Collaborative Clustering with Heterogeneous Algorithms
|
|
- Griselda Belinda Murphy
- 5 years ago
- Views:
Transcription
1 Collaborative Clustering with Heterogeneous Algorithms Jérémie Sublime, Nistor Grozavu, Younès Bennani and Antoine Cornuéjols AgroParisTech, UMR INRA MIA rue Claude Bernard, Paris Cedex 5, France Université Paris 13, Sorbonne Paris Cité, LIPN UMR CNRS av. J-B Clément, Villetaneuse, France Abstract The aim of collaborative clustering is to reveal the common underlying structures found by different algorithms while analyzing data. The fundamental concept of collaboration is that the clustering algorithms operate locally but collaborate by exchanging information about the local structures found by each algorithm. In this framework, the one purpose of this article is to introduce a new method which allows to reinforce the clustering process by exchanging information between several results acquired from different clustering algorithms. The originality of our proposed approach is that the collaboration step can use clustering results obtained from any type of algorithm during the local phase. This article gives the theoretical foundations of our approach as well as some experimental results. The proposed approach has been validated on several data sets and the results have shown to be very competitive. I. INTRODUCTION Data Clustering is an important task in the process of knowledge extraction from databases. It aims to discover the intrinsic structure of a set of objects by forming clusters that share similar characteristics. The difficulty of this task has significantly increased over the past two decades when the available data sets saw their volume explode. This increasing complexity as well as new types of data sets distributed over several sites has make it difficult for a single clustering algorithm to give competitive results. This problem can however be tackled more easily by having several clustering algorithms working together. Most of the existing approaches to this problem rely on consensus-based methods [1], [2], [3], [4], [5] which aggregate (or fuse) the clustering results of all algorithms. Then a voting technique or a clustering technique is applied in order to get a unique final clustering result. Other methods to solve this kind of problem have recently appeared in the form of collaborative clustering techniques. Collaborative Clustering is an emerging problem in data mining and only few works on this subject can be found in the literature, (e.g. [6], [7], [8], [9]). The fundamental concept of collaboration is that the clustering algorithms operate locally (on individual subsets) and then collaborate by exchanging information about their structures and results in order to improve their models [6], [9], [10]. Collaborative techniques therefore differ from consensus-based methods in the sense that they aim at a mutual improvement of all participating algorithms whereas the consensus methods aims at providing a single result for all the algorithms. While the idea of collaboration is clearly more interesting, unlike consensus-based methods, so far all proposed collaborative techniques imply that the collaborating algorithms must be from the same algorithm family. In this article we propose a solution to previous limits in the form of a collaborative framework that works with most clustering algorithms regardless of their algorithm family. The exact context of our framework is known as horizontal collaboration [9]: All algorithms are working either on subsets representing the same data in different feature spaces, or on the exact same data searching for a different number of clusters, or a mix of both. The rest of the article is organized as follows: In Section 2, we present our proposed collaborative approach. Section 3 shows the adaptation of our collaborative framework to several clustering algorithms. In Section 4, we present some test of the proposed approach on different data sets. Finally the paper s with a conclusion and future works to ext the range of the proposed method. II. A. Principles COLLABORATIVE CLUSTERING WITH HETEROGENEOUS ALGORITHMS In this section we consider a group of clustering algorithms C = {c 1,..., c j } and a data set X = {x 1,..., x N }, x i R d. We suppose that our clustering algorithms are working on the same data, but may or may not have access to all the columns attributes. We note S = {s 1,..., s N }, s i 1..K, the solution vector that each local algorithm will provide. S = arg max(p (S X, Θ)) = arg max (P (X S, Θ)P (S)) S S (1) We make the assumption that all our clustering algorithms uses clusters that can be modeled by a law of parameters Θ.
2 Each algorithm may be based on a different statistical models (gaussian, multinomial, poisson, etc.) and may search for a different number of clusters. We make the hypothesis that our collaborative method can only be applied to algorithms that are trying to optimize an equation similar to equation (1). Such algorithms include: The K-Means algorithm and its variants, all Expectation Minimization based algorithms [11], several unsupervised neural network algorithms such as the SOM algorithm and the GTM algorithm, and a few algorithms from image segmentation such as the Iterated Conditional Modes [12]. Equation (1) can be solved using a local minimization process as shown below: P (X S, Θ)P (S) = t P (x t s t, θ st )P (s t ) (2) In equation (2), the main issue comes from P (s), the probability of occurrence of each cluster which is a priori unknown. Each specific algorithms implies its own hypothesis about P (s) : Equiprobability of all clusters. The hypothesis assume that all clusters have the same probability of occurrence. This is the hypothesis made in the K-Means algorithm and some version of the EM algorithm. A posteriori measured probability. This method is followed by most probabilistic unsupervised classifiers. The main idea is to measure the occurrence of all clusters after each iteration. The local probability hypothesis. This is mostly used in computer vision and works as follows : Instead of measuring a global P (s) for the whole data set, P (s) will dep on the neighborhood configuration of the observed data, and thus its value will be decided deping on the clusters assigned to the neighbor data (think about pixel depencies). For our collaborative process, we decided to follow a hypothesis similar to the one made in computer vision: during the collaborative step, we will consider that each cluster occurrence probability P (s) will not be a global probability. Instead, P (s) will be bound to the clustering choices made by the other algorithms for the same data point. First, in the local step each clustering algorithm will process the data it has access to and produce a clustering result in the form of a solution vector. The result of this local step will be a matrix similar to Table I where each column is a solution vector proposed by an algorithm. Then, from these solutions vectors, we will compute probabilistic confusion matrices that will be used in the collaborative step to evaluate P (s). This idea is inspired from another work [13], where the same confusion matrices are computed for a consensus-based method. The first step of our collaboration process is to map the clusters of the different algorithms. There is no reason or obligation for two given clustering algorithms c i and c j to be looking for the same number of clusters. TABLE I. EXAMPLE OF RESULTS AFTER THE LOCAL STEP X\C c 1 c 2... c j x x x x N We note Ψ ci cj the probabilistic confusion matrix mapping the clusters from algorithm c i to the clusters of algorithm c j. Ψ ci cj s a,s b the probability of having a data being put in the cluster s b of the clustering algorithm c j if it is in the cluster s a of algorithm c i. Once all the matrices Ψ ci cj have been computed by browsing all the results vectors, it is possible to start the collaborative step. For a given clustering algorithm c i, given the matrices Ψ and the results of the others algorithms from the local step, the local equation to be maximized regarding to s during the collaborative step becomes as shown in Equations (3) and (4). P (x s, θ s )P (s) = P (x s, θ s ) P (s s x,cj ) (3) What we propose here is to use the solution vectors produced by the other algorithms in order to improve the local results. In equation (3), we note s x,cj the cluster decided for data x by algorithm c j. Then, we assume that the P (s s x,cj ) are indepent from each others. Given this assumption, we can use the matrices Ψ previously retrieved and get the following equation: P (x s, θ s )P (s) = 1 Z P (x s, θ s) Ψ cj ci s x,cj,s (4) The general process of our collaborative method is shown in Algorithm 1. Algorithm 1: Collaborative Clustering for Heterogeneous Algorithms : General Framework Local step: for each clustering algorithm do Apply the clustering algorithm on the data X. Initialize the local parameters Θ Compute all Ψ ci cj matrices Collaboration step: while the results are not stable do for each clustering algorithm do Run one iteration of the modified algorithm using Equation 4. Update the solution vectors S Update the local parameters Θ Update all Ψ ci cj matrices
3 It is important to highlight that since the consensus matrices Ψ are relying on the original algorithm results to be built, there is no way that the collaborative process will have any positive effects on results that are extremely poor. In such a case the collaboration matrices built using the contingency probability would bring additional noise and may make the results even worse. Our method is inted to improve the results of clustering algorithms having already decent performances by their own. In this case, our collaboration model will help to detect and fix cluster labels that are unlikely given the results of the other algorithms included in the collaborative process. Furthermore, the process becomes more effective when the number of clusters to be found, the size of the data set, and the number of collaborating clustering algorithms increases. During our experiments, we have even detected that the worst case scenario would be to have only two mildly performing clustering algorithms working together and searching for only two classes. B. Exemples of collaborative schemes In this subsection, we introduce a few cases where our collaborative framework can be used. Fig. 1. Collaborative clustering algorithms working in parallel Figure 1 shows an example of 3 collaborative clustering algorithms working in parallel and improving each others results. This is the scenario of horizontal collaboration that we have been mentioning since the beginning of this article. Concrete examples of such a collaborations are: several algorithms working on several sites of a distributed data set, several algorithms accessing the same data features but looking for a different number of clusters (satellite image multiscale analysis for example), algorithms using different models working on a difficult dataset distributed or not. In all three cases, a mutual improvement on the results would be beneficial. Another possible scenario that our model can handle is shown in Figure 2. In this case, the EM algorithm uses the information from two others classifiers to improve its own results. Such a one sided information transfer would be beneficial if the two other clustering algorithms are specialized algorithms capable of detecting very specific elements of the observed data. The more generic EM algorithm working with those two clustering algorithms would then benefit from Fig. 2. Collaborative EM reinforced by two algorithms the information provided by the two specialized clustering algorithms. In this case, the process would be analog to a reinforcement learning process rather than a collaborative one. III. EXEMPLES OF APPLICATIONS OF OUR COLLABORATIVE FRAMEWORK In this section, we show how to use our collaborative framework with some Gaussian-Mixture Model based clustering algorithms. First, we show an application to the Expectation Maximization (EM) Algorithm [11] when considering a gaussian distribution, then we adapt our framework to the K-Means algorithm, and finally we show a more complex adaptation with to the ICM Algorithm for MRF-based image segmentation. While in this article we only show our method applied to clustering algorithms using Gaussian emission laws, it could be adapted to any algorithm using any other probabilistic model. This model is also compatible with fuzzy clustering algorithms albeit at the cost of more computation power to determine the confusion matrices. A. Collaborative EM and K-Means Algorithms In the following example, we suppose the clusters s 1..K to follow a Gaussian distribution. We consequently have θ s = {µ s, Σ s } where µ s is the average vector of cluster s and Σ s its co-variance matrix. Given this model, we can consider the K- Means algorithm to be a degenerate case of the EM algorithm with a co-variance matrix being the Identity. Then, in the case of the Expectation-Maximization algorithm, when the results of the collaborating algorithms are fixed, Equation (3) can be developed as follows: P (x s, θ s )P (s) = 1 Z N (µ s, Σ s, x) where: Ψ cj ci s x,cj,s (5) Z is a normalizing constant or partition function that is indepent of s. The Ψ cj ci are the consensus matrices mapping the other algorithms clusters to the ones of the current algorithm c i.
4 An expression similar to Equation (5) can be found for the K-Means Algorithm: P (x s, θ s )P (s) = 1 Z N (µ s, I d, x) Ψs cj ci x,cj,s (6) The original EM algorithm in its collaborative step is then modified as shown in Algorithm 2 Algorithm 2: Collaboration step for the EM Algorithm Retrieve the initial Ψ matrices while the results are not stable up to small variations do E-Step: S = arg max S P (X S, Θ)P (S): for each x X do s = arg max s P (x s, θ s )P (s) using Equation (5) M-Step: Θ = arg max Θ P (X S, Θ) Update Ψ B. Collaborative Hidden Markov Fields Model In this subsection we introduce an adaptation of our model to the specific case of the Hidden-Markov Model for computer vision [14] using the Expectation Maximization and Iterated Conditional Modes framework. In this case, the original algorithm already uses a Markov random Field based model to take into consideration the inner depencies between the data. Therefore, by adding our collaborative term, we get the following formula: 1 Z N (µ s, Σ s, x) v V x exp U N (s v,v) Ψ cj ci s x,cj,s (7) The previous equation can then be turned in to an MRF style energy function by applying a log function on the previous Equation. The result is shown in Equation (8): U MRF (s, x, Ψ) = 1 2 (x µ s) T Σ 1 s (x µ s ) + log( Σ s (2π) d ) + U N (s v, s) v V x Where: log(ψ cj ci ) s c j x,s V x is the set containing all the neighboring data of x in the data set U N (s v, s) is an inner neighborhood energy formula based on the clique potentials. The couple Expectation Maximization and Iterated Conditional Modes Algorithm (ICM) [15] is then modified as shown in Algorithm 3. (8) Algorithm 3: Collaborative EM+ICM Framework for MRF segmentation Local step: for each collaborating subset do Apply a clustering algorithm on the subset Initialize Θ and S Initialize Ψ for each collaboration link Collaboration step: while the results are not stable do for each x X do Minimize U MRF (s, x, Ψ) using Equation (8) Update Ψ return (S, Θ) IV. A. Data sets and indexes EXPERIMENTAL RESULTS In order to evaluate our approach, we have applied our algorithm on four data sets of different sizes and complexity: The Iris data set, Wisconsin Diagnostic Breast Cancer (wdbc) and the Spam Base data set from [16], as well as the VHR Strasbourg data set from [17]. We will give more details on the experiments done on this last data set, as it is by far the most difficult data set here. Iris data set (Iris): This data set has 150 instances of iris flowers described by 4 attributes. The flowers can be classified in 3 categories : Iris Setosa, Iris Versicolour and Iris Virginica. Wisconsin Diagnostic Breast Cancer (WDBC): This data set contains 569 instances having 32 parameters (an id, a diagnosis, and 30 real-valued attributes). Each observation is labeled as benign (357 observations) or malignant (212 observations). Spam Base: The Spam Base data set contains observations described by 57 attributes and a label column: Spam or not Spam (1 or 0). VHR Strasbourg: This data set contains observations described by 27 attributes. These data have been extracted from a very high resolution image from the French city of Strasbourg. This data set is much more complex because each data has neighborhood irregular depencies (between 1 and 15 neighbors deping on the data). A partial expert-based groundtruth was provided with this data set but it has a few caveats : First, it covers only 90% of the data. Then, the ground truth was hand-made by expert geographers and the areas they classified do not always match the preprocessed regions of our segmentation. It also features classes are quite difficult to detect such as specific types of crops, or multiples class of forest areas deping on the size of the area covered by the said forest. For these reasons, the indexes computed using this ground truth such as the Rand Index have to be taken with caution.
5 As criteria to validate our approach we have used an internal validation criterion: the Davies-Bouldin Index ; And two external criteria: The clustering accuracy for the data sets from the UCI repository, and the visual aspect of the result and Rand Index for the VHR Strasbourg data set. Davies-Bouldin Index The Davies-Bouldin Index (DB) [18] is a metric to evaluate clustering algorithms. It is an internal validity index that assess the clusters internal compactness and determines if they are well separated from each other. Rand Index The Rand Index [19] is a measure of agreement frequently used in cluster validation. It aims at measuring the agreement between two partitions one given by the clustering process and the other defined by external criteria such as an expert groundtruth. The Rand Index measures the relation between pairs of data set elements without using information from classes (labels) and can be used to detect problems with a clustering algorithm. In the following subsection, we use the Rand Index to compare our results on the VHR Strasbourg data set with the expert-based ground-truth. B. Results In the following, the DB index and accuracy index are shown before and after collaboration in order to analyze the impact of the collaboration on the clustering results. a) Iris: We have split the Iris data set into 2 subsets of size containing respectively the two first and two last attributes. Then, we ran the regular EM algorithm on the two subsets to assess the results of the algorithms to detect the 3 different varieties of Iris flowers. We also ran an extra EM algorithm looking for only two clusters (merging Versicolour and Virginica) on the first subset containing the first two attributes. And then we tried our collaborative EM using these 3 results to collaborate on the two subsets using the same method as the one shown in Figure 1. We note EM 1:2 the results of the regular EM algorithm working on the two first attributes, EM 3:4 the results on the two last attributes, and EM 1:2bis the results on the two first attributes looking for only 2 classes. The notation Algo x (y+z) is used to indicate that algorithm x is collaborating with algorithms y and z. TABLE II. EXPERIMENTAL RESULTS OF THE COLLABORATIVE APPROACH ON THE IRIS DATA SET data set algorithm Accuracy DB Index EM 1: % EM 3: % EM 1:2bis 86.67% EM 1:2 (3:4+1:2bis) 80.67% EM 1:2 3: % EM 3:4 1:2bis 73.33% EM 3:4 1: % EM 3:4 (1:2+1:2bis) 74.00% As one can see, the most interesting result on this data set are those achieved when EM 1:2 collaborates with EM 3:4 and EM 1:2bis. A 5% gain in the accuracy of the final clustering can be observed after the collaboration process: EM 12 (34+12bis). On the other hand the 11% gain observed in EM 1:2 3:4 is more problematic as we can clearly see that the results from the best clustering algorithm are simply replacing the one from the weaker clustering algorithm. The same phenomenon can be observed when the collaboration is done in the opposite direction: EM 3:4 1:2, the results from the weaker clustering algorithm are replaced. However this is not true for EM 34 1:2bis, where the weaker clustering algorithm EM 1:2bis affects the results of EM 3:4. Our interpretation is that in the later case EM 1:2bis can bring some diversity to EM 3:4, while EM 1:2 cannot. Furthermore, when we look at the collaborations EM 3:4 1:2bis, EM 3:4 1:2 and EM 3:4 (1:2+1:2bis), while two of them are deteriorating the results in term of classification purity, we can see that in the three cases the DB Index either remains the same or improves after the collaborative process. It means that in spite of a worst purity result the separation of the clusters is better after the collaboration process. Finally it is worth noting that EM 12 (34+12bis) and EM 3:4 (1:2+1:2bis) give different results, which proves that this collaboration framework is a directional process. TABLE III. EXPERIMENTAL RESULTS OF THE COLLABORATIVE APPROACH ON THE WDBC DATA SET data set Algorithm Purity DB Index EM 1: % EM 11: % EM 21: % EM 1:10 (11:20+21:30) 90.07% EM 11:20 (1:10+21:30) 74.87% EM 21:30 (1:10+11:20) 93.32% b) WDBC: In order to apply the proposed approach on the WDBC data set, we have split it into 3 subsets of size The first subset was featuring the first 10 attributes, the second one attributes 11 to 20 and the last one attributes 21 to 30. Then we ran regular version of the EM algorithm on the 3 subsets and have evaluated the purity and DB Index of the clustering results. And finally, we have run our collaborative EM algorithm on the 3 subsets again and we have used the results of the regular EM on the other 2 subsets for the collaboration process. This process is analog to the one shown in Figure 1. We note EM 1:10 the results of the regular EM algorithm on the 10 first attributes, EM 11:20 the results on the 10 middle attributes and EM 21:30 the results on the last 10 attributes. In EM 1:10 (11:20+21:30), we have the average EM algorithm collaborating with both a weaker and better performing EM, and as we can see in this case the results are almost unchanged, the purity and the DB Index is slightly better. In EM 11:20 (1:10+21:30) the weakest EM algorithm collaborates with two better EM algorithms, and as could be expected both the purity and DB Index are improved. In EM 21:30 (1:10+11:20) as the strongest EM algorithm is collaborating with two weaker ones, we can see that the purity is negatively impacted. However, due to the diversity between the different collaborating algorithms results, this negative impact remains minimal on the DB Index which actually slightly improves. For these 3 results the collaborative process proved to be
6 efficient for all algorithms when it comes to the DB Index. All algorithms scores for the DB Index were improved even when collaborating with weaker algorithms. It is however not the case for the purity which behaved in more predicable way: collaborations with stronger algorithms improved the purity while collaboration with weaker ones deteriorated it. It highlight the importance of choosing carefully the algorithms to collaborate with both in term of result quality and diversity. TABLE IV. EXPERIMENTAL RESULTS OF THE COLLABORATIVE APPROACH ON THE SPAM BASE DATA SET data set Algorithm Purity DB Index EM % EM % GT M a 86.96% GT M b 86.59% EM 1 (GT Ma+GT Mb ) 69.20% EM 2 (GT Ma+GT Mb ) 74.87% c) Spam Base: For our experiments on the Spam Base data set, during the local phase, the EM algorithm proved to give irregular and average results (EM 1 and EM 2 ). Therefore, instead of splitting the data set and risking even worse performances, we decided to reinforce the EM Algorithm results by enabling a collaboration with two results from the GTM (Generative Topographic Mappings) algorithms applied twice on the data set with different parameters [20]. This reinforcement process is analog to the one shown in Figure 2. With the help of the results acquired from the GTM algorithm, our collaborative EM s results have been greatly improved. In the case of EM 1 (GT Ma+GT M b ), from an EM algorithm giving results slightly better than random, we were able to get a sharp improvement in both the purity of the clustering and the DB Index. While the improvements shown in EM 2 (GT Ma+GT M b ) are less drastic, the positive effect of the collaborative reinforcement process cannot be denied. These results show the importance of the quality of the initial segmentation acquired during the local phase as it has to be taken into account for the collaboration. d) VHR Strasbourg: Unlike for the previous experiments, the exact number of clusters for this data set is unknown and deps on the level of detail that one wishes to have : From very large landscape elements such as cities, water bodies or forest areas, down to small details such as trees or cars, there are several possible scales of analysis that would lead to a different number of clusters. Expert geographers provided us a ground-truth with 15 classes for this data set, but it can be reduced to between 7 and 9 potential clusters when merging elements that couldn t possibly be distinguished by an unsupervised algorithm. Therefore, instead of splitting the data, we decided to go for a multi-scale approach on this data set and ran the Semantic Rich ICM algorithm described in [21] looking for different number of clusters: 7, 8 and 9 clusters to match with the expert ground truth, and 5 clusters as it gave the best results from a purely unsupervised point of view. We then proceeded to run the collaborative version for the MRF algorithms searching for 7 and 9 clusters collaborating with the 3 others results. This process is both a reinforcement process and a collaborative process. The first important remark is that unlike for our previous experiments, here the supervised index (Rand Index) always disagrees with the unsupervised Davies-Bouldin Index : Good results from a clustering point of view are here the worst ones when we compare them to our ground truth. As one can see, SRICM9 is the best result from a supervised point of view, while SRICM5 and SRICM7 are the best one from an unsupervised point of view. The collaboration SRICM7 SRICM5+SRICM 8+SRICM 9 improves the DB- Index value despite the poor DB-Index values of SRICM 9 and SRICM8. In this case, the collaboration leads to better clusters at the cost of a segmentation farther from the ground truth. In the case of SRICM9 SRICM5+SRICM 8+SRICM 7 however, the collaboration does not bring any positive effect and deteriorate both DB-Index and Rand Index results. Our current explanation for these diverging results in very similar cases is that it is caused by our asymmetric probabilistic confusion matrices. Since these matrices are asymmetric they probably lead to a collaborative process that is equally asymmetric and does not lead to the same improvement or deterioration of all the collaborating algorithms. For a better analysis of the results, all images shown in section C shall be visualized in color and full scale version. Furthermore, these images are only small extracts from the full image. When we look at the 7-cluster segmentations in Figures 4 and 7, the following changes can be noticed after the collaborative process: First, not all the clusters have the same prototypes, the dark green cluster (dense vegetation) from the segmentation without collaboration does not exist anymore and has been merged along with other vegetation clusters to form a single cluster for vegetation areas. The clusters representing the buildings are not exactly the same either. The main improvement can be seen when looking at the 3 big buildings near the top left corner of the image. Their segmentation is almost homogeneous after the collaboration process. A similar improvement can be seen on the right of the central stadium where the segmentation proposed after the collaborative process is correct whereas the one before is not. Another noticeable amelioration is that there are less unwanted areas wrongly labeled as water after the collaborative process. These observation are confirming what we had already hinted from the DB index result: the collaborative process improved the quality of the segmentation in this case. On the other hand, SRICM 9 (see Figure 6) who was the best segmentation (from a supervised and visual point of view) before the collaborative process has been deteriorated by the 3 weaker algorithms with which the collaborative process was enabled. The most obvious deterioration is the apparent merging of the red and orange clusters thus merging together roads and building in Figure 8. Furthermore, many smaller clusters have seen their population shrink. It is already well documented that MRF-based algorithms for image segmentation t to have the sparser clusters absorbed by the larger ones. This phenomenon is amplified by the fact that the collaboration process that we use is analog to using a second Markov Random Field on top of the original Markov Random Field used during the local step.
7 TABLE V. DB I NDEX R AND I NDEX RESULTS FOR THE DIFFERENT ALGORITHMS ON THE VHR S TRASBOURG DATA SET Algorithm SRICM 5 clusters SRICM 7 clusters SRICM 8 clusters SRICM 9 clusters SRICM 7 SRICM5 +SRICM8 +SRICM9 SRICM 9 SRICM5 +SRICM8 +SRICM7 Figure Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 DB Index Rand Index C. VHR Strasbourg Figures Fig. 3. A 5-cluster segmentation using the regular SRICM algorithm. Fig. 4. A 7-cluster segmentation using the SRICM algorithm. Fig. 5. Fig cluster segmentation using the SRICM algorithm. Fig. 7. Result of the 7-cluster segmentation after collaboration Fig. 8. Result of the 9-cluster segmentation after collaboration 8-cluster segmentation using the SRICM algorithm.
8 V. CONCLUSION In this study we have proposed a new collaborative framework for heterogeneous clustering algorithms in the context of horizontal collaboration. We have shown that our framework is flexible and can also be used in the context of reinforcement learning. The strength of our approach is that it does not need the subsets, prototypes or models used by the different collaborating algorithms to be shared during the collaboration step. Only the solution vectors produced by all algorithms need to be share. Our framework is therefore much more generic than the previously proposed collaborative methods. Our theoretical model has been validated on several data sets and the experimental results have shown competitive performances. However these results were obtained with at most 4 collaborating algorithms and have to be exted to contexts with more collaborating clustering algorithms and more data sets. This would certainly be our first perspective to ext this work. One could also argue that in our experiments, the collaborative process never s up with a results better than the one of the best algorithm result from the local phase. While it did not happen during our experiment, it does not mean that it cannot happen at all. Furthermore, the concept of best result in unsupervised learning is highly depent on the index considered and it is impossible to know in advance which algorithm will give the best results. Therefore, in most cases and especially with distributed data, our approach remains valid. Given this issue of strong and weak collaborators, another possible extension to this work would be to study the impact of clustering results quality and diversity in order to weight or choose with which algorithm to collaborate. ACKNOWLEDGMENT This work has been supported by the ANR Project COCLICO, ANR-12-MONU [5] W. Pedrycz and K. Hirota, A consensus-driven fuzzy clustering, Pattern Recogn. Lett., vol. 29, no. 9, pp , [6] W. Pedrycz, Collaborative fuzzy clustering, Pattern Recognition Letters, vol. 23, no. 14, pp , [7] M. Ghassany, N. Grozavu, and Y. Bennani, Collaborative clustering using prototype-based techniques, International Journal of Computational Intelligence and Applications, vol. 11, no. 3, [8] B. Depaire, R. Falcon, K. Vanhoof, and G. Wets, PSO Driven Collaborative Clustering: a Clustering Algorithm for Ubiquitous Environments, Intelligent Data Analysis, vol. 15, pp , January [9] M. Ghassany, N. Grozavu, and Y. Bennani, Collaborative clustering using prototype-based techniques. International Journal of Computational Intelligence and Applications, vol. 11, no. 3, [10] N. Grozavu and Y. Bennani, Topological collaborative clustering, Austr. J. Intelligent Information Processing Systems, vol. 12, no. 3, [11] A. P. Dempster, N. M. Laird, and D. B. Rubin, Maximum likelihood from incomplete data via the em algorithm, Journal of the Royal Statistical Society. Series B, vol. 39, no. 1, pp. 1 38, [12] J. Besag, On the statistical analysis of dirty pictures, Journal of the Royal Statistical Society. Series B, vol. 48, no. 3, pp , [13] G. Forestier, P. Gancarski, and C. Wemmert, Collaborative clustering with background knowledge. Data & Knowledge Engineering, vol. 69, no. 2, pp , [14] S. Roth and M. J. Black, Fields of experts, in Markov Random Fields for Vision and Image Processing, A. Blake, P. Kohli, and C. Rother, Eds. MIT Press, 2011, pp [15] Y. Zhang, M. Brady, and S. M. Smith, Segmentation of brain mr images through a hidden markov random field model and the expectation maximization algorithm. IEEE Trans. Med. Imaging, vol. 20, no. 1, pp , [16] A. Frank and A. Asuncion, UCI machine learning repository, [Online]. Available: [17] S. Rougier and A. Puissant, Improvements of urban vegetation segmentation and classification using multi-temporal pleiades images, 5th International Conference on Geographic Object-Based Image Analysis, p. 6, [18] D. L. Davies and D. W. Bouldin, A cluster separation measure, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 1, no. 2, pp , Feb [19] W. Rand, Objective criteria for the evaluation of clustering methods, Journal of the American Statistical Association., pp , [20] C. M. Bishop, M. Svensen, and C. K. I. Williams, Gtm: The generative topographic mapping, Neural Computation, vol. 10, pp , [21] J. Sublime, A. Cornuéjols, and Y. Bennani, A new energy model for the hidden markov random fields, in ICONIP 2014, Part II, Lecture Notes in Computer Science, vol. 8835, 2014, pp REFERENCES [1] A. Strehl and J. Ghosh, Cluster ensembles a knowledge reuse framework for combining multiple partitions. Journal of Machine Learning Research, vol. 3, pp , [2] A. D. Silva, Y. Lechevallier, F. de A. T. de Carvalho, and B. Trousse, Mining web usage data for discovering navigation clusters. in ISCC, P. Bellavista, C.-M. Chen, A. Corradi, and M. Daneshmand, Eds. IEEE Computer Society, 2006, pp [3] A. Cornuéjols and C. Martin, Unsupervised object ranking using not even weak experts, in ICONIP (1), ser. Lecture Notes in Computer Science, B.-L. Lu, L. Zhang, and J. T. Kwok, Eds., vol Springer, 2011, pp [4] K. Jong, J. Mary, A. Cornuéjols, E. Marchiori, and M. Sebag, Ensemble feature ranking, in PKDD, ser. Lecture Notes in Computer Science, J.-F. Boulicaut, F. Esposito, F. Giannotti, and D. Pedreschi, Eds., vol Springer, 2004, pp
A New Energy Model for the Hidden Markov Random Fields
A New Energy Model for the Hidden Markov Random Fields Jérémie Sublime 1,2, Antoine Cornuéjols 1, and Younès Bennani 2 1 AgroParisTech, INRA - UMR 518 MIA, F-75005 Paris, France {jeremie.sublime,antoine.cornuejols}@agroparistech.fr
More informationConsensus clustering by graph based approach
Consensus clustering by graph based approach Haytham Elghazel 1, Khalid Benabdeslemi 1 and Fatma Hamdi 2 1- University of Lyon 1, LIESP, EA4125, F-69622 Villeurbanne, Lyon, France; {elghazel,kbenabde}@bat710.univ-lyon1.fr
More informationLecture on Modeling Tools for Clustering & Regression
Lecture on Modeling Tools for Clustering & Regression CS 590.21 Analysis and Modeling of Brain Networks Department of Computer Science University of Crete Data Clustering Overview Organizing data into
More informationAn Information Theory based Approach to Multisource Clustering
An Information Theory based Approach to Multisource Clustering Pierre-Alexandre Murena 1,2, Jérémie Sublime 3,4, Basarab Matei 4 and Antoine Cornuéjols 2 1 LTCI - Télécom ParisTech, Paris, France 2 UMR
More informationComparison of supervised self-organizing maps using Euclidian or Mahalanobis distance in classification context
6 th. International Work Conference on Artificial and Natural Neural Networks (IWANN2001), Granada, June 13-15 2001 Comparison of supervised self-organizing maps using Euclidian or Mahalanobis distance
More informationNormalized Texture Motifs and Their Application to Statistical Object Modeling
Normalized Texture Motifs and Their Application to Statistical Obect Modeling S. D. Newsam B. S. Manunath Center for Applied Scientific Computing Electrical and Computer Engineering Lawrence Livermore
More informationConsensus Clustering. Javier Béjar URL - Spring 2019 CS - MAI
Consensus Clustering Javier Béjar URL - Spring 2019 CS - MAI Consensus Clustering The ensemble of classifiers is a well established strategy in supervised learning Unsupervised learning aims the same goal:
More informationLab 9. Julia Janicki. Introduction
Lab 9 Julia Janicki Introduction My goal for this project is to map a general land cover in the area of Alexandria in Egypt using supervised classification, specifically the Maximum Likelihood and Support
More informationUnderstanding Clustering Supervising the unsupervised
Understanding Clustering Supervising the unsupervised Janu Verma IBM T.J. Watson Research Center, New York http://jverma.github.io/ jverma@us.ibm.com @januverma Clustering Grouping together similar data
More informationAutomatic Group-Outlier Detection
Automatic Group-Outlier Detection Amine Chaibi and Mustapha Lebbah and Hanane Azzag LIPN-UMR 7030 Université Paris 13 - CNRS 99, av. J-B Clément - F-93430 Villetaneuse {firstname.secondname}@lipn.univ-paris13.fr
More informationCOLLABORATIVE CLUSTERING USING PROTOTYPE-BASED TECHNIQUES
International Journal of Computational Intelligence and Applications Vol. 11, No. 3 (2012) 1250017 (17 pages) #.c Imperial College Press DOI: 10.1142/S1469026812500174 COLLABORATIVE CLUSTERING USING PROTOTYPE-BASED
More informationDocument image binarisation using Markov Field Model
009 10th International Conference on Document Analysis and Recognition Document image binarisation using Markov Field Model Thibault Lelore, Frédéric Bouchara UMR CNRS 6168 LSIS Southern University of
More informationSome questions of consensus building using co-association
Some questions of consensus building using co-association VITALIY TAYANOV Polish-Japanese High School of Computer Technics Aleja Legionow, 4190, Bytom POLAND vtayanov@yahoo.com Abstract: In this paper
More informationUndirected Graphical Models. Raul Queiroz Feitosa
Undirected Graphical Models Raul Queiroz Feitosa Pros and Cons Advantages of UGMs over DGMs UGMs are more natural for some domains (e.g. context-dependent entities) Discriminative UGMs (CRF) are better
More informationSemi-Supervised Clustering with Partial Background Information
Semi-Supervised Clustering with Partial Background Information Jing Gao Pang-Ning Tan Haibin Cheng Abstract Incorporating background knowledge into unsupervised clustering algorithms has been the subject
More informationUnsupervised Learning and Clustering
Unsupervised Learning and Clustering Selim Aksoy Department of Computer Engineering Bilkent University saksoy@cs.bilkent.edu.tr CS 551, Spring 2008 CS 551, Spring 2008 c 2008, Selim Aksoy (Bilkent University)
More informationSupervised Variable Clustering for Classification of NIR Spectra
Supervised Variable Clustering for Classification of NIR Spectra Catherine Krier *, Damien François 2, Fabrice Rossi 3, Michel Verleysen, Université catholique de Louvain, Machine Learning Group, place
More informationProbabilistic Graphical Models Part III: Example Applications
Probabilistic Graphical Models Part III: Example Applications Selim Aksoy Department of Computer Engineering Bilkent University saksoy@cs.bilkent.edu.tr CS 551, Fall 2014 CS 551, Fall 2014 c 2014, Selim
More informationComputational Statistics The basics of maximum likelihood estimation, Bayesian estimation, object recognitions
Computational Statistics The basics of maximum likelihood estimation, Bayesian estimation, object recognitions Thomas Giraud Simon Chabot October 12, 2013 Contents 1 Discriminant analysis 3 1.1 Main idea................................
More informationCLASSIFICATION WITH RADIAL BASIS AND PROBABILISTIC NEURAL NETWORKS
CLASSIFICATION WITH RADIAL BASIS AND PROBABILISTIC NEURAL NETWORKS CHAPTER 4 CLASSIFICATION WITH RADIAL BASIS AND PROBABILISTIC NEURAL NETWORKS 4.1 Introduction Optical character recognition is one of
More informationContextual priming for artificial visual perception
Contextual priming for artificial visual perception Hervé Guillaume 1, Nathalie Denquive 1, Philippe Tarroux 1,2 1 LIMSI-CNRS BP 133 F-91403 Orsay cedex France 2 ENS 45 rue d Ulm F-75230 Paris cedex 05
More informationUnsupervised Learning and Clustering
Unsupervised Learning and Clustering Selim Aksoy Department of Computer Engineering Bilkent University saksoy@cs.bilkent.edu.tr CS 551, Spring 2009 CS 551, Spring 2009 c 2009, Selim Aksoy (Bilkent University)
More informationChapter 8 The C 4.5*stat algorithm
109 The C 4.5*stat algorithm This chapter explains a new algorithm namely C 4.5*stat for numeric data sets. It is a variant of the C 4.5 algorithm and it uses variance instead of information gain for the
More informationDetecting Clusters and Outliers for Multidimensional
Kennesaw State University DigitalCommons@Kennesaw State University Faculty Publications 2008 Detecting Clusters and Outliers for Multidimensional Data Yong Shi Kennesaw State University, yshi5@kennesaw.edu
More informationINF 4300 Classification III Anne Solberg The agenda today:
INF 4300 Classification III Anne Solberg 28.10.15 The agenda today: More on estimating classifier accuracy Curse of dimensionality and simple feature selection knn-classification K-means clustering 28.10.15
More informationShort Survey on Static Hand Gesture Recognition
Short Survey on Static Hand Gesture Recognition Huu-Hung Huynh University of Science and Technology The University of Danang, Vietnam Duc-Hoang Vo University of Science and Technology The University of
More informationImproving the Efficiency of Fast Using Semantic Similarity Algorithm
International Journal of Scientific and Research Publications, Volume 4, Issue 1, January 2014 1 Improving the Efficiency of Fast Using Semantic Similarity Algorithm D.KARTHIKA 1, S. DIVAKAR 2 Final year
More informationNorbert Schuff VA Medical Center and UCSF
Norbert Schuff Medical Center and UCSF Norbert.schuff@ucsf.edu Medical Imaging Informatics N.Schuff Course # 170.03 Slide 1/67 Objective Learn the principle segmentation techniques Understand the role
More informationClustering with Reinforcement Learning
Clustering with Reinforcement Learning Wesam Barbakh and Colin Fyfe, The University of Paisley, Scotland. email:wesam.barbakh,colin.fyfe@paisley.ac.uk Abstract We show how a previously derived method of
More informationClustering and Dissimilarity Measures. Clustering. Dissimilarity Measures. Cluster Analysis. Perceptually-Inspired Measures
Clustering and Dissimilarity Measures Clustering APR Course, Delft, The Netherlands Marco Loog May 19, 2008 1 What salient structures exist in the data? How many clusters? May 19, 2008 2 Cluster Analysis
More informationSelection of Scale-Invariant Parts for Object Class Recognition
Selection of Scale-Invariant Parts for Object Class Recognition Gy. Dorkó and C. Schmid INRIA Rhône-Alpes, GRAVIR-CNRS 655, av. de l Europe, 3833 Montbonnot, France fdorko,schmidg@inrialpes.fr Abstract
More informationInternational Journal of Data Mining & Knowledge Management Process (IJDKP) Vol.7, No.3, May Dr.Zakea Il-Agure and Mr.Hicham Noureddine Itani
LINK MINING PROCESS Dr.Zakea Il-Agure and Mr.Hicham Noureddine Itani Higher Colleges of Technology, United Arab Emirates ABSTRACT Many data mining and knowledge discovery methodologies and process models
More informationCollaborative clustering based on Algorithmic Information Theory
Collaborative clustering based on Algorithmic Information Theory Pierre-Alexandre Murena 1, Jérémie Sublime 2, Basarab Matei 3, et Antoine Cornuéjols 4 1 LTCI - Télécom ParisTech - Université Paris-Saclay,
More informationTime Stamp Detection and Recognition in Video Frames
Time Stamp Detection and Recognition in Video Frames Nongluk Covavisaruch and Chetsada Saengpanit Department of Computer Engineering, Chulalongkorn University, Bangkok 10330, Thailand E-mail: nongluk.c@chula.ac.th
More informationLesion Segmentation and Bias Correction in Breast Ultrasound B-mode Images Including Elastography Information
Lesion Segmentation and Bias Correction in Breast Ultrasound B-mode Images Including Elastography Information Gerard Pons a, Joan Martí a, Robert Martí a, Mariano Cabezas a, Andrew di Battista b, and J.
More informationClassification. Vladimir Curic. Centre for Image Analysis Swedish University of Agricultural Sciences Uppsala University
Classification Vladimir Curic Centre for Image Analysis Swedish University of Agricultural Sciences Uppsala University Outline An overview on classification Basics of classification How to choose appropriate
More informationUsing Machine Learning to Optimize Storage Systems
Using Machine Learning to Optimize Storage Systems Dr. Kiran Gunnam 1 Outline 1. Overview 2. Building Flash Models using Logistic Regression. 3. Storage Object classification 4. Storage Allocation recommendation
More informationMRF Based LSB Steganalysis: A New Measure of Steganography Capacity
MRF Based LSB Steganalysis: A New Measure of Steganography Capacity Debasis Mazumdar 1, Apurba Das 1, and Sankar K. Pal 2 1 CDAC, Kolkata, Salt Lake Electronics Complex, Kolkata, India {debasis.mazumdar,apurba.das}@cdackolkata.in
More informationShared Kernel Models for Class Conditional Density Estimation
IEEE TRANSACTIONS ON NEURAL NETWORKS, VOL. 12, NO. 5, SEPTEMBER 2001 987 Shared Kernel Models for Class Conditional Density Estimation Michalis K. Titsias and Aristidis C. Likas, Member, IEEE Abstract
More informationData Mining Chapter 9: Descriptive Modeling Fall 2011 Ming Li Department of Computer Science and Technology Nanjing University
Data Mining Chapter 9: Descriptive Modeling Fall 2011 Ming Li Department of Computer Science and Technology Nanjing University Descriptive model A descriptive model presents the main features of the data
More informationAn Efficient Model Selection for Gaussian Mixture Model in a Bayesian Framework
IEEE SIGNAL PROCESSING LETTERS, VOL. XX, NO. XX, XXX 23 An Efficient Model Selection for Gaussian Mixture Model in a Bayesian Framework Ji Won Yoon arxiv:37.99v [cs.lg] 3 Jul 23 Abstract In order to cluster
More informationExpectation Maximization (EM) and Gaussian Mixture Models
Expectation Maximization (EM) and Gaussian Mixture Models Reference: The Elements of Statistical Learning, by T. Hastie, R. Tibshirani, J. Friedman, Springer 1 2 3 4 5 6 7 8 Unsupervised Learning Motivation
More informationA Unified Framework to Integrate Supervision and Metric Learning into Clustering
A Unified Framework to Integrate Supervision and Metric Learning into Clustering Xin Li and Dan Roth Department of Computer Science University of Illinois, Urbana, IL 61801 (xli1,danr)@uiuc.edu December
More informationFuzzy Segmentation. Chapter Introduction. 4.2 Unsupervised Clustering.
Chapter 4 Fuzzy Segmentation 4. Introduction. The segmentation of objects whose color-composition is not common represents a difficult task, due to the illumination and the appropriate threshold selection
More informationIntroduction to Trajectory Clustering. By YONGLI ZHANG
Introduction to Trajectory Clustering By YONGLI ZHANG Outline 1. Problem Definition 2. Clustering Methods for Trajectory data 3. Model-based Trajectory Clustering 4. Applications 5. Conclusions 1 Problem
More informationMachine Learning. Unsupervised Learning. Manfred Huber
Machine Learning Unsupervised Learning Manfred Huber 2015 1 Unsupervised Learning In supervised learning the training data provides desired target output for learning In unsupervised learning the training
More informationChapter 7 UNSUPERVISED LEARNING TECHNIQUES FOR MAMMOGRAM CLASSIFICATION
UNSUPERVISED LEARNING TECHNIQUES FOR MAMMOGRAM CLASSIFICATION Supervised and unsupervised learning are the two prominent machine learning algorithms used in pattern recognition and classification. In this
More informationOutlier Ensembles. Charu C. Aggarwal IBM T J Watson Research Center Yorktown, NY Keynote, Outlier Detection and Description Workshop, 2013
Charu C. Aggarwal IBM T J Watson Research Center Yorktown, NY 10598 Outlier Ensembles Keynote, Outlier Detection and Description Workshop, 2013 Based on the ACM SIGKDD Explorations Position Paper: Outlier
More informationThe Data Mining Application Based on WEKA: Geographical Original of Music
Management Science and Engineering Vol. 10, No. 4, 2016, pp. 36-46 DOI:10.3968/8997 ISSN 1913-0341 [Print] ISSN 1913-035X [Online] www.cscanada.net www.cscanada.org The Data Mining Application Based on
More informationAutomatic Modularization of ANNs Using Adaptive Critic Method
Automatic Modularization of ANNs Using Adaptive Critic Method RUDOLF JAKŠA Kyushu Institute of Design 4-9-1 Shiobaru, Minami-ku, Fukuoka, 815-8540 JAPAN Abstract: - We propose automatic modularization
More informationTowards conflict resolution in collaborative clustering
Towards conflict resolution in collaborative clustering Germain Forestier, Cédric Wemmert and Pierre Gancarsi LSIIT - CNRS - University of Strasbourg - UMR 7005 Pôle API, Bd Sébastien Brant - 671 Illirch,
More informationPredicting Popular Xbox games based on Search Queries of Users
1 Predicting Popular Xbox games based on Search Queries of Users Chinmoy Mandayam and Saahil Shenoy I. INTRODUCTION This project is based on a completed Kaggle competition. Our goal is to predict which
More informationChapter 3: Supervised Learning
Chapter 3: Supervised Learning Road Map Basic concepts Evaluation of classifiers Classification using association rules Naïve Bayesian classification Naïve Bayes for text classification Summary 2 An example
More informationCOSC160: Detection and Classification. Jeremy Bolton, PhD Assistant Teaching Professor
COSC160: Detection and Classification Jeremy Bolton, PhD Assistant Teaching Professor Outline I. Problem I. Strategies II. Features for training III. Using spatial information? IV. Reducing dimensionality
More informationProbabilistic Abstraction Lattices: A Computationally Efficient Model for Conditional Probability Estimation
Probabilistic Abstraction Lattices: A Computationally Efficient Model for Conditional Probability Estimation Daniel Lowd January 14, 2004 1 Introduction Probabilistic models have shown increasing popularity
More information1. Introduction. 2. Motivation and Problem Definition. Volume 8 Issue 2, February Susmita Mohapatra
Pattern Recall Analysis of the Hopfield Neural Network with a Genetic Algorithm Susmita Mohapatra Department of Computer Science, Utkal University, India Abstract: This paper is focused on the implementation
More informationProcessing Missing Values with Self-Organized Maps
Processing Missing Values with Self-Organized Maps David Sommer, Tobias Grimm, Martin Golz University of Applied Sciences Schmalkalden Department of Computer Science D-98574 Schmalkalden, Germany Phone:
More informationGene Clustering & Classification
BINF, Introduction to Computational Biology Gene Clustering & Classification Young-Rae Cho Associate Professor Department of Computer Science Baylor University Overview Introduction to Gene Clustering
More informationCluster Validation. Ke Chen. Reading: [25.1.2, KPM], [Wang et al., 2009], [Yang & Chen, 2011] COMP24111 Machine Learning
Cluster Validation Ke Chen Reading: [5.., KPM], [Wang et al., 9], [Yang & Chen, ] COMP4 Machine Learning Outline Motivation and Background Internal index Motivation and general ideas Variance-based internal
More informationECG782: Multidimensional Digital Signal Processing
ECG782: Multidimensional Digital Signal Processing Object Recognition http://www.ee.unlv.edu/~b1morris/ecg782/ 2 Outline Knowledge Representation Statistical Pattern Recognition Neural Networks Boosting
More informationFeature Selection. CE-725: Statistical Pattern Recognition Sharif University of Technology Spring Soleymani
Feature Selection CE-725: Statistical Pattern Recognition Sharif University of Technology Spring 2013 Soleymani Outline Dimensionality reduction Feature selection vs. feature extraction Filter univariate
More informationA MRF Shape Prior for Facade Parsing with Occlusions Supplementary Material
A MRF Shape Prior for Facade Parsing with Occlusions Supplementary Material Mateusz Koziński, Raghudeep Gadde, Sergey Zagoruyko, Guillaume Obozinski and Renaud Marlet Université Paris-Est, LIGM (UMR CNRS
More informationPattern Recognition. Kjell Elenius. Speech, Music and Hearing KTH. March 29, 2007 Speech recognition
Pattern Recognition Kjell Elenius Speech, Music and Hearing KTH March 29, 2007 Speech recognition 2007 1 Ch 4. Pattern Recognition 1(3) Bayes Decision Theory Minimum-Error-Rate Decision Rules Discriminant
More informationRegion-based Segmentation
Region-based Segmentation Image Segmentation Group similar components (such as, pixels in an image, image frames in a video) to obtain a compact representation. Applications: Finding tumors, veins, etc.
More informationColor-Based Classification of Natural Rock Images Using Classifier Combinations
Color-Based Classification of Natural Rock Images Using Classifier Combinations Leena Lepistö, Iivari Kunttu, and Ari Visa Tampere University of Technology, Institute of Signal Processing, P.O. Box 553,
More informationColor Image Segmentation
Color Image Segmentation Yining Deng, B. S. Manjunath and Hyundoo Shin* Department of Electrical and Computer Engineering University of California, Santa Barbara, CA 93106-9560 *Samsung Electronics Inc.
More informationBoosting Algorithms for Parallel and Distributed Learning
Distributed and Parallel Databases, 11, 203 229, 2002 c 2002 Kluwer Academic Publishers. Manufactured in The Netherlands. Boosting Algorithms for Parallel and Distributed Learning ALEKSANDAR LAZAREVIC
More informationECS 234: Data Analysis: Clustering ECS 234
: Data Analysis: Clustering What is Clustering? Given n objects, assign them to groups (clusters) based on their similarity Unsupervised Machine Learning Class Discovery Difficult, and maybe ill-posed
More informationFUSION OF MULTITEMPORAL AND MULTIRESOLUTION REMOTE SENSING DATA AND APPLICATION TO NATURAL DISASTERS
FUSION OF MULTITEMPORAL AND MULTIRESOLUTION REMOTE SENSING DATA AND APPLICATION TO NATURAL DISASTERS Ihsen HEDHLI, Josiane ZERUBIA INRIA Sophia Antipolis Méditerranée (France), Ayin team, in collaboration
More informationA Visualization Tool to Improve the Performance of a Classifier Based on Hidden Markov Models
A Visualization Tool to Improve the Performance of a Classifier Based on Hidden Markov Models Gleidson Pegoretti da Silva, Masaki Nakagawa Department of Computer and Information Sciences Tokyo University
More informationEnhancing Forecasting Performance of Naïve-Bayes Classifiers with Discretization Techniques
24 Enhancing Forecasting Performance of Naïve-Bayes Classifiers with Discretization Techniques Enhancing Forecasting Performance of Naïve-Bayes Classifiers with Discretization Techniques Ruxandra PETRE
More informationPart I. Hierarchical clustering. Hierarchical Clustering. Hierarchical clustering. Produces a set of nested clusters organized as a
Week 9 Based in part on slides from textbook, slides of Susan Holmes Part I December 2, 2012 Hierarchical Clustering 1 / 1 Produces a set of nested clusters organized as a Hierarchical hierarchical clustering
More informationMachine Learning. B. Unsupervised Learning B.1 Cluster Analysis. Lars Schmidt-Thieme, Nicolas Schilling
Machine Learning B. Unsupervised Learning B.1 Cluster Analysis Lars Schmidt-Thieme, Nicolas Schilling Information Systems and Machine Learning Lab (ISMLL) Institute for Computer Science University of Hildesheim,
More informationMIT Samberg Center Cambridge, MA, USA. May 30 th June 2 nd, by C. Rea, R.S. Granetz MIT Plasma Science and Fusion Center, Cambridge, MA, USA
Exploratory Machine Learning studies for disruption prediction on DIII-D by C. Rea, R.S. Granetz MIT Plasma Science and Fusion Center, Cambridge, MA, USA Presented at the 2 nd IAEA Technical Meeting on
More informationA Population Based Convergence Criterion for Self-Organizing Maps
A Population Based Convergence Criterion for Self-Organizing Maps Lutz Hamel and Benjamin Ott Department of Computer Science and Statistics, University of Rhode Island, Kingston, RI 02881, USA. Email:
More informationMachine Learning nearest neighbors classification. Luigi Cerulo Department of Science and Technology University of Sannio
Machine Learning nearest neighbors classification Luigi Cerulo Department of Science and Technology University of Sannio Nearest Neighbors Classification The idea is based on the hypothesis that things
More informationImproving interpretability in approximative fuzzy models via multi-objective evolutionary algorithms.
Improving interpretability in approximative fuzzy models via multi-objective evolutionary algorithms. Gómez-Skarmeta, A.F. University of Murcia skarmeta@dif.um.es Jiménez, F. University of Murcia fernan@dif.um.es
More informationCOMBINING HIGH SPATIAL RESOLUTION OPTICAL AND LIDAR DATA FOR OBJECT-BASED IMAGE CLASSIFICATION
COMBINING HIGH SPATIAL RESOLUTION OPTICAL AND LIDAR DATA FOR OBJECT-BASED IMAGE CLASSIFICATION Ruonan Li 1, Tianyi Zhang 1, Ruozheng Geng 1, Leiguang Wang 2, * 1 School of Forestry, Southwest Forestry
More informationSwarm Based Fuzzy Clustering with Partition Validity
Swarm Based Fuzzy Clustering with Partition Validity Lawrence O. Hall and Parag M. Kanade Computer Science & Engineering Dept University of South Florida, Tampa FL 33620 @csee.usf.edu Abstract
More informationCluster Analysis. Angela Montanari and Laura Anderlucci
Cluster Analysis Angela Montanari and Laura Anderlucci 1 Introduction Clustering a set of n objects into k groups is usually moved by the aim of identifying internally homogenous groups according to a
More informationUsing a genetic algorithm for editing k-nearest neighbor classifiers
Using a genetic algorithm for editing k-nearest neighbor classifiers R. Gil-Pita 1 and X. Yao 23 1 Teoría de la Señal y Comunicaciones, Universidad de Alcalá, Madrid (SPAIN) 2 Computer Sciences Department,
More informationClustering CS 550: Machine Learning
Clustering CS 550: Machine Learning This slide set mainly uses the slides given in the following links: http://www-users.cs.umn.edu/~kumar/dmbook/ch8.pdf http://www-users.cs.umn.edu/~kumar/dmbook/dmslides/chap8_basic_cluster_analysis.pdf
More informationRemote Sensing & Photogrammetry W4. Beata Hejmanowska Building C4, room 212, phone:
Remote Sensing & Photogrammetry W4 Beata Hejmanowska Building C4, room 212, phone: +4812 617 22 72 605 061 510 galia@agh.edu.pl 1 General procedures in image classification Conventional multispectral classification
More informationLearning to Segment Document Images
Learning to Segment Document Images K.S. Sesh Kumar, Anoop Namboodiri, and C.V. Jawahar Centre for Visual Information Technology, International Institute of Information Technology, Hyderabad, India Abstract.
More informationNorth Asian International Research Journal of Sciences, Engineering & I.T.
North Asian International Research Journal of Sciences, Engineering & I.T. IRJIF. I.F. : 3.821 Index Copernicus Value: 52.88 ISSN: 2454-7514 Vol. 4, Issue-12 December-2018 Thomson Reuters ID: S-8304-2016
More informationNearly-optimal associative memories based on distributed constant weight codes
Nearly-optimal associative memories based on distributed constant weight codes Vincent Gripon Electronics and Computer Enginering McGill University Montréal, Canada Email: vincent.gripon@ens-cachan.org
More informationA Comparison of Resampling Methods for Clustering Ensembles
A Comparison of Resampling Methods for Clustering Ensembles Behrouz Minaei-Bidgoli Computer Science Department Michigan State University East Lansing, MI, 48824, USA Alexander Topchy Computer Science Department
More informationComparision between Quad tree based K-Means and EM Algorithm for Fault Prediction
Comparision between Quad tree based K-Means and EM Algorithm for Fault Prediction Swapna M. Patil Dept.Of Computer science and Engineering,Walchand Institute Of Technology,Solapur,413006 R.V.Argiddi Assistant
More informationClustering. Mihaela van der Schaar. January 27, Department of Engineering Science University of Oxford
Department of Engineering Science University of Oxford January 27, 2017 Many datasets consist of multiple heterogeneous subsets. Cluster analysis: Given an unlabelled data, want algorithms that automatically
More informationA Multiclassifier based Approach for Word Sense Disambiguation using Singular Value Decomposition
A Multiclassifier based Approach for Word Sense Disambiguation using Singular Value Decomposition Ana Zelaia, Olatz Arregi and Basilio Sierra Computer Science Faculty University of the Basque Country ana.zelaia@ehu.es
More informationA Content Based Image Retrieval System Based on Color Features
A Content Based Image Retrieval System Based on Features Irena Valova, University of Rousse Angel Kanchev, Department of Computer Systems and Technologies, Rousse, Bulgaria, Irena@ecs.ru.acad.bg Boris
More informationTextural Features for Image Database Retrieval
Textural Features for Image Database Retrieval Selim Aksoy and Robert M. Haralick Intelligent Systems Laboratory Department of Electrical Engineering University of Washington Seattle, WA 98195-2500 {aksoy,haralick}@@isl.ee.washington.edu
More informationData: a collection of numbers or facts that require further processing before they are meaningful
Digital Image Classification Data vs. Information Data: a collection of numbers or facts that require further processing before they are meaningful Information: Derived knowledge from raw data. Something
More informationContents. Preface to the Second Edition
Preface to the Second Edition v 1 Introduction 1 1.1 What Is Data Mining?....................... 4 1.2 Motivating Challenges....................... 5 1.3 The Origins of Data Mining....................
More informationCOMS 4771 Clustering. Nakul Verma
COMS 4771 Clustering Nakul Verma Supervised Learning Data: Supervised learning Assumption: there is a (relatively simple) function such that for most i Learning task: given n examples from the data, find
More informationSTA 4273H: Statistical Machine Learning
STA 4273H: Statistical Machine Learning Russ Salakhutdinov Department of Statistics! rsalakhu@utstat.toronto.edu! http://www.utstat.utoronto.ca/~rsalakhu/ Sidney Smith Hall, Room 6002 Lecture 12 Combining
More informationMotivation. Technical Background
Handling Outliers through Agglomerative Clustering with Full Model Maximum Likelihood Estimation, with Application to Flow Cytometry Mark Gordon, Justin Li, Kevin Matzen, Bryce Wiedenbeck Motivation Clustering
More informationNearest Clustering Algorithm for Satellite Image Classification in Remote Sensing Applications
Nearest Clustering Algorithm for Satellite Image Classification in Remote Sensing Applications Anil K Goswami 1, Swati Sharma 2, Praveen Kumar 3 1 DRDO, New Delhi, India 2 PDM College of Engineering for
More informationBackground Subtraction in Video using Bayesian Learning with Motion Information Suman K. Mitra DA-IICT, Gandhinagar
Background Subtraction in Video using Bayesian Learning with Motion Information Suman K. Mitra DA-IICT, Gandhinagar suman_mitra@daiict.ac.in 1 Bayesian Learning Given a model and some observations, the
More informationCHAPTER 6 PERCEPTUAL ORGANIZATION BASED ON TEMPORAL DYNAMICS
CHAPTER 6 PERCEPTUAL ORGANIZATION BASED ON TEMPORAL DYNAMICS This chapter presents a computational model for perceptual organization. A figure-ground segregation network is proposed based on a novel boundary
More information