derived from similarity algorithms, such as Pearson
correlation coefficient, mean squared distance, and cosine. These methods are not much efficient, particularly in the cold
user conditions. Applications/Improvement: This paper presents a new user based similarity calculation model to
enhance the recommendation performance and to estimate the similarities for each user. The proposed model incorporates
two traditional similarity measures namely Pearson Correlation Coefficient and Jaccard Coefficient.
Keywords: Collaborative Filtering, Recommendation System, Similarity Measures, Web Service
1. Introduction The collaborative filtering has become the most
frequently used method to suggest items for users. It
In recent days people have their own smart phones, tablet makes suggestion in accordance to similar users with the
PC’s and other handy terminals like palmtops and so they active user or the similar items with the items which are
spend more time in surfing all kinds of social networking rated by the active user. The collaborative filtering includes
media (such as G+, Facebook, etc.) and e-commerce sites model-based method and memory-based method. The
(such as Myntra, Flipkart, etc). The voluminous informa- model-based method first defines a model to explain the
tion available makes them overwhelmed and indecisive. interest of users and, consequently to forecast the ratings
Users spend much time and energy in probing for their of items. The memory-based method first defines the
anticipated information. Still, they do not get acceptable similarities among users and then selects the most similar
outcomes. Luckily, the user preferences can be recorded users as the neighbors of the user to make recommen-
for latter reference on the social networking sites and dation. Finally, it gives the suggestions according to the
e-commerce sites, which makes easier to study the behav- neighbors. The memory-based method gives considerable
ior of users. Recommender systems are used to suggest recommended precision, but the computing time grows
information of user anticipations and offer personalized rapidly with the increasing number of items and users. In
services by analyzing the user’s behaviors, for instance, some circumstances, it is hard to take action in real-time.
the recommendation of the products in Amazon, photos The model-based technique tends to be faster in predic-
in Flickr, and results in the query based Web search. tion time than the memory-based technique, because
*Author for correspondence
Performance Comparison of Different Similarity Measures for Collaborative Filtering Technique
the creation of the model can be completed in a con- Advantages of this system are solves the data sparse
siderable amount of time and this technique is executed problem and improves consistency. Disadvantage is that
off-line. The limitation of the model-based technique is it is computed using less number of datasets. A tech-
that the recommendation performance is not as good for nique that does Qos prediction, Time-aware and Web
the memory-based technique. In addition to collabora- service has been proposed2. The similarity measure used
tive filtering, semantic recommendation, content-based is adjusted Cosine-based similarity. Advantages of this
technique, social recommendation are also applied in system are to improve prediction accuracy and missing
prediction of user preference. value prediction for QoS. Disadvantages are it does not
This paper concentrates on the recommended per- include many QoS factors and relationships among QoS
formance in memory-based collaborative filtering factors into consideration and it does not incorporate
algorithms. The core of collaborative filtering technique is QoS factors into QoS prediction. A technique that does
to compute similarities among users or items. The generic Greedy Filtering, K-nearest neighbor graph and Fast
traditional similarity measures, such as Pearson correla- collaborative filtering has been proposed3. The similar-
tion coefficient, mean squared distance, and cosine are ity measures used are Pearson Correlation Coefficient
not enough to capture the effective similar users, particu- (PCC) and adjusted Cosine similarity. Advantages of
larly for cold user who only rates a small number of items. this system are it decreases the execution time and
This paper presents a better heuristic similarity measure improve the recommendation quality. Disadvantage is
model. The new similarity model incorporates two simi- that it is computationally expensive. A technique that
larity measures namely Pearson Correlation Coefficient does Novel approach has been proposed4. The similarity
and Jaccard Coefficient. In order to evaluate fy the new measure used is Pearson Correlation Coefficient (PCC).
similarity measure, experiments are conducted on web An advantage of this system is predictions are of high
service data set. In comparison with many state-of-the- precision. Disadvantages are that if the number of users
art similarity measures, new model can show improved and items become huge, a huge amount of time will be
recommended performance and uses the better ratings in consumed. A technique that does K-nearest neighbor-
cold user conditions. hood and K-means has been proposed5. The similarity
Collaborative Filtering (CF), as a category of person- measure used is Pearson Correlation Coefficient (PCC).
alized recommendation method, has been commonly Advantages of this system are it improves the accu-
used in variety of domains. Though, collaborative filter- racy, improves lower time consuming level, solves the
ing suffers from a few issues, like cold start, data sparsity, cold start issue, and solves time and space complexity.
scalability problems. These issues seriously lessen the Disadvantage is that it is difficult to access user profile.
user experience. This paper concentrates on how to get A technique that does Novel algorithm has been pro-
the better prediction accuracy. Collaborative filtering rec- posed6. The similarity measure used is cosine similarity.
ommends items to users according to their preferences. Advantages of the system are it is more robust and it
Therefore, the past database of users’ preference must be improves prediction accuracy. Disadvantages are it has
available. However, the database is always very sparse, incomplete comparisons with the previous methods and
that is, user only rates a lesser number of items. Up to the analysis of the fusion model is reduced. A technique
now, there are many researchers who have focused on the that does QoS-aware ranking-oriented hybrid Web ser-
prediction accuracy and proposed some solutions. vice recommendation approach has been proposed7. The
To improve the precision, many researchers have similarity measure used is Pearson Correlation Coefficient
proposed some new similarity measures. A technique (PCC). Advantages of this system are it has higher accu-
that does Recommendation System and Collaborative racy rate, predicts the missing QoS values in a given
Filtering has been proposed1. With respect to the data dataset and improves interpretability. Disadvantages are
sparsity issue, the approach of user-item based collab- it is computationally expensive and involves more math-
orative filtering algorithm is proposed along with an ematical formulas. A technique that does Context-aware
iterative technique and a three step updating algorithm approach, which is a cloud based mobile multimedia has
to form a constant recommendation scheme. Finally, been proposed8. The similarity measure used is Pearson
based on the scalability of the neighborhood size, cosine Correlation Coefficient (PCC). Advantages of this sys-
similarity is used to calculate a similarity among users. tem are it is used to develop a real-world applications
2 Vol 9 (29) | August 2016 | www.indjst.org Indian Journal of Science and Technology
K. G. Saranya, G. Sudha Sadasivam and M. Chandralekha
and it improves services provided by service provid- 2. The New Similarity Model
ers. Disadvantage is that it is restricted to the relatively
small datasets. A technique that does DBSCAN cluster- This section presents the drawbacks of the existing
ing algorithm9 is used to perform a clustering on set of similarity measures. Then, it introduces the motivation and
items, and then obtains the user’s prediction rating of the hypothesis of the proposed similarity measure approach.
target item using weighted slope one scheme. The simi- Finally, this paper presents the mathematic formaliza-
larity measures used are Pearson Correlation Coefficient tion of the proposed novel similarity measure approach.
(PCC) and Cosine similarity. Advantages of this system In this system, Web service users are represented as
are it improves accuracy, solves the problem of sparsity, m represents the total number
scalability and cold start and it is more robust to noise. of users; set of web service items with similar function-
Disadvantage is that it considers only limited number ality are represented as where n
of datasets. A technique that does TrustSVD, a trust- represents total number of web services.
based matrix factorization method has been proposed10.
This technique considers both explicit and implicit rat- 2.1 The Disadvantages of Existing Similarity
ings and trust information during rating prediction on Measures
unknown items. A weighted- -regularization technique
was adapted and used to further regularize the user- and The Pearson Correlation Coefficient (PCC), Jaccard
item-specific latent feature vectors studied and the simi- Uniform Operator Distance (JacUOD) and Bhattacharyya
larity measure used is Pearson Correlation Coefficient. coefficient are the most widely used similarity measures
Advantages of this system are it solves the problem of in collaborative filtering.
data sparsity and cold start and it is outperformed in
predictive accuracy. Disadvantage is that it does not 2.1.1 Pearson Correlation Coefficient (PCC)
consider the influence of trusters and trustees. A tech- In many recommendation systems, Pearson Correlation
nique that does Ensemble Method has been proposed11. Coefficient (PCC)15 measure has been applied to compute
The similarity measures used are modified Pearson the similarity between the users and items. PCC measure
Correlation Coefficient (PCC) and modified Cosine- based similarity between two users is computed using the
based similarity measures. Advantages of this system are following formula
it has less computational cost and linear space complex-
ity, running time complexity. Disadvantage is that it does (1)
not focus on the application of the ensemble methods. A
technique that does Behavior Factorization has been pro- where represents co-invoked set of web services
posed12. The similarity measure used is Jaccard similarity. by user v and u. and represent the average
Advantage of this system is it improves the performance. QoS values of all the web services invoked by user u and
Disadvantages are it does not work well when users have v respectively. The similarity value computed using above
very sparse or no data and it concentrate on social media equation falls within the range of [-1, 1]. The larger the
platform like Google+. A technique that does the com- similarity value represents, that two uses are more similar
parison of least mean square algorithm and fractional to each other. However, the main draw back of this equa-
least mean square has been proposed13. It is observed tion is that it does not consider the personal influence of
that fractional LMS has proved very well in case of deter- web services on similarity calculation. i.e., Co-invoked
ministic signal because of the higher rate of convergence web services are given equivalent weights in the computa-
and smaller amount of errors occur in it though LMS tion of similarity between two users. Therefore, a weighted
algorithm has better performance for random signals. A PCC has been developed which incorporates the personal
technique that does Karl’s Pearson Coefficient (KPC) has influences of web services into similarity computation
been proposed14. In order to make the system person- between two users. Weight of web service i based on QoS
alized, Felder’s learning styles catalogue is applied and deviation is calculated using the following steps.
to build many e-learning systems where the reliability of
their recommended learning styles are analyzed using • QoS Normalization: This step transforms each QoS
KPC. value of web service i, r (u, i), to a real number between
Vol 9 (29) | August 2016 | www.indjst.org Indian Journal of Science and Technology 3
Performance Comparison of Different Similarity Measures for Collaborative Filtering Technique
0 and 1. This could be done by comparing it with the Moreover, this approach provides low similarity value
maximum and minimum QoS values of web service regardless of the similar ratings made by two users on
i. Here two cases are to be considered. If the QoS cri- items and if the co-rated items present in the user-item
terion concerned is positive then r(u,i) is normalized rating matrix is very few, then it will not provide a reliable
using Equation (2);, if the QoS criterion is negative similarity value.
then r(u,i) is normalized using Equation (3). The working principle for Pearson Correlation
Coefficient is computed using the following formula
(2)
(8)
(3)
= Set of users invoked both web services i and j.
where set of QoS values of web service I is represented as r r(u,i) = web service i’s QoS value.
(i). n(u,i) is set to 1, in the case of . r(u,j) =web service j’s QoS value.
= web service i’s average QoS value.
• Computation of Standard Deviation using = web service j’s average QoS value.
Normalized QoS Values: This is computed using the weight associated with user u (Standard deviation
following formula of the normalized QoS values of web services invoked by
user u).
When the similarity measure is calculated, the follow-
(4) ing matrix is obtained from Table 1.
Table 2 denotes the final values when the similar-
where is the average QoS value of Web service i, is ity measure is calculated using Pearson Correlation
a threshold for the number of users that have invoked i, Coefficient (PCC)
i.e., . If is very small, the standard deviation is likely
2.1.2 Jaccard Uniform Operator Distance
to be overestimated by the original standard deviation
computation formula. The is used to address the above
(JacUOD)
issue. JacUOD16 approach investigates the characteristics of
similarity measurement for different multidimensional
• Weigh Generation: the weight of a Web Service i is
obtained using the following formula. Table 1. Example user-item rating matrix where the
value 99 corresponds to null (not rated)
(5)
Users\
Item1 Item2 Item3 Item4 Item5
The value of weight is always in the range (0, 1). Items
After weight generation, the similarity between user u User1 -7.82 8.79 -9.66 -8.16 -7.52
and v is computed using the following formula. User2 4.08 -0.29 6.36 4.37 -2.38
User3 99 99 99 99 9.03
(6) User4 99 8.35 99 99 1.8
User5 8.5 4.61 -4.17 -5.39 1.36
The above formula incorporates both the personal
influence of Web services and user rating value during Table 2. From the user similarity matrix in Table 1,
user similarity measurement. It implies that the weights applying PCC
of the web services with larger values will contribute more
Users\
during the similarity computation between two users. User 1 User 2 User 3 User 4 User 5
Users
In the next step, similarity between the web services I User1 1 -0.18665 0.154549 0.380217 0.298611
ad j is calculated using the standard PCC measure and the
User2 1 -0.21076 -0.05925 -0.14916
same is expressed in the following formula.
User3 1 0.306739 0.239582
(7) User4 1 0.360049
User5 1
4 Vol 9 (29) | August 2016 | www.indjst.org Indian Journal of Science and Technology
K. G. Saranya, G. Sudha Sadasivam and M. Chandralekha
vector spaces, and leads to better prediction accu- n = number of partitions.
racy. JacUOD approach can be adopted as a promising = ratio of the number of items with rating
approach for hybrid recommender systems to provide value i by user u to the total number of items rated by
more accurate similarity measurement. Data sparsity in user u.
user profile decreases the performance and quality of any = ratio of the number of items with rating
recommender systems. This similarity measurement only value i by user v to the total number of items rated by
focuses on rating-based collaborative filtering approaches. user v.
Moreover, JacUOD approach is ineffective for ranking- The working principle for Bhattacharyya coefficient17
based collaborative filtering approaches and also suffers is discussed below
from few or no overlapping items. Let (1,0,2,0,1,0,2,0) and (0,1,0,2,0,1,0,2)
The working principle for JacUOD is computed using be the rating vectors of user and .The ratings lie in
the following formula {1,2}, hence the number of partitions is 2. BC coefficient
between user and is computed the following for-
mula
(9)
(11)
= =1
It can be noted that there is no common rated items
between and . This measure could not calculate
user similarity in this situation. It can also be noted that
and both have preferences for giving low ratings. In
addition to that, two users have an identical rating distri-
bution which can be inferred that and are similar
in rating habits. Let = (0, 5, 0, 4, 0, 5, 0, 3) be the rating
When the similarity measure is calculated, the vector of user . A problem occurs where BC coefficient
following matrix is obtained from Table 1. between and equals 0 because there is no overlap
Table 3 denotes the final values when the similarity at all in every partition and still have certain similarities
measure is calculated using Jaccard Uniform Operator as their ratings are relatively centralized in distribution.
Distance (JacUOD) Moreover, this approach cannot be used to find a similar-
ity between pair of users if they rate on few or no similar
2.1.3 Bhattacharyya Coefficient items, also not scalable and computation is very complex.
The formula for Bhattacharyya coefficient is computed When the similarity measure is calculated, the following
using the following formula matrix is obtained from Table 1.
Table 4 denotes the final values when the similarity
(10) measure is calculated using Bhattacharyya Coefficient
(BC)
Table 3. From the user similarity matrix in Table 1,
applying JacUOD 2.2 Discussions on the New Similarity
Users\ Measure Model
User 1 User 2 User 3 User 4 User 5
Users The new similarity measure incorporates Pearson
User1 1 0.247754 0.230845 0.305642 0.307504 Correlation Coefficient (PCC) and Jaccard Coefficient.
User2 1 0.256345 0.279499 0.389933 The mathematical formalization of the proposed novel
User3 1 0.441945 0.353995 similarity measure can be calculated using the following
User4 1 0.398219 formula:
User5 1 (12)
Vol 9 (29) | August 2016 | www.indjst.org Indian Journal of Science and Technology 5
Performance Comparison of Different Similarity Measures for Collaborative Filtering Technique
Table 4. From the user similarity matrix in Table 1, Third, each user becomes comparable, that is each user
applying Bhattacharyya coefficient has different similarities. This can be seen in the above
Users\ matrix. Each pair has different similarities. However, this
User 1 User 2 User 3 User 4 User 5 is not the case in existing similarity measures. This also
Users
User1 1 0.86023206 0.7473063 0.73827124 0.889582 can be seen from the above matrix.
User2 1 0.6999996 0.67838615 0.9539392
User3 1 0.8247859 0.7038481
3. Results and Discussion
User4 1 0.7111426 3.1 Data set
User5 1
The data set of web service (www.wsdream.com) is used
in our experiments. This data set consists of 339 users
(13) and 5825 web services as user – item matrix. This matrix
also includes trough put and response time for each web
= set of users invoked both web service i and j. service. This Web service dataset is used for web ser-
r (u, i) = web service i’s QoS value. vice recommendation system. 80% of users are used for
r (u,j) = web service j’s QoS value. training while 20% is used for testing.
= web service i’s average QoS value.
= web service j’s average QoS value.
3.2 Evaluation Metrics
The performance of the new similarity measure is evaluated
(14) using two metrics called precision and recall. The main
When the similarity measure is calculated, the draw back of using these measures is that if number of
following matrix is obtained from Table 1. items increases in the top-N recommendation list then
Table 5 denotes the final values when the similarity recall increases while the precision decreases. Therefore,
measure is calculated by incorporating Pearson F-Measure which combines precision and recall is used
Correlation Coefficient (PCC) and Jaccard Coefficient. to measure the accuracy of predicting number of near-
First, from the above matrix we can see that the est neighbors and performance of the recommendation
similarity between User 1 and User 3 is higher when com- system.
pared to the similarity between User 1 and User 2. However,
this is not accurate in PCC, Jaccard and Bhattacharyya Experiments were conducted on web service data set
coefficient. This indicates that the new similarity measure and the proposed similarity measure is compared with
model is able to overcome the drawback of low similarity other traditional similarity measures. K-Neighbors and
regardless of similar ratings by two users. the number of recommendations are the two parameters
Second, the similarity between User 3 and User 5 is which can impact the performance of recommendation
also higher than the similarity between User 4 and User systems. The results are compared with different values of
5. However, the misleading still exists in PCC, Jaccard and these two parameters.
Bhattacharyya coefficient similarity. This demonstrates
that the new similarity measure can avoid the misleading. 3.2.1 Performance of different Similarity
Measures on Web Service Data Set
Table 5. From the user similarity matrix in Table 1,
applying combined similarity measure 3.2.1.1 K-Neighbors
Users\ When the k value increases, the precision value increases
User 1 User 2 User 3 User 4 User 5 while the recall values get increases as shown in the Figure
Users
User1 1 0.02089 0.05520 0.00475 0.02440 1, Figure 2, Figure 3.
User2 1 0.0183 0.00464 0.03561 3.2.1.2 Number of Recommendations
User3 1 0.00636 0.02500
When the k value increases, the precision value increases
User4 1 0.01531 while the recall values get increases as shown in the Figure
User5 1 4, Figure 5, Figure 6.
6 Vol 9 (29) | August 2016 | www.indjst.org Indian Journal of Science and Technology
K. G. Saranya, G. Sudha Sadasivam and M. Chandralekha
Figure 1. Comparison of precision against k-neighbors on Figure 4. Comparison of precision against number of
web service data set. recommendations on web service data set.
Figure 2. Comparison of recall against k-neighbors on Figure 5. Comparison of recall against number of
web service data set. recommendations on web service data set.
Figure 3. Comparison of F-measure against k-neighbors Figure 6. Comparison of F-measure against number of
on web service data set. recommendations on web service data set.
Vol 9 (29) | August 2016 | www.indjst.org Indian Journal of Science and Technology 7
Performance Comparison of Different Similarity Measures for Collaborative Filtering Technique
4. Conclusion 7. Chen M, Ma Y. A hybrid approach to web service recom-
mendation based on QoS-aware rating and ranking. arXiv
The paper first analyzes the disadvantages of the existing preprint arXiv:1501.04298; 2015.
similarity measures. In order to deal all these shortages, a 8. Lee G-Y, Tseng W-P. An enhanced memory-based
novel similarity measure approach which combines PCC collaborative filtering approach for context-aware rec-
and Jaccard is proposed. Experiments were conducted on ommendation. Proceedings of the World Congress on
web services data set to demonstrate the performance of Engineering; 2015.
9. Haipeng, et al. An improved collaborative filtering rec-
the new similarity measure. Experimental results shows
ommendation algorithm combining item clustering and
the effectiveness of the novel similarity measure and it
slope one scheme. Proceedings of the International Multi
can overcome the drawbacks of the traditional similarity Conference of Engineers and Computer Scientists; 2015.
measures. 10. Guo G, Zhang J, Yorke-Smith N. TrustSVD: Collaborative
filtering with both the explicit and implicit influence of user
5. References trust and of item ratings. Proceedings of the 29th AAAI
Conference on Artificial Intelligence (AAAI); 2015.
1. Luo B, et al. Recommendation scheme via improved itera- 11. Bar A, et al. Improving simple collaborative filtering mod-
tively collaborative filtering algorithm with neighborhood els using ensemble methods. Multiple Classifier Systems.
scale research. 2013 5th International Conference on IEEE Springer Berlin Heidelberg. 2013; 7872:1–12. DOI:
Computational and Information Sciences (ICCIS); Shiyang. 10.1007/978-3-642-38067-9_1.
2013 Jun 21-23. p. 609-12. DOI: 10.1109/ICCIS.2013.167. 12. Zhao Z, et al. Improving user topic interest profiles by behav-
2. Yu C, Huang L. Time-aware collaborative filtering for ior factorization. Proceedings of the 24th International
qos-based service recommendation. 2014 International Conference on World Wide Web, International World
Conference on IEEE Web Services (ICWS); Anchorage, AK. Wide Web Conferences Steering Committee; 2015. DOI:
2014 Jun 27-Jul 2. p. 265–72. DOI: 10.1109/ICWS.2014.47. 10.1145/2736277.2741656.
3. Park Y, et al. Fast collaborative filtering with a k-nearest 13. Imran R. Adaptive filtering algorithms for channel equaliza-
neighbor graph. 2014 International Conference on IEEE tion in wireless communication. Indian Journal of Science
Big Data and Smart Computing (BIGCOMP); 2014. DOI: and Technology. 2015 Aug; 8(17). DOI: 10.17485/ijst/2015/
10.1109/BIGCOMP.2014.6741414. v8i17/57805.
4. Mase H, Hayato O. A collaborative filtering incorporat- 14. Muruganandam S, Srinivasan N. Appraisal of felder - sil-
ing hybrid-clustering technology. 2012 International verman learning style model with discrete data sets. Indian
Conference on IEEE Systems and Informatics (ICSAI); Journal of Science and Technology. 2016 Mar; 9(10). DOI:
Yantai. 2012 May 19-20. p. 2342–6. DOI: 10.1109/ 10.17485/ijst/2016/v9i10/88992.
ICSAI.2012.6223524. 15. Liu J, Tang M. Zheng Z. Liu XF, Lyu S. Location-aware and
5. Xiong L, et al. A novel nearest neighborhood algorithm for personalized collaborative filtering for web service recom-
recommender systems. 2012 3rd Global Congress on IEEE mendation. IEEE Transactions on Services Computing.
Intelligent Systems (GCIS); Wuhan. 2012 Nov 6-8. p. 156–9. 2015. DOI: 10.1109/TSC.2015.2433251.
DOI: 10.1109/GCIS.2012.58. 16. Li M, Kai Z. A collaborative filtering algorithm combined
6. Wang J, De Vries AP, Reinders MJT. Unifying user-based with user habits for rating. International Conference on
and item-based collaborative filtering approaches by Logistics Engineering, Management and Computer Science;
similarity fusion. ACM Proceedings of the 29th Annual 2015.
International ACM SIGIR Conference on Research 17. Sun H-F, et al. JacUOD: A new similarity measurement for
and Development in Information Retrieval; 2006. DOI: collaborative filtering. Journal of Computer Science and
10.1145/1148170.1148257. Technology. 2012. DOI: 10.1016/j.ins.2007.07.024.
8 Vol 9 (29) | August 2016 | www.indjst.org Indian Journal of Science and Technology
