International Journal of Emerging Trends & Technology in Computer Science (IJETTCS)
Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com
Volume 2, Issue 2, March – April 2013 ISSN 2278-6856

Attribution Non-Commercial (BY-NC)

3 views

IJETTCS-2013-04-18-120

International Journal of Emerging Trends & Technology in Computer Science (IJETTCS)
Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com
Volume 2, Issue 2, March – April 2013 ISSN 2278-6856

Attribution Non-Commercial (BY-NC)

- DataMining and Emotional Intelli - ClusterAnalysis
- 06832486-data
- Wang_TrajStat
- Beyond Resilients Under Controllers and Over Controllers
- Features Mapping Based Human Gait Recognition
- Human Haptic
- Syllabus
- 2010_icpr
- Nguyen Armitage SurveysAndTutorials2008
- Ieee 2011 Java Projects Titles
- A Compression-Based Algorithm for Chinese Word Segmentation
- A Fuzzy Diagnostic Model and Its Application in Automotive
- RBF Classnote Mtech Spring2013
- MDS-UPDRS
- Separation of Music Signals by Harmonic Structure Modeling.pdf
- K-MEANS CLUSTERING FOR ANALYZING PRODUCTIVITY IN LIGHT OF R & D SPILLOVER
- 2014-ELL-LLR
- Report
- 10 - Cluster Analysis
- New Distance Measure of Single-Valued Neutrosophic Sets and Its Application

You are on page 1of 7

Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com Volume 2, Issue 2, March April 2013 ISSN 2278-6856

B R Prakash1 and Hanumanthappa. M 2 Research Scholar, 2 Professor, Department of Computer Science & Applications, Bangalore University, Bangalore. INDIA. clusters with intuitive descriptions with limited input. The inputs to SRC algorithms are snippets which are typically one or two sentences long. Common requirements for SRC methods can be defined as the following [1]: Algorithms should be fast, preferably linear-time and incremental. Algorithms should produce good results even with short text fragments. Algorithms should allow overlapping clusters. Search results often have multiple topics. Cluster labels should be intuitive. They should allow the users to understand the reason behind a clustering. Unlike other document clustering tasks, the use of short snippets as input is a pressing factor in SRC problem since the information gained from term frequencies tend to become noisy and meaningless for smaller sample size. In text retrieval the vector space model is the most commonly used. Documents in this model are represented as vectors of weighted terms called bag of words. For weighting, many approaches are used: binary schema (in which the weight of a term is 0 if the term does not appear in the document, 1 otherwise), the tf-idf scoring and so on. A set of objects O = {o1,o2,o3,o4,on} are represented with m-dimensional vectors which are stored in a matrix M of n rows and m columns, where n is the number of objects in the corpus while m is the number of features of the objects. These vector spaces endowed with a distance function define a metric space. The most widely used distance functions are[2]: Cosine similarity: it is defined as the cosine of the angle between Oa and Ob. More formally

1

similar objects. Each group, called cluster, consists of objects that are similar between themselves and dissimilar to objects of other groups. In other words, the goal of a good document clustering scheme is to minimize intra-cluster distances between documents, while maximizing inter-cluster distances further we improve the Furthest Point First algorithm from both the computational cost point of view and the output clustering quality. We call this algorithm M-FPF. Another crucial shortcoming of FPF is that it selects a set of centers not representative of the clusters. This phenomenon must be imputed to the fact that, when FPF creates a new center, it selects the furthest point from the previous selected centers and thus the new center can likely be close to a boundary of the subspace containing the data set. We modify M-FPF to use medoids instead of centers. Our results show improvement in the clustering quality. Keywords: Clustering, document clustering, Cluster Labeling, Information retrieval, FPF Algorithm.

1. INTRODUCTION

Data clustering is a statistical data analysis technique that is used to partition a data set into a set of meaningful groups according to given criteria. Different clustering algorithms have implicit weaknesses and strengths against different definitions of cluster meaningfulness. However, the main goal of clustering algorithms remains the same: sum of the distances between data points in the cluster (intra-cluster distance) should be minimal and sum of the distances between clusters (inter-cluster distance) should be maximal. Document clustering is often a harder task than clustering numerical data because one only has a numerical representation of textual data, defined mainly by word frequency. For human languages, most statistical correlations are noise according to context and this makes document clustering even harder. Over the years, mapping documents to vectors defined by term frequency weights to compute distances in Vector Space Model (VSM) remained as an effective strategy. With reliable frequency information and proper noise removal, good results are achieved with many clustering algorithms. Search Result Clustering (SRC) is sub-topic of document clustering in which the clustering methods are expected to produce Volume 2, Issue 2 March April 2013

A distance can be easily derived from cosine similarity by: The most important property of cosine similarity is that it does not depend on the length of the vectors: = Page 299

Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com Volume 2, Issue 2, March April 2013 ISSN 2278-6856

for > 0. This property makes the cosine similarity widely used in text information retrieval. Jaccard coefficient: in its basic form it is defined as: when a data object belongs to each cluster with a degree of membership. 2.1 Partitional clustering When the data representation and the distance function d have been chosen, partitional clustering reduces to a problem of minimization of a given target function. The most widely used are: K-center minimizes the maximum cluster radius

where oa ob is the set of features in common between oa and ob and oa ob is the total set of features Many variants of the Jaccard coefficient were proposed in the literature. The most interesting is the Generalized Jaccard Coefficient (GJC) that takes into account also the weight of each term. It is defined as

GJC is proven to be a metric [Charikar, 2002]. The Generalized Jaccard Coefficient defines a very flexible distance that works well with both text and video data. Minkowski distance: it is defined as:

It is the standard family of distances for geometrical problems. Varying the value of the parameter p, we obtain different well known distance functions. When p = 1 the Minkowski distance reduces to the Manhattan distance. For p = 2 we have the well known Euclidean distance

where C = {c1, . . . , ck} are k clusters such that Cj is the center of the j-th cluster and j is its centroid. For all these functions it is known that finding the global minimum is NP-hard. Thus, heuristics are always employed to find a local minimum. 2.2 FPF algorithm for the k-center problem We first describe the original Furthest Point First (FPF) algorithm proposed by Gonzalez [5] that represents the basic algorithm we adopted in this paper.

2. CLUSTERING

Clustering algorithms can be classified according with many different characteristics. One of the most important is the strategy used by the algorithm to partition the space[3,4]: Partitional clustering: given a set O = {o1, ,on} of n data objects, the goal is to create a partition C = {c1, . . . , ck} such that:

Basic algorithm Given a set O of n points, FPF increasingly computes the set of centers c1 . . . ck O, where Ck is the solution to the k-center problem and C1 = {c1} is the starting set, built by randomly choosing c1 in O. At a generic iteration 1 < i k, the algorithm knows the set of centers Ci 1 (computed at the previous iteration) and a mapping that associates, to each point p O, its closest center (p) Ci 1. Iteration i consists of the following two steps: 1. Find the point p O for which the distance to its closest center, d(p, (p)), is maximum; make p a new center ci and let Ci = Ci 1 {ci }. 2. Compute the distance of ci to all points in O \ Ci 1 to update the mapping of points to their closest center. After k iterations, the set of centers ck = {c1, . . . , ck} and the mapping define the clustering. Cluster Ci is the set of points {p O \ Ck such that (p) =ci}, for i [1, k]. Each iteration can be done in time O(n), hence the overall cost of the algorithm is O(kn). Experiments have shown Page 300

Hierarchical clustering: given a set O = {o1,o2,o3,o4,on} of n data objects, the goal is to build a tree-like structure (called dendrogram) H = {h1, . . . , hq} with q n, such that: given two clusters ci hm and cj hl with hl ancestor of hm, one of the following conditions hold: ci cj or ci cj = for all i, j i,m, l [1, q]. Partitional clustering is said hard if a data object is assigned uniquely to one cluster, soft or fuzzy Volume 2, Issue 2 March April 2013

Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com Volume 2, Issue 2, March April 2013 ISSN 2278-6856

that the random choice of c1 to initialize C1 does not affect neither the effectiveness nor the efficiency of the algorithm. FPF: Data: Let O be the input set, k the number of clusters Result: C, k-partition of O C = x such that x is an arbitrary element of O; for i = 0; i < k; i + + do Pick the element x of O \ C furthest from the closest element in C; Ci = Ci = x; end forall x O \ C do Let i such that d(ci, x) < d(cj , x), j i Ci. append (x); end Algorithm 1: The furthest point first algorithm for the kcenter problem. Wherever Times is specified, Times Roman or Times New Roman may be used. If neither is available on your word processor, please use the font closest in appearance to Times. Avoid using bit-mapped fonts if possible. TrueType 1 or Open Type fonts are preferred. Please embed symbol fonts, as well, for math, etc. {o1, . . . , on}, we denote with S(oi, oj) the similarity of the objects oi and oj according to a given similarity function. We say that oi and oj are mates if they belong to the same cluster. We define: Homogeneity of a clustering: the average similarity between mates. Let M be the number of mate pairs:

Separation of a clustering: the average similarity between non-mates. As M is the number of mate pairs, the number of non-mates pairs is given by .

3. CLUSTERING VALIDATION

Since the clustering task has an ambiguous definition, the assessment of the quality of results is also not well defined. There are two main philosophies for evaluating the clustering quality: Internal criterion: is based on the evaluation of how the output clustering approximates a certain objective function, External criterion: is based on the comparison between the output clustering and a predefined handmade classification of the data called ground truth. When a ground truth is available, it is usually preferable to use an external criterion to assess the clustering effectiveness, because it deals with real data while an internal criterion measures how well founded the clustering is according with such mathematical definition. 3.1 Internal measures There is a wide number of indexes used to measure the overall quality of a clustering. Some of them (i.e. the mean squared error) are also used as goal functions for the clustering algorithms. Homogeneity and separation According with the intuition, the more a cluster contains homogeneous objects the more it is a good cluster. Nevertheless the more two clusters are well separated the more they are considered good clusters. Following the intuition, homogeneity and separation [6] attempt to measure how compact and well distanciated clusters are among them. More formally given a set of objects O = Volume 2, Issue 2 March April 2013

3.2 External measures In the following, we denote with GT(S) = {GT1, . . . ,GTk} the ground truth partition formed by a collection of classes; and with C = {c1, . . . , ck} the outcome of the clustering algorithm that is a collection of clusters. F-measure: The F-measure was introduced in [Larsen and Aone, 1999] and is based on the precision and recall that are concepts well known in the information retrieval literature [Kowalski, 1997], [Van Rijsbergen, 1979]. Given a cluster cj and a class GTi we have:

Note that precision and recall are real numbers in the range [0, 1]. Intuitively precision measures the probability that an element of the class GTi falls in the cluster cj while recall is the probability that an element of the cluster cj is also an element of the class GTi. The Fmeasure F(GTi, cj) of a cluster cj and a class GTi is the harmonic mean of precision and recall:

where n is the sum of the cardinality of all the classes. The value of F is in the range [0, 1] and a higher value indicates better quality. Entropy: Entropy is a widely used measure in information theory. In a nutshell we can use the relative entropy to measure the amount of uncertainty that we have about the ground Page 301

Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com Volume 2, Issue 2, March April 2013 ISSN 2278-6856

truth provided the available information is the computed clustering. Given a cluster cj and a class GTi, we can define NCE ranges in the interval [0, 1], and a greater value implies better quality of cj . The complementary normalized entropy of C is the weighted average of the contributions of the single clusters in C. Let n = P|C| j1 |cj | be the sum of the cardinalities of the clusters of C. Note that when clusters may overlap it holds thatn n. Thus

where n is the number of elements of the whole clustering. The value of E is in the range [0, log n] and a lower value indicates better quality. Accuracy: While the entropy of a clustering is an average of the entropy of single clusters, a notion of accuracy is obtained using simply the maximum operator:

In this we improve the Furthest Point First algorithm from both the computational cost point of view and the output clustering quality. Since theoretically the FPF algorithm as proposed by Gonzalez [9] is optimal (unless P = NP), only heuristics can be used to obtain better results and, in the worst case, it is not possible to go behind the theoretical bounds. We profiled FPF and analyzed the most computational expensive parts of the algorithm. We found that most of the distance computation are devoted to find the next furthest point. We observed that there are cases such that some distance computations can be avoided without changing the final clustering algorithm. FPF clustering quality can be improved modifying part of the clustering schema. We describe an approach that use the random sampling technique to improve clustering output quality, we call this algorithm M-FPF. Another crucial shortcomings of FPF is that it selects a set of centers not representative of the clusters. This phenomenon must be imputed to the fact that, when FPF creates a new center, it selects the furthest point from the previous selected centers and thus the new center can likely be close to a boundary of the subspace containing the data set. We modify M-FPF to use medoids instead of centers. The triangular inequality to improve the FPF speed We observed that most of the running time of the FPF algorithm is devoted to compute distances for finding the closest center to each point. More precisely at a generic iteration 1 < i k, after finding the center k, n k distances must be computed to decide whether or not to assign a point to the new center. If this is done in a straightforward manner it takes O(n) time per iteration, thus the total computational cost of the algorithm is O(nk). Exploiting the triangular inequality, in certain conditions we can avoid to compute the distances among all the points in a cluster and the new center being sure that they are closer to their center. Unfortunately the worst case time complexity still remain O(nk) because the number of saved distance computations depends on data distribution and thus, it can not be predicted in advance. The modifications discussed here do not change the FPF output, they only speed up the algorithm. In the FPF algorithm, any center ci and its associated set of closest points Ci. Store Ci as a ranked list, in order of

The accuracy A is in the range [0, 1] and a higher value indicates better quality. Normalized mutual information: The normalized mutual information comes information theory and is defined as follows:

from

where P(c) represents the probability that a randomly selected object oj belongs to c, and P(c, c) represents the probability that a randomly selected object oj belongs to both c and c. The normalization, achieved by the 2 log |C||GT| factor, is necessary in order to account for the fact that the cardinalities of C and GT are in general different [7]. Higher values of NMI mean better clustering quality. NMI is designed for hard clustering. Normalized complementary entropy: In order to evaluate soft clustering, the normalized complementary entropy [8] is often used. Here we describe a version of normalized complementary entropy in which we have changed the normalization factor so as to take overlapping clusters into account. The entropy of a cluster cj C is

Page 302

Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com Volume 2, Issue 2, March April 2013 ISSN 2278-6856

decreasing distance to ci. When a new center cj is selected, scan Ci in decreasing order of distance, and stop scanning when, for a point p Ci, it is the case that d(p, ci) 1 2d(cj , ci). By the triangular inequality, any point p that satisfies this condition cannot be closer to cj than to ci. This rule filters out from the scan points whose neighbor cannot possibly be cj , thus significantly speeding up the identification of neighbors. Note that all distances between pairs of centers must be available; this implies an added O(k2) cost for computing and maintaining these distances. Note that this modified algorithm works in any metric space, hence in any vector space. Using a random sample The efficiency of the algorithm is further improved by applying FPF algorithm not to the whole data set but only to a random sample of size of the input points (sample size suggested in [10]). Note that given that k n, it is always true that n n. Then we add the remaining (n n) points to the cluster of their closest center, one by one.

i = getCenter (Ci ); end forall p in O \ R do assign p to cluster Ci such that d(p, i ) < d(p, j), j i; end Algorithm 3: M-FPF.

Using medoids as centers The concept of medoid was introduced by Kaufman and Rousseeuw in [11]. Medoids have two main advantages with respect to centroids: first of all, they are elements of the input and not artificial objects. This make medoids available also in those environments in which the concept of centroid is not well defined or results artificious. Nevertheless, in many environments (i.e texts) centroids tends to become dense objects with a high number of features more of which of poor meaning. This makes centroids to lose representativeness and compute distances with them becomes more expensive with respect to distances between normal objects. The main drawback of the original definition is that the clustering algorithm (Partition Around Medoids) and the computation of medoids is expensive, to overcome this disadvantage many different re-definitions of medoids were introduced in the literature. In the context of the Furthest Point First heuristic where some input points are elected as cluster centers and are used to determinate which input points belong to the cluster, the restrictions of the use of centroids are not present. However, we observed that, although the objects selected from FPF as centers determine the points belonging to the cluster, they are not centers in the sense suggested by the human intuition. M-FPF-MD: Data: Let O be the input set, k the number of desired clusters Result: C: a k-partition of O Initialize R with a random sample of size |O|k elements of O; C = FPF(R, k); forall Ci C do ti = getRandomPoint (Ci ); ai = ci such that max d(ci , ti ) for each ci Ci ; bi = ci such that max d(ci , ai ) for each ci Ci ; mi = ci such that min |d(ci , ai ) d(ci , bi )| + |d(ci , ai ) + d(ci , bi ) d(ai , bi )|; end forall p in O \ R do assign p to cluster Ci such that d(p,mi ) < d(p,mj ), j i; if d(p, bi ) > d(ai , bi ) then ai = p ; if d(p, ai ) > d(ai , bi ) then bi = p ; Page 303

Figure 2.5. Exploiting the triangular inequality. Also in the operation of insertion of the (n n) remaining points, the bottleneck is the time spent computing distances to the point to the closest center. According with [Phillips, 2002] this operation can be made more efficiently exploiting the triangular inequality, even if the worst case running time does not change. Consider to have available the distances between all the pairs of centers of the clustering. Let p be the new point to be inserted in the clustering, by the triangular inequality if 1 2d(ci, cj) > d(ci, p) then d(ci, p) < d(cj , p). It means that the computation of the distance d(cj , p) can be safely avoided. We will refer to this algorithm as M-FPF. M-FPF: Data: Let O be the input set, k the number of desired clusters Result: C: a k-partition of O Initialize R with a random sample of size elements of O; C = FPF(R, k); forall Ci C do Volume 2, Issue 2 March April 2013

Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com Volume 2, Issue 2, March April 2013 ISSN 2278-6856

if d(p, bi ) > d(ai , bi ) or d(p, ai ) > d(ai , bi ) then mi = ci such that min |d(ci , ai ) d(ci , bi )| + |d(ci , ai ) + d(ci , bi ) d(ai , bi )|; end end Algorithm 4: M-FPF-MD. Snippets, the comparable. final clustering quality is always

5. TEXT CLUSTERING

We described clustering algorithms based on the Furthest Point First heuristic by Gonzalez [5, 12]. In this section we compare all these algorithms (FPF, M-FPF and MFPF-MD) in the setting of text clustering. We compared them using two different metric spaces: the well known Cosine Similarity (CS) and the Generalized Jaccard Coefficient (GJC). Moreover we are interested also in comparing our algorithms against k-means which is the most used algorithm for text clustering. In order to compare clustering algorithms we used two different data sets: Reuters: is one of the most used data sets in information retrieval. After an initial cleaning to remove multilabelled documents, we obtain a corpus of 8538 documents of various lengths organized in 65 mutually exclusive classes. We further remove 10 classes each containing only a single document. Snippets: is a collection of 30 small corpora of 200 web snippets returned by the Open Directory Project (ODP) search engine in correspondence of the 30 most popular queries according with Google statistics. The two considered datasets present many differences: Reuters contains much more documents that are much longer than those in Snippets. Documents in Reuters can have a large size variability and are completely flat. Instead, Snippets contains semi-structured documents (a field with the snippet title and one with the snippet body). Since both the datasets we used are endowed with a manual classification, it was possible to establish a ground truth. We used four measures to validate the clustering correspondence with the ground truth: the Normalized Mutual Information (NMI), the Normalized Complementary Entropy (NCE), the F-measure and the Accuracy. All these measures are in the range [0, 1] where higher values mean better quality FPF evaluation In table 3.1 we show a comparison of the three clustering algorithms for the k center problem. Each clustering was run using two different distance functions with the goal of comparing also the metrics for text clustering. As shown in table 3.1, when clustering Reuters data, all the algorithms have a very different behavior depending on the used distance. In all the cases cosine similarity performs better than GJC. Instead, when clustering Volume 2, Issue 2 March April 2013

Reuters data put in evidence also some differences among the clustering algorithms. In fact FPF seems to achieve a worse quality with respect to the others. When clustering Snippets, these differences become much less evident. In fact, the F-measure and accuracy are always comparable; NMI, in the case FPF, is quite higher than that of M-FPF and M-FPF-MD. In contrast NCE of FPF is lower than that of the other algorithms. Table 1 Comparison of FPF, M-FPF and M-FPF-MD on the Reuters and Snippets datasets with Generalized Jaccard Coefficient and Cosine Similarity

FPF CS

0.29 0.76 0.45 0.704 0.807 0.538 0.47 0.678

Dat aset

Measu re

M-FPF GJC

0.080 2 0.633 0.377 0.579 0.774 0.498 0.486 0.656

M-FPF-MD GJC

0.08 8 0.63 7 0.33 0.57 7 0.5 0.77 6 0.43 9 0.64 9

CS

0.339 0.8 0.561 0.76 0.524 0.798 0.448 0.662

CS

0.281 0.751 0.439 0.696 0.472 0.738 0.386 0.602

GJC

0.08 9 0.59 6 0.21 7 0.51 4 0.45 6 0.73 3 0.40 4 0.61 2

Snippets

Reuters

F-mea Acc.

6. CONCLUSIONS

To give a global overview we report all performances and running time results for respectively the Reuters and Snippets datasets. After analyzing the data, it is clear that the perfect clustering algorithm still does not exist and this motivates the effort of researchers. Hence some observations should be done. First of all, the on Reuters data, k-means outperforms algorithms for kcenter by a factor of roughly 10% but requires at least two orders of magnitude more time. Using GJC, time and performance differences become more evident. Clearly the best clustering/metric scheme is strictly dependent on the problem requirements and data size. In off-line applications probably computational efficiency should be sacrificed in favor of quality, but this is not always true. REFERENCES [1.] Geraci, F., Pellegrini, M., Sebastiani, F., Maggini, M.: Cluster generation and cluster labeling for web snippets: A ast and accurate hierarchical solution.

Page 304

Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com Volume 2, Issue 2, March April 2013 ISSN 2278-6856

Technical Report IIT TR-1/2006, Institute for Informatics and Telematics of CNR (2006) [2.] Nearest-neighbor searching and metric space dimensions. In Gregory Shakhnarovich, Trevor Darrell, and Piotr Indyk, editors, Nearest-Neighbor Methods for Learning and Vision: Theory and Practice, pages 1559. MIT Press, 2006. [3.] Flavio Chierichetti, Alessandro Panconesi, Prabhakar Raghavan, Mauro Sozio, Alessandro Tiberi, and Eli Upfal. Finding near neighbors through cluster pruning. In Proceedings of ACM PODS, 2008. [4.] Paolo Ferragina and Antonio Gulli. A personalized search engine based on Web-snippet hierarchical clustering. In Special Interest Tracks and Poster Proceedings of WWW-05, 14th International Conference on the World Wide Web, pages 801810, Chiba, JP, 2006. [5.] Karina Figueroa, Edgar Chavez, Gonzalo Navarro, and Rodrigo Paredes. On the least cost for proximity searching in metric spaces. In 5th International Workshop on Experimental Algorithms (WEA), volume 4007 of Lecture Notes in Computer Science, pages 279290. Springer, 2006. [6.] Filippo Geraci, Marco Pellegrini, Paolo Pisati, and Fabrizio Sebastiani. A scalable algorithm for highquality clustering of Web snippets. In Proceedings of SAC-06, 21st ACM Symposium on Applied Computing, pages 10581062, Dijon, FR, 2007. [7.] Filippo Geraci, Mauro Leoncini, Manuela Montangero, Marco Pellegrini, and M. Elena Renda. Fpf-sb: a scalable algorithm for microarray gene expression data clustering. In Proceedings of 1st International Conference on Digital Human Modeling, 2008. [8.] Stanislaw Osinski and Dawid Weiss. Conceptual clustering using Lingo algorithm: Evaluation on Open Directory Project data. In Proceedings of IIPWM-04, 5th Conference on Intelligent Information Processing and Web Mining, pages 369 377, Zakopane, PL, 2004. [9.] M. Hanumanthappa, B R Prakash ,Mamatha M Improving the efficiency of document clustering and labeling using Modified FPF algorithm Proceeding of international conference on problem solving and soft computing, 2011 [10.] Carrot2: Design of a flexible and efficient web information retrieval framework, in Lecture Notes in Computer Science 3528.Springer, 2005, pp. 439 444. [11.] Leonard Kaufman and Peter J. Rousseeuw. Finding groups in data: an introduction to cluster analysis. Wiley, 1990. A Wiley-Interscience publication. [12.] Takazumi Matsumoto and Edward Hung Fuzzy Clustering and Relevance Ranking of Web Search Results with Differentiating Cluster Label Generation. [13.] Chavez and G. Navarro. An effective clustering algorithm to index high dimensional metric spaces. Volume 2, Issue 2 March April 2013 In Proceedings of the6th International Symposium on String Processing and Information Retrieval (SPIRE2000), pages 7586. IEEE CS Press, 2000

Page 305

- DataMining and Emotional Intelli - ClusterAnalysisUploaded byjeffconnors
- 06832486-dataUploaded byahmadlukyr
- Wang_TrajStatUploaded byapi-26352014
- Beyond Resilients Under Controllers and Over ControllersUploaded byEniko
- Features Mapping Based Human Gait RecognitionUploaded byEditor IJRITCC
- Human HapticUploaded byRajasekar Panneerselvam
- SyllabusUploaded byNikhil Mawandia
- 2010_icprUploaded bySameeran Amar Nath
- Nguyen Armitage SurveysAndTutorials2008Uploaded byMahadev
- Ieee 2011 Java Projects TitlesUploaded byieeexploreprojects
- A Compression-Based Algorithm for Chinese Word SegmentationUploaded bytonbear
- A Fuzzy Diagnostic Model and Its Application in AutomotiveUploaded byCABean1
- RBF Classnote Mtech Spring2013Uploaded byNishantKumar
- MDS-UPDRSUploaded bydquebradas
- Separation of Music Signals by Harmonic Structure Modeling.pdfUploaded byRick Rite
- K-MEANS CLUSTERING FOR ANALYZING PRODUCTIVITY IN LIGHT OF R & D SPILLOVERUploaded byAlejandro Carver
- 2014-ELL-LLRUploaded byShahid KI
- ReportUploaded byLuis Franço
- 10 - Cluster AnalysisUploaded bySusi Lestari
- New Distance Measure of Single-Valued Neutrosophic Sets and Its ApplicationUploaded byMia Amalia
- A Good All-Around Semi-supervised LearningUploaded bytsabitafawaati
- tun analytics submission lucUploaded byapi-438499109
- Face Recognition using Content Based Image Retrieval for Intelligent SecurityUploaded byIJAERS JOURNAL
- Top 50 R Interview Questions You Must Prepare for 2018 _ EdurekaUploaded byAbhinav Soni
- Contoh ProposalUploaded bysasmito skom
- Kdd1999 CactusUploaded byKaunteya Shaw
- 04 LEC Data Science KmeansUploaded byViram Shah
- Cluster AnalysisUploaded byshivendra kumar pandit
- Sentiment Ci 2006Uploaded byMansoor Ahmed
- clusteranalysisanddatamining.pdfUploaded byYianni

- A Review on Artificial Intelligence in Education for Learning Disabled ChildrenUploaded byAnonymous vQrJlEN
- PERFORMANCE EVALUATION OF SHUNT WOUND DC MOTOR IN PROCESS CONTROLUploaded byAnonymous vQrJlEN
- C.Ri.S.P. - Cyber Risk Analysis in Industrial Process System EnvironmentUploaded byAnonymous vQrJlEN
- Heart Disease Prediction Using Hybrid ClassifierUploaded byAnonymous vQrJlEN
- IRIS SCAN BASED EMERGENCY SYSTEMUploaded byAnonymous vQrJlEN
- IJAIEM-2019-06-30-13.pdfUploaded byAnonymous vQrJlEN
- Prediction of Cyberbulling Using Random Forest ClassifierUploaded byAnonymous vQrJlEN
- Proficient Technique for Classification of ECG SignalUploaded byAnonymous vQrJlEN
- Performance Evaluation of No-frill AirlinesUploaded byAnonymous vQrJlEN
- An empirical study of Factors contributing to the adoption of Free Desktop Open Source Software in Africa: A case of Kenya university studentsUploaded byAnonymous vQrJlEN
- Effects Of Cybercrime In IndiaUploaded byAnonymous vQrJlEN
- Impact of Slurry Temperature on Biogas Production from Chicken Waste and Food Wastes by Anaerobic Digestion MethodUploaded byAnonymous vQrJlEN
- CONSUMPTION OF SANITARY NAPKINS&EFFECT ON ENVIRONMENTALS USTAINABILITYUploaded byAnonymous vQrJlEN
- Classification of Power Quality Disturbances Using Wavelet Transform and Halfing AlgorithmUploaded byAnonymous vQrJlEN
- FACTORS INFLUENCING ONLINE SHOPPING –AN EXPLORATORY REVIEW IN THE INDIAN PERSPECTIVEUploaded byAnonymous vQrJlEN
- Socio-economic Conditions of the Agricultural Labourers: An AnalysisUploaded byAnonymous vQrJlEN
- Analysis two level of real time load flow 132/33/11kv sub-station using ETAP softwareUploaded byAnonymous vQrJlEN
- Music as Persuasive Communication StrategyinAdvertising and BrandingUploaded byAnonymous vQrJlEN
- IJAIEM-2019-06-25-9.pdfUploaded byAnonymous vQrJlEN
- Study of Smoke Levels Using crude Jatropha Oil with Preheated Condition and at Normal Condition Injection in High Grade Low Heat Rejection Combustion Chamber of Diesel Engine at Different Operating PressuresUploaded byAnonymous vQrJlEN
- Digital Literacy and Academic Performance at Higher Education: Effects and Challenges Authors: Dr. Jyotsana ThakurUploaded byAnonymous vQrJlEN
- An Approach to Study Vibration Phenomena Using Field Data Based Mathematical Modeling in Spinning MachineUploaded byAnonymous vQrJlEN
- Prediction of Personality through Hand Geometry and Big Five Factor ModelUploaded byAnonymous vQrJlEN
- Development of Naval Alloy by Casting ProcessUploaded byAnonymous vQrJlEN
- AUTOMATIC WATER FLOW REGULATORUploaded byAnonymous vQrJlEN
- 3D PRINTED INTEGRATED FULLFUNCTION PROSTHETIC HAND-ARM SYSTEMUploaded byAnonymous vQrJlEN
- Measuring effectiveness of TBRP method for different capacity Data centersUploaded byAnonymous vQrJlEN
- Offline Signature Verification using LBP TechniqueUploaded byAnonymous vQrJlEN

- HW3Uploaded byArpit Gulati
- ICCC14 Metric LearningUploaded byLanna Larissa
- Docs Slides Lecture11Uploaded byIrham Dodot
- Using Domain Knowledge for Ontology-Guided Entity Extraction from Noisy, Unstructured Text DataUploaded byvvsailesh
- sm_otsmUploaded byDeepak Kumar
- Studio 3 Precision SettingsUploaded byOncemore Suwira
- Hierarchical Structure Generation for Malayalam using Hybrid ApproachUploaded byAnonymous vQrJlEN
- Machine Learning Yearning V0.5 01Uploaded bygburtini
- DOCUMENTUploaded byPINACLE
- WM2Uploaded bysivagami
- Undergraduate ThesisUploaded byAmiya Patanaik
- We Ka TutorialUploaded byFasih Sendu
- Identification and Classification of Leaf Diseases in Turmeric PlantsUploaded byAnonymous 7VPPkWS8O
- Histogram-Based Color Image RetrievalUploaded byNadisha Fathima
- Brill's Rule-based Part of Speech Tagger for KadazanUploaded byidescitation
- 2_live-2934-4846-jair.pdfUploaded byCristinush15
- 10.1.1.155Uploaded byFasy Awan
- White paperUploaded byKannadas Puthuval
- Analyzing Android Encrypted NetworkUploaded byChanchal Kumar
- A hybrid approach to fuzzy name search incorporating language-based and textbased principlesUploaded byHorng J P Wu
- An Immune Inspired Multilayer IDSUploaded byijcsis
- heart diseaseUploaded byapi-352507964
- draisbach_icde12Uploaded byRizki Okta
- Introduction IRUploaded byAdemir
- 126943521-HOTS-Questions-for-Math.pdfUploaded byElain yuuuo
- Novelty Detection via Topic Modeling in Research ArticlesUploaded byCS & IT
- Intech-semantic Matchmaking AlgorithmsUploaded by9835415933
- Jirkm2-5-Experiment on Cross-lingual Information Retrieval Using Malay-English Bilingual Dictionary Nurul AmelinaUploaded byJournal of Information Retrieval and Knowledge Management
- Search Engines Information Retrieval in Practice.pdfUploaded bypraveenaprabhu
- Natural Language Support for Competency Evaluation of Web-OntologiesUploaded bymkppan

## Much more than documents.

Discover everything Scribd has to offer, including books and audiobooks from major publishers.

Cancel anytime.