Contribution of Data Complexity Features on Dynamic Classifier Selection
|
|
- Augusta Walton
- 5 years ago
- Views:
Transcription
1 Contribution of Data Complexity Features on Dynamic Classifier Selection André L. Brun, Alceu S. Britto Jr, Luiz S. Oliveira, Fabricio Enembreck, and Robert Sabourin Postgraduate Program in Informatics Pontifical Catholic University of Parana (PUCPR) Curitiba, Parana, Brazil {abrun, alceu, Postgraduate Program in Informatics Federal University of Parana (UFPR) Curitiba, Parana, Brazil École de Technologie Supérieure (ÉTS) Quebec University Montreal, Quebec, Canada Abstract Different dynamic classifier selection techniques have been proposed in the literature to determine among diverse classifiers available in a pool which should be used to classify a test instance. The individual competence of each classifier in the pool is usually evaluated taking into account its accuracy on the neighborhood of the test instance in a validation dataset. In this work we investigate the possible contribution of considering during the classifier evaluation the use of features related to the problem complexity. Since usually the pool generation technique does not assure diversity, the idea is to consider diversity during the selection. Basically, we select a classifier trained in subset of data showing similar complexity than that observed in neighborhood of the test instance. We expect that this similarity in terms of complexity allow us to select a more competent classifier. Experiments on 30 classification problems representing different levels of difficulty have shown that the proposed selection method is comparable to well known dynamic selection strategies. When compared with other DS approaches it was able to win on 123 over 150 experiments. This promising results indicate that further investigation must be done to increase diversity in terms of data complexity during the process of pool generation. I. INTRODUCTION Multiple Classifier Systems (MCS) have been proposed in the literature as an alternative for the hard task of training a monolithic approach in which a single classifier must be capable of learning the wide variability usually found in a pattern recognition problem [1]. The rationale behind this is that the classifiers in an MCS are diverse in the sense that they make different errors, and consequently, they show some complementarity [2]. Many researchers have focused on Multiple Classifier Systems and, consequently, new solutions have been dedicated to each of the three possible MCS phases: a) generation, b) selection, and c) integration. In the first phase, a pool of classifiers is generated; in the second, a subset of these classifiers is selected, while in the last phase, a final decision is made based on the prediction(s) of the selected classifier(s). In this context, the selection phase has received special attention of the research community. We can find approaches based on the selection of a single or an ensemble of classifiers from an initial pool, created using learning algorithms oriented by diversity and accuracy [3][4][5]. Such a selection can be performed in a static or dynamic fashion. In the former, the selection is done during training stage of the MCS. The same selected classifiers are used for all testing samples. In the later, the selection is executed during the testing stage of the MCS. For each testing sample a specific selection is executed. The success of a dynamic selection method depends on the adoption of a good criterion to evaluate the competence of classifiers in recognizing the test pattern to be labeled. The authors in [6] present a taxonomy to categorize the methods according to the criterion used for the fitness evaluation of the classifiers. According to them, the methods can be separated into two major groups: those based on individual competence and those which consider the relationship between the classifiers that compose the pool. Despite the high number of different strategies and aspects used to evaluate the performance of the classifiers in the pool, it is possible to observe the common use of accuracy evaluation which is usually combined with other sources of information. It this paper, we evaluate the contribution of features related to the level of difficulty of a classification problem extracted from data complexity analysis in the process of evaluating the competence of each classifier given a test instance. It is worth noticing that the complexity here, or difficulty, is
2 not restricted to only the number of instances, classes and features. In fact, it involves important aspects inherent to a classification problem that are estimated from complexity measures computed from the problem data. Among different aspects, the complexity measures usually attempt to describe and quantify how overlapping are two classes, how the border region between classes behaves, or even how is the spatial distribution of each class. The study presented here is inspired on works that try to find the most promising learning inducer for a specific classification problem taking into account its difficulty [7]. However, the idea is to investigate whether the level of difficulty estimated from the neighborhood of the test pattern in a validation dataset can contribute to compute the competence of the classifiers in the pool of an MCS. In summary, the research questions to be answered here are two, as follows: a) Could the use of information related to data complexity analysis contribute to estimated the competence of classifiers in an MCS based on dynamic selection? and b) Is there some relation between the distribution of training datasets in the complexity space and the observed performance of the proposed selection strategy? To answer those questions we have used an experimental protocol based on 30 classification problems representing different levels of difficulty. A pool of classifiers is constructed for each problem and for each training subset a complexity signature is created considering a set of complexity measures used to describe the problem difficulty. During the operational phase, similar signature is obtained for the neighborhood of the test instance using a validation dataset. The similarity between these complexity signatures, plus additional feature used to describe the test instance complexity are combined with accuracy information and applied to evaluate the competence of each classifier in the pool. In addition, we compared the obtained results with 5 different dynamic selection methods available in the literature. The experiments have shown that selecting the classifiers based on complexity analysis is a promising strategy. The proposed DS approach was able to win the single best (SB) on 28 over 30 experiments, and the combination of all classifiers on 23 over 30 experiments. When compared with other DS approaches it was able to win on 123 over 150 experiments. This work is organized into 6 sections. In Section 2, we present the complexity measures used to estimated the problem difficulty. Section 3 describes the proposed DS method, while Sections 4 and 5 present the experimental results and a further analysis, respectively. Finally, in Section 6 one may find the conclusion and future work. II. COMPLEXITY MEASURES The level of difficulty of a classification problem can be estimated using complexity measures applied on the data. The complexity measures proposed in the literature are usually classified into three categories [7], [8], as follows: a) classes overlapping; b) classes separability; and c) classes geometry, topology and density. In our work, we have considered three measures, one from each category. The rationale behind that is to use measures based on different concepts, showing low correlation. From the first category, we have used the F 1 (Fisher s Discriminant Ratio). This measure expresses how separable are two classes according to a specific feature. The F 1 metric can be interpreted as the distance between the center of two classes, so that the larger its value, larger the separation between the classes. Its calculation consists on comparing the means and the standard deviations for each class feature, in order to evaluate discrepancy level. Equation 1 presents how it is computed for each attribute specifically. The elements µ 1, µ 2, σ 1, and σ 2 correspond to the means and standard deviations for classes 1 and 2, respectively, for a specific feature space. The value adopted for F 1 will be the highest among all features. F 1 = (µ 1 µ 2 ) 2 σ 12 σ 2 2 (1) From the second category, we have used the N2 measure (the ratio of intra/inter class nearest neighbor distance). By applying N 2 we expect to determine how two classes are separable analyzing the existence and form of the border between classes. The idea is characterized by calculating the Euclidean distance between each element of the set to the nearest neighbor in the same class and to the closest neighbor outside the class. Then the distances between the elements of the same class are added and divided by the sum of the distances between instances of different classes, as shown in Equation 2, where δ(n 1 = (x i ), x i ) represents the distance between the instance i and its nearest neighbor from the same class. At the same time, δ(n 1 (x i), x i ) consists in the distance between the element i to the nearest element that belongs to a different class, while n is the number of instances. n i=1 N2 = δ(n 1 = (x i ), x i ) n i=1 δ(n 1 (x i), x i ) Finally, from the third category, we have selected N4 (the nonlinearity of the one-nearest neighbor classifier) which is based on the same idea of another measure named L3 (The nonlinearity of a linear classifier). According Ho & Basu [7] and Ho, Basu & Law [8], given a training set, L3 first creates a test set by linearly interpolating between randomly chosen pairs within the same class (the training set) also with random coefficients (different weights for each element). Then L3 correspond to the value of the error rate of the training data set versus test set applying a linear classifier. The measure N 4 follows the same principle, however, when calculating the error rate on the test set, instead of adopting a linear classifier, the 1NN classifier is employed. III. (2) N4 = 1NN error (dataset) (3) PROPOSED DS METHOD Different approaches can be found in the literature for dynamic selection of classifiers. The main difference among them is the criterion used to estimate the competence of the classifiers for a given test instance during the selection scheme. One may find measures of competence based on pure accuracy (overall local accuracy or local class accuracy) [9], [10], ranking of classifiers [11], probabilistic information [10],
3 Fig. 1. Training Stage - Pool Generation. Fig. 2. Operational Stage - Classifiers Selection. [12], classifier behavior calculated on output profiles [13], and oracle information [14][15]. Moreover, we may find measures that consider interactions among classifiers, such as diversity [16][17], ambiguity [18] or other grouping approaches [19]. In this work, our assumption is that the complexity of the neighborhood of the test instance computed on a validation dataset when combined with accuracy information can be used to estimate the competence of the classifiers in the pool. For this purpose the competence of a classifier is estimated considering its accuracy on a local region of the feature space for which we also compute the level of difficulty. Figures 1 and 2 present an overview of the proposed method, while its steps are described in the Algorithm 1. In the training stage, the training dataset of a given classification problem is used to generate a pool of initial classifiers using an ensemble learning technique (Bagging or Boosting) to provide diversity and accuracy (Fig. 1 - A). Afterwards, for each subset of data generated, a vector composed of M complexity measures is computed (Fig. 1 - B). This feature set is used as an M-complexity signature for each data subset (DS i ). During the operational phase (Fig. 2), the dynamic selection is done by estimating the competence of each classifier based on three features. In order to describe them let us consider DS i as the dataset used for training the classifier C i, while sig DSi is the M-complexity signature (F1, N2 and N4 measure values) computed from DS i. In addition, γ t as the K-neighborhood of the test instance t, while sig γt as the M-complexity signature computed from γ t. f 1i - Similarity in terms of complexity: Given a testing sample t, the first step is to define its neighborhood γ t in the validation dataset (Fig. 2 - A). Afterwards, sig γt is computed from γ t (Fig. 2 - B). The similarity between the complexity signature sig γt with each training dataset complexity signature sig DSi is done by means of Euclidean distance as denoted in Eq. 4. With this, we can determine the classifier trained on dataset showing similar complexity than that observed on the neighborhood of the test instance. f 1i = δ(sig γt, sig DSi ) (4) f 2i - Distance from the predicted class: Let us to consider y j as the class predicted by the classifier C i for the test instance t, DS i as the dataset used to train C i, and α ij as the centroid of the predicted class y j in the training dataset DS i. We compute the distance of the test instance t to the centroid α ij as shown in Eq. 5. The idea is to better describe the complexity space, since the complexity measures may show similar values to represent the difficulty between two classes even when they are differently distributed in the feature space.
4 f 2i = δ(t, α ij ) (5) f 3i - Local class accuracy: Consists on the local class accuracy of each classifier C i considering the class predicted (y j ) for the test instance t. This local accuracy is estimated on the neighborhood (γ t ) as denoted by Eq. 6. f 3i = Accuracy(C i, y j, γ t ) (6) The features are computed for each classifier. The final competence value of the classifier C i can be obtained by combining the three features. We have evaluated to combine them by using product and sum, both strategies have shown similar results. The final combination using the sum of them is presented in Eq. 7. Comp_ Ci = (1 f 1i) + (1 f 2i) + f 3i (7) where f 1i and f 2i correspond to the normalized metrics of f 1i and f 2i, respectively. They were normalized using the MinMax scaling as denoted on Eq. 8 for the feature f 1i. f 1i = f 1i f 1imin f 1imax f 1imin (8) The most promising classifier C is obtained (Fig. 2 - D) as described in the Eq. 9. C = argmax(comp_ Ci ) (9) Algorithm 1: DSOC - DS on Complexity Input: the pool C of M classifiers; training, validation and testing sets, Tr, Va and Te; and the neighborhood size K Output: C, the most promising classifier for each testing sample t in Te 1 for each classifier C i in the pool do 2 Compute the complexity signature sig DSi from data subset DS i ; 3 end 4 for each test t i in Te do 5 Find the γ t as the K-nearest-neighboors of the t i in Va; 6 Compute the complexity signature of γ t ; 7 for each classifier C i in the pool do 8 Compute features f 1, f 2, f 3 ; 9 Normalize f 1 and f 2 ; 10 Comp_ Ci = (1 f 1i ) + (1 f 2i ) + f 3i; 11 end 12 C = argmax(comp_ Ci ); 13 Use the classifier C to classify t i ; 14 end IV. EXPERIMENTS This section presents the experiments undertaken to evaluate the proposed DS method. A set of 30 datasets were used in these comparative experiments. Sixteen coming from the UCI machine learning repository [20], four from the KEEL (Knowledge Extraction based on Evolutionary Learning) repository [21], four from the Ludmila Kuncheva Collection of real medical data [22], four from the STATLOG project [23] and two artificial datasets generated with the Matlab PRTools toolbox. These datasets present only numeric features with no missing values. In addition, they have been frequently used to evaluate DS methods in the literature. Table I presents the details of each dataset. The experiments were conduced using 20 replications. For each replication, the datasets were randomly divided on a distribution of 50% of the dataset elements for training, 25% for validation and 25% for testing. For each problem a pool with 100 perceptrons was created using the Bagging technique [3]. Bags containing 10% or 20% of the training samples were used, depending on the size of the training sets (20% was used for the smaller datasets). The perceptron was used as base classifier since it represents an unstable and weak inducer. In addition, weak classifiers may emphasize the differences in terms of performance among the DS schemes [15]. The size of the neighborhood used to compute the complexity descriptors for each test instance was defined as 30. This value ensures for the used datasets the presence of elements of at least two classes in the neighborhood, thus it would be possible to calculate the metrics of complexity. However, in order to determine this size for the neighborhood we have carried out experiments on thirteen UCI problems, where the number of neighbors were varied from 20 to 50. As described before, we used three complexity metrics: F 1, N2 and N4. The idea was to use a descriptor of each category described in Section 2. To drive our choice, we carried out a study on thirteen UCI databases, in which we analyzed the correlation among all the 14 complexity metrics available in the DCoL library [24]. It was found that these three measures present low Pearson correlation among themselves, indicating that they may explain various phenomena together. In order to evaluate the contribution of the adoption of data complexity measures in the selection process, we compared the performance of the proposed method with 5 DS techniques already established in the literature, and also with the Single Best (SB) and combination of all classifiers (ALL). With respect to DS methods, we have implemented methods based on single classifier selection (LCA [9], OLA [9], and a Priori [25]) and also methods based on ensemble selection (KNORA- E and KNORA-U [15]). For these methods a neighborhood size K = 7 was used. This value was proved to be the most appropriate in previous studies [15], [26]. The average performance of each approach for each classification problem is shown in Table II. The boldfaced values in that table represents the highest accuracy for each problem. In the last column of Table II we have the oracle performance for each problem considering the given pool of classifiers. Such an upper limit in terms of performance is estimated considering the assumption that if at least one classifier can well recognize a given test pattern then the pool is also able to recognize it. It is possible to observe that the proposed DS method
5 TABLE I. MAIN FEATURES OF THE DATASETS USED IN THE EXPERIMENTS Dataset Instances Training Test Validation Features Classes % Bag Source Adult UCI Banana PRTools Blood UCI CTG UCI Diabetes UCI Ecoli UCI Faults UCI German STATLOG Glass UCI Haberman UCI Heart STATLOG ILPD UCI Segmentation UCI Ionosphere UCI Laryngeal LKC Laryngeal LKC Lithuanian PRTools Liver UCI Magic KEEL Mammo KEEL Monk KEEL Phoneme ELENA Sonar UCI Thyroid LKC Vehicle STATLOG Vertebral UCI WBC UCI WDVG UCI Weaning LKC Wine UCI TABLE II. COMPARISON OF THE PROPOSED METHOD OF DYNAMIC SELECTION BASED ON COMPLEXITY (DSOC) WITH THE SINGLE BEST (SB) CLASSIFIER IN THE POOL, THE COMBINATION OF ALL CLASSIFIERS (ALL), SOME DYNAMIC SELECTION METHODS LIKE OLA, LCA, A PRIORI, KNORA-U (KU), KNORA-E (KE), AND THE ORACLE PERFORMANCE. THE RESULTS ARE THE AVERAGE AND CORRESPONDING STANDARD DEVIATION OF 20 REPLICATIONS. THE BEST RESULTS ARE IN BOLDFACE. Dataset SB ALL OLA LCA a Priori KU KE DSOC Oracle Adult 83.6 (2.3) 86.7 (2.4) 82.4 (2.8) 82.3 (2.5) 80.6 (4.8) 76.6 (2.3) 71.0 (3.2) 85.6 (2.5) 99.7 (0.4) Banana 85.3 (1.4) 84.1 (1.4) 89.2 (1.9) 89.5 (1.9) 86.1 (2.5) 89.2 (1.4) 84.4 (1.9) 87.4 (2.4) 89.8 (1.9) Blood 76.4 (0.3) 76.4 (0.2) 74.2 (3.0) 74.2 (3.2) 69.0 (16.4) 76.4 (0.2) 76.4 (0.2) 72.7 (2.7) 100 (-) CTG 69.8 (11.1) 86.6 (1.7) 87.9 (1.1) 88.4 (1.2) 84.1 (1.6) 85.3 (0.9) 81.3 (1.0) 88.8 (1.1) 99.9 (0.1) Diabetes 66.0 (1.3) 64.5 (1.4) 69.9 (2.9) 70.0 (2.4) 58.6 (7.8) 65.5 (0.4) 65.1 (-) 69.4 (3.5) 92.3 (7.2) Ecoli 63.7 (3.9) 42.1 (0.6) 77.9 (3.8) 79.9 (2.9) 55.1 (9.2) 64.0 (4.2) 42.1 (0.6) 80.5 (3.7) 97.1 (1.7) Faults 31.2 (14.2) 63.5 (2.8) 64.9 (2.5) 66.4 (1.6) 51.4 (2.9) 53.6 (2.0) 36.7 (2.3) 67.6 (1.5) 99.2 (0.4) German 59.5 (5.0) 75.7 (2.5) 68.7 (2.9) 70.0 (2.9) 66.7 (3.4) 70.1 (0.3) 70.0 (-) 72.8 (2.4) 100 (-) Glass 56.6 (6.6) 58.0 (5.2) 59.9 (6.9) 60.7 (8.6) 46.4 (9.4) 49.3 (5.8) 33.6 (1.7) 63.1 (6.2) 99.8 (0.6) Haberman 75.3 (2.9) 73.7 (-) 75.3 (3.9) 74.9 (3.8) 73.9 (1.4) 73.8 (0.3) 73.7 (-) 76.4 (3.5) 88.8 (5.9) Heart 79.1 (4.9) 83.8 (3.2) 76.9 (3.3) 75.7 (4.3) 75.8 (6.3) 70.8 (4.1) 68.2 (3.6) 82.1 (3.4) 100 (-) ILPD 68.1 (3.5) 70.6 (3.5) 66.9 (3.0) 67.7 (3.2) 64.6 (6.0) 71.7 (-) 71.7 (-) 66.6 (2.9) 100 (0.1) Image 16.1 (5.6) 36.3 (1.1) 68.6 (3.0) 70.9 (2.6) 47.9 (3.8) 49.9 (1.9) 27.8 (1.2) 70.3 (2.4) 77.8 (2.7) Ionosphere 78.3 (2.8) 72.0 (2.6) 80.3 (2.6) 86.1 (3.2) 72.1 (4.9) 79.5 (6.2) 56.3 (15.3) 86.9 (3.2) 98.2 (1.9) Laryngeal (3.8) 78.6 (4.7) 79.4 (5.0) 79.8 (4.9) 76.2 (4.3) 69.2 (3.9) 66.9 (3.3) 82.4 (5.2) 99.9 (0.4) Laryngeal (5.1) 66.5 (3.3) 65.4 (5.3) 66.2 (4.9) 61.5 (5.5) 57.1 (4.0) 50.1 (3.6) 67.7 (4.0) 99.6 (0.7) Lithuanian 67.9 (6.5) 50.8 (0.5) 95.9 (1.1) 95.8 (1.2) 85.9 (2.7) 72.3 (3.2) 50.0 (-) 95.7 (2.5) 99.9 (0.5) Liver 65.6 (3.4) 59.5 (2.7) 64.5 (4.5) 66.7 (3.8) 54.1 (6.5) 49.9 (4.6) 41.9 (-) 66.0 (3.2) 100 (-) Magic 60.2 (9.5) 78.3 (0.6) 80.7 (0.6) 80.6 (1.7) 77.4 (0.6) 77.9 (0.5) 77.3 (0.5) 80.9 (0.8) 90.0 (0.5) Mammo 64.2 (14.4) 81.0 (2.6) 78.9 (2.1) 78.8 (1.6) 77.5 (3.5) 75.9 (2.4) 72.6 (2.4) 80.6 (2.4) 98.3 (1.0) Monk 78.4 (4.2) 80.5 (2.6) 86.5 (3.3) 86.5 (3.2) 77.5 (3.5) 63.8 (3.9) 55.1 (3.2) 89.3 (3.2) 100 (-) Phoneme 62.2 (6.6) 76.3 (0.8) 81.6 (1.1) 82.0 (0.9) 76.1 (1.4) 75.0 (0.6) 72.9 (0.8) 80.6 (1.1) 96.5 (0.5) Sonar 61.4 (9.0) 54.6 (1.9) 68.9 (6.8) 70.3 (6.8) 53.6 (5.9) 53.6 (1.3) 53.2 (0.9) 71.0 (7.7) 100 (-) Thyroid 93.3 (1.9) 94.4 (1.3) 94.3 (1.9) 95.6 (1.2) 90.1 (20.8) 72.0 (3.5) 21.4 (4.8) 94.5 (1.4) 100 (-) Vehicle 26.4 (3.8) 36.0 (5.7) 59.1 (3.7) 59.4 (3.2) 35.3 (5.9) 46.5 (3.1) 25.7 (0.2) 58.2 (7.5) (-) Vertebral 80.9 (2.9) 81.3 (3.6) 81.5 (4.9) 81.8 (4.3) 76.5 (5.3) 75.7 (2.9) 68.7 (1.3) 81.8 (3.7) 100 (-) WBC 85.3 (0.3) 53.6 (0.2) 92.7 (3.0) 93.2 (3.2) 81.7 (16.4) 88.3 (0.2) 62.9 (0.2) 92.5 (2.5) 100 (-) WDVG 44.6 (0.3) 83.4 (1.1) 80.1 (1.1) 80.4 (1.0) 77.2 (1.7) 65.2 (1.2) 61.5 (1.1) 82.4 (1.3) 99.9 (0.1) Weaning 76.9 (5.6) 79.3 (5.0) 77.0 (3.3) 76.9 (4.1) 71.7 (4.9) 58.4 (1.7) 53.5 (2.2) 82.9 (3.6) 100 (-) Wine 59.2 (8.7) 32.8 (1.1) 70.0 (5.2) 70.2 (4.9) 61.5 (6.0) 66.9 (4.3) 32.8 (1.2) 69.4 (6.4) 100 (-)
6 Fig. 3. Pairwise comparison of DSOC with all methods. The blue bars represent the number of problems where the adoption of complexity outperformed the competitor method, while red bars refer to the number of losses of the proposed approach Fig. 4. Graphical representation of Nemenyi test comparing all methods. The values presented next the method names correspond to its average rank considering the 30 classification problems surpassed the performance of the single best classifier on 28 over 30 experiments, and with respect to the combination of all classifiers in 23 over 30 experiments. When compared only with other DS approaches, it won in 123 over 150 experiments. Figure 3 shows the pairwise comparison with all tested methods. The blue bars represent the number of cases where there s a contribution in using the complexity measures and the red bars represents the numbers of problems where the proposed method loses. In one situation there is a tie between LCA (Vertebral dataset) and our approach. In order to compare the approaches behavior was performed Friedman s test with a confidence of 95% and a degree of freedom of 7 (once we are comparing 8 methods). For all thirty problems the null hypothesis was rejected, indicating that there s a significant difference in the accuracy among the approaches. A Nemenyi post-hoc test was performed to delineate the rank of the algorithms in all problems. The results are shown in Figure 4. It was possible to notice that our approach achieved the best ranking position in general. However the distance to OLA and LCA is smaller than the critical distance. Based on those results we can answer the first research question presented in this paper, which is related to whether the information related to data complexity analysis can contribute to estimated the competence of classifiers in an MCS based on dynamic selection. In fact, we can see some interesting contribution, however we should grasp why it could not be more significant. In the next section we discuss that, while we try to answer our second research question. V. FURTHER ANALYSIS Despite some interesting results observed in the last section, it is important to notice that the proposed method was not the best always. Thus, it is necessary to understand why in some cases there is some gain in performance and in other cases some loss is observed. With this in mind, we analyzed when the two situations occur. Considering that the proposed approach takes into account the similarity between the complexities of the data subset on which the classifier is trained and the neighborhood of test instance, we analyzed the behavior of F 1, N2 and N4 complexity descriptors. To this end, we quantified each complexity descriptor into one hundred bins. Thus, it was possible to compare both distributions, the one related to the subset used for training the classifier and that estimated from neighborhood of the test instances in the validation dataset. Figure 5 shows the behavior related to one replication selected from the 20 performed with the Monk (left side) and Sonar (right side) datasets. The proposed DS method have shown for the Monk dataset 5.5 percentage points of improvement (a significant gain) when compared to the observed second position in the rank (LCA method). On the other hand, we observed for the sonar dataset 3.4 percentage points
7 (a) (b) (c) (d) (e) (f) Fig. 5. Overlapping between complexity distributions, in red the distribution estimated from the neighborhood of each test instance, and in blue the distribution estimated from the training subsets: (a),(c) and (e) are related to the measures F1, N2 and N4 for the monk dataset; similarly (b),(d) and (f) are related to the sonar dataset. of loss (a significant loss) when compared to the observed second position in the rank (again the LCA method). One can observe the overlapping between complexity distributions, in red the distribution estimated from the neighborhood of each test instance, and in blue the distribution estimated from the training subsets. The distributions on the left side, Figures 5(a), 5(c) and 5(e) are related to the measures F 1, N2 and N4 for the Monk dataset for which the proposed method has shown promising results. Similarly, Figures 5(b), 5(d) and 5(f) are related to the Sonar dataset, for which the proposed method is not indicated. As one may see, when the overlap between distributions is more evident the contribution of the proposed DS approach is more significant. This may justify some efforts to investigate strategies to modify the ensemble learning algorithm by using complexity features to drive the generation of the training subsets that will be used to create the pool of classifiers. The rationale behind that is to better cover the complexity space of the problem in hand. VI. CONCLUSION We have evaluated the contribution of data complexity information to measure the competence of the classifiers in a dynamic selection method. For this purpose, local accuracy was combined with information related to the similarity of the data complexity estimated from the neighborhood of the test instance and from the subset of data used for training each classifier in the pool. The experiments on 30 datasets considering 20 replications have shown that the dynamic selection using complexity descriptors is a promising strategy when compared with the combination of all classifiers in the pool, and with 5 different DS methods available in the literature. Although promising results were achieved, there is still the need for further study on the influence of the data complexity on the selection process. Additional investigation is necessary to verify if we can improve the results by considering during the pool generation some strategy to better exploit the problem complexity space. An alternative would be to generate the pool of classifiers taking into account data subsets with different complexities. ACKNOWLEDGMENT This work was partially supported by CNPq (Conselho Nacional de Desenvolvimento Científico e Tecnológico), Brazil, and by CAPES (Coordenação de Aperfeiçoamento de Pessoal de Nível Superior), Brazil.
8 REFERENCES [1] V. Gunes, M. Manard, P. Loonis, and S. Petit-Renaud, Combination, cooperation and selection of classifiers: A state of the art, International Journal of Pattern Recognition and Artificial Intelligence, vol. 17, no. 08, pp , [2] L. I. Kuncheva and C. J. Whitaker, Measures of diversity in classifier ensembles and their relationship with the ensemble accuracy, Machine Learning, vol. 51, no. 2, pp , [3] L. Breiman, Bagging predictors, Machine Learning, vol. 24, no. 2, pp , [4] Y. Freund and R. E. Schapire, Experiments with a new boosting algorithm, in Proceedings of the 13th International Conference on Machine Learning, 1996, pp [5] T. K. Ho, The random subspace method for constructing decision forests, IEEE Trans. Pattern Anal. Mach. Intell., vol. 20, no. 8, pp , Aug [6] A. S. Britto Jr., R. Sabourin, and L. E. S. Oliveira, Dynamic selection of classifiers - a comprehensive review, Pattern Recognition, vol. 47, no. 11, pp , [7] T. K. Ho and M. Basu, Complexity measures of supervised classification problems, Pattern Analysis and Machine Intelligence, IEEE Transactions on, vol. 24, no. 3, pp , Mar [8] T. Ho, M. Basu, and M. Law, Measures of geometrical complexity in classification problems, in Data Complexity in Pattern Recognition, ser. Advanced Information and Knowledge Processing, M. Basu and T. Ho, Eds. Springer London, 2006, pp [9] K. Woods, W. P. Kegelmeyer, Jr., and K. Bowyer, Combination of multiple classifiers using local accuracy estimates, IEEE Trans. Pattern Anal. Mach. Intell., vol. 19, no. 4, pp , Apr [10] L. Didaci, G. Giacinto, F. Roli, and G. L. Marcialis, A study on the performances of dynamic classifier selection based on local accuracy estimation, Pattern Recognition, vol. 38, no. 11, pp , [11] M. Sabourin, A. Mitiche, D. Thomas, and G. Nagy, Classifier combination for hand-printed digit recognition, in Document Analysis and Recognition, 1993, Proceedings of the Second International Conference on, 1993, pp [12] G. Giacinto and F. Roli, Methods for dynamic classifier selection, in Proceedings of the 10th International Conference on Image Analysis and Processing, ser. ICIAP 99. Washington, DC, USA: IEEE Computer Society, 1999, pp [13] G. Giacinto, F. Roli, and G. Fumera, Selection of classifiers based on multiple classifier behaviour, in Proceedings of the Joint IAPR International Workshops on Advances in Pattern Recognition. London, UK, UK: Springer-Verlag, 2000, pp [14] L. Kuncheva and J. Rodriguez, Classifier ensembles with a random linear oracle, Knowledge and Data Engineering, IEEE Transactions on, vol. 19, no. 4, pp , [15] A. Ko, R. Sabourin, and A. Britto Jr., From dynamic classifier selection to dynamic ensemble selection, Pattern Recognition, vol. 41, no. 5, pp , [16] A. Santana, R. Soares, A. Canuto, and M. C. P. d. Souto, A dynamic classifier selection method to build ensembles using accuracy and diversity, in Neural Networks, SBRN 06. Ninth Brazilian Symposium on, Oct 2006, pp [17] Y. Yan, X.-C. Yin, Z.-B. Wang, X. Yin, C. Yang, and H.-W. Hao, Sorting-based dynamic classifier ensemble selection, in Document Analysis and Recognition (ICDAR), th International Conference on, Aug 2013, pp [18] E. dos Santos, R. Sabourin, and P. Maupin, Ambiguity-guided dynamic selection of ensemble of classifiers, in Information Fusion, th International Conference on, July 2007, pp [19] J. Xiao and C. He, Dynamic classifier ensemble selection based on gmdh, in Computational Sciences and Optimization, CSO International Joint Conference on, vol. 1, April 2009, pp [20] K. Bache and M. Lichman, UCI machine learning repository, [Online]. Available: [21] J. Alcalá-Fdez, A. Fernãndez, J. Luengo, J. Derrac, S. García, L. Sãnchez, and F. Herrera, Keel data-mining software tool: Data set repository, integration of algorithms and experimental analysis framework, Journal of Multiple-Valued Logic and Soft Computing, vol. 17, no. 2-3, pp , 2011, cited By 275. [22] L. Kuncheva, Statlog: Comparison of classification algorithms on large real-world problems, [Online]. Available: mas00a/activities/real_data.htm [23] R. D. King, C. Feng, and A. Sutherland, Statlog: Comparison of classification algorithms on large real-world problems, [24] A. Orriols-Puig, N. Macià, and T. K. Ho, Documentation for the data complexity library in c++, Barcelona, Spain, Tech. Rep., [25] G. Giacinto and F. Roli, Adaptive selection of image classifiers, in Image Analysis and Processing, ser. Lecture Notes in Computer Science, A. Bimbo, Ed., vol Springer Berlin Heidelberg, 1997, pp [26] R. M. Cruz, R. Sabourin, G. D. Cavalcanti, and T. Ing Ren, Meta-des: A dynamic ensemble selection framework using meta-learning, Pattern Recogn., vol. 48, no. 5, pp , May 2015.
Analyzing different prototype selection techniques for dynamic classifier and ensemble selection
Analyzing different prototype selection techniques for dynamic classifier and ensemble selection Rafael M. O. Cruz and Robert Sabourin École de technologie supérieure - Université du Québec Email: cruz@livia.etsmtl.ca,
More informationAnalyzing different prototype selection techniques for dynamic classifier and ensemble selection
Analyzing different prototype selection techniques for dynamic classifier and ensemble selection Rafael M. O. Cruz and Robert Sabourin École de technologie supérieure - Université du Québec Email: cruz@livia.etsmtl.ca,
More informationFrom dynamic classifier selection to dynamic ensemble selection Albert H.R. Ko, Robert Sabourin, Alceu Souza Britto, Jr.
From dynamic classifier selection to dynamic ensemble selection Albert H.R. Ko, Robert Sabourin, Alceu Souza Britto, Jr Eider Sánchez Contenidos 1. Introduction 2. Proposed dynamic ensemble selection KNORA
More informationOn dynamic ensemble selection and data preprocessing for multi-class imbalance learning
On dynamic ensemble selection and data preprocessing for multi-class imbalance learning Rafael M. O. Cruz Ecole de Technologie Supérieure Montreal, Canada rafaelmenelau@gmail.com Robert Sabourin Ecole
More informationarxiv: v2 [cs.lg] 11 Sep 2015
A DEEP analysis of the META-DES framework for dynamic selection of ensemble of classifiers Rafael M. O. Cruz a,, Robert Sabourin a, George D. C. Cavalcanti b a LIVIA, École de Technologie Supérieure, University
More informationDynamic Selection of Ensembles of Classifiers Using Contextual Information
Dynamic Selection of Ensembles of Classifiers Using Contextual Information Paulo R. Cavalin 1, Robert Sabourin 1, and Ching Y. Suen 2 1 École de Technologie Supérieure, 1100 Notre-dame ouest, Montreal(QC),
More informationDynamic Ensemble Construction via Heuristic Optimization
Dynamic Ensemble Construction via Heuristic Optimization Şenay Yaşar Sağlam and W. Nick Street Department of Management Sciences The University of Iowa Abstract Classifier ensembles, in which multiple
More informationIntegration Base Classifiers Based on Their Decision Boundary
Integration Base Classifiers Based on Their Decision Boundary Robert Burduk (B) Department of Systems and Computer Networks, Wroclaw University of Science and Technology, Wybrzeze Wyspianskiego 27, 50-370
More informationMultiple Classifier Fusion using k-nearest Localized Templates
Multiple Classifier Fusion using k-nearest Localized Templates Jun-Ki Min and Sung-Bae Cho Department of Computer Science, Yonsei University Biometrics Engineering Research Center 134 Shinchon-dong, Sudaemoon-ku,
More informationDynamic Class Prediction with Classifier Based Distance Measure
Dynamic Class Prediction with Classifier Based Distance Measure Şenay Yaşar Sağlam 1 W. Nick Street 2 1 University of Iowa, 108 John Pappajohn Business Building, Iowa City, IA 52242, U.S.A. Email: senay-yasarsaglam@uiowa.edu
More informationTakagi-Sugeno-Kanga Fuzzy Fusion In Dynamic Multi-Classifier System
Proceedings of the 2 nd World Congress on Electrical Engineering and Computer Systems and Science (EECSS 16) Budapest, Hungary - August 16-17, 2016 Paper No. MVML 108 DOI: 10.11159/mvml16.108 Takagi-Sugeno-Kanga
More informationA review of data complexity measures and their applicability to pattern classification problems. J. M. Sotoca, J. S. Sánchez, R. A.
A review of data complexity measures and their applicability to pattern classification problems J. M. Sotoca, J. S. Sánchez, R. A. Mollineda Dept. Llenguatges i Sistemes Informàtics Universitat Jaume I
More informationOPTIMIZATION OF BAGGING CLASSIFIERS BASED ON SBCB ALGORITHM
OPTIMIZATION OF BAGGING CLASSIFIERS BASED ON SBCB ALGORITHM XIAO-DONG ZENG, SAM CHAO, FAI WONG Faculty of Science and Technology, University of Macau, Macau, China E-MAIL: ma96506@umac.mo, lidiasc@umac.mo,
More informationKernel Combination Versus Classifier Combination
Kernel Combination Versus Classifier Combination Wan-Jui Lee 1, Sergey Verzakov 2, and Robert P.W. Duin 2 1 EE Department, National Sun Yat-Sen University, Kaohsiung, Taiwan wrlee@water.ee.nsysu.edu.tw
More informationAvailable online at ScienceDirect. Procedia Computer Science 35 (2014 )
Available online at www.sciencedirect.com ScienceDirect Procedia Computer Science 35 (2014 ) 388 396 18 th International Conference on Knowledge-Based and Intelligent Information & Engineering Systems
More informationAn Optimized Hill Climbing Algorithm for Feature Subset Selection: Evaluation on Handwritten Character Recognition
An Optimized Hill Climbing Algorithm for Feature Subset Selection: Evaluation on Handwritten Character Recognition Carlos M. Nunes, Alceu de S. Britto Jr.,2, Celso A. A. Kaestner and Robert Sabourin 3
More informationUnsupervised Feature Selection Using Multi-Objective Genetic Algorithms for Handwritten Word Recognition
Unsupervised Feature Selection Using Multi-Objective Genetic Algorithms for Handwritten Word Recognition M. Morita,2, R. Sabourin 3, F. Bortolozzi 3 and C. Y. Suen 2 École de Technologie Supérieure, Montreal,
More informationImproving accuracy and speed of Optimum-Path Forest classifier using combination of disjoint training subsets
Improving accuracy and speed of Optimum-Path Forest classifier using combination of disjoint training subsets Moacir P. Ponti-Jr. 1 and João P. Papa 2 1 Institute of Mathematical and Computer Sciences,
More informationDesign of Nearest Neighbor Classifiers Using an Intelligent Multi-objective Evolutionary Algorithm
Design of Nearest Neighbor Classifiers Using an Intelligent Multi-objective Evolutionary Algorithm Jian-Hung Chen, Hung-Ming Chen, and Shinn-Ying Ho Department of Information Engineering and Computer Science,
More informationA Modular Reduction Method for k-nn Algorithm with Self-recombination Learning
A Modular Reduction Method for k-nn Algorithm with Self-recombination Learning Hai Zhao and Bao-Liang Lu Department of Computer Science and Engineering, Shanghai Jiao Tong University, 800 Dong Chuan Rd.,
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 informationImproving the Random Forest Algorithm by Randomly Varying the Size of the Bootstrap Samples for Low Dimensional Data Sets
Improving the Random Forest Algorithm by Randomly Varying the Size of the Bootstrap Samples for Low Dimensional Data Sets Md Nasim Adnan and Md Zahidul Islam Centre for Research in Complex Systems (CRiCS)
More informationAn Empirical Comparison of Ensemble Methods Based on Classification Trees. Mounir Hamza and Denis Larocque. Department of Quantitative Methods
An Empirical Comparison of Ensemble Methods Based on Classification Trees Mounir Hamza and Denis Larocque Department of Quantitative Methods HEC Montreal Canada Mounir Hamza and Denis Larocque 1 June 2005
More informationPackage ECoL. January 22, 2018
Type Package Version 0.1.0 Date 2018-01-22 Package ECoL January 22, 2018 Title Compleity Measures for Classification Problems Provides measures to characterize the compleity of classification problems
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 informationBRACE: A Paradigm For the Discretization of Continuously Valued Data
Proceedings of the Seventh Florida Artificial Intelligence Research Symposium, pp. 7-2, 994 BRACE: A Paradigm For the Discretization of Continuously Valued Data Dan Ventura Tony R. Martinez Computer Science
More informationThe Role of Biomedical Dataset in Classification
The Role of Biomedical Dataset in Classification Ajay Kumar Tanwani and Muddassar Farooq Next Generation Intelligent Networks Research Center (nexgin RC) National University of Computer & Emerging Sciences
More informationMULTIVARIATE TEXTURE DISCRIMINATION USING A PRINCIPAL GEODESIC CLASSIFIER
MULTIVARIATE TEXTURE DISCRIMINATION USING A PRINCIPAL GEODESIC CLASSIFIER A.Shabbir 1, 2 and G.Verdoolaege 1, 3 1 Department of Applied Physics, Ghent University, B-9000 Ghent, Belgium 2 Max Planck Institute
More informationCombined Weak Classifiers
Combined Weak Classifiers Chuanyi Ji and Sheng Ma Department of Electrical, Computer and System Engineering Rensselaer Polytechnic Institute, Troy, NY 12180 chuanyi@ecse.rpi.edu, shengm@ecse.rpi.edu Abstract
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 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 informationA free software tool for automatic tuning of segmentation parameters
A free software tool for automatic tuning of segmentation parameters Pedro Achanccaray a, *, Victor Ayma a, Luis Jimenez b, Sergio Garcia b, Patrick Happ a, Raul Feitosa a,c, Antonio Plaza b a Dept. of
More informationInternational Journal of Research in Advent Technology, Vol.7, No.3, March 2019 E-ISSN: Available online at
Performance Evaluation of Ensemble Method Based Outlier Detection Algorithm Priya. M 1, M. Karthikeyan 2 Department of Computer and Information Science, Annamalai University, Annamalai Nagar, Tamil Nadu,
More informationAn Ensemble of Classifiers using Dynamic Method on Ambiguous Data
An Ensemble of Classifiers using Dynamic Method on Ambiguous Data Dnyaneshwar Kudande D.Y. Patil College of Engineering, Pune, Maharashtra, India Abstract- The aim of proposed work is to analyze the Instance
More informationLorentzian Distance Classifier for Multiple Features
Yerzhan Kerimbekov 1 and Hasan Şakir Bilge 2 1 Department of Computer Engineering, Ahmet Yesevi University, Ankara, Turkey 2 Department of Electrical-Electronics Engineering, Gazi University, Ankara, Turkey
More informationEfficient Pruning Method for Ensemble Self-Generating Neural Networks
Efficient Pruning Method for Ensemble Self-Generating Neural Networks Hirotaka INOUE Department of Electrical Engineering & Information Science, Kure National College of Technology -- Agaminami, Kure-shi,
More informationThe Combining Classifier: to Train or Not to Train?
The Combining Classifier: to Train or Not to Train? Robert P.W. Duin Pattern Recognition Group, Faculty of Applied Sciences Delft University of Technology, The Netherlands duin@ph.tn.tudelft.nl Abstract
More informationScalable Coding of Image Collections with Embedded Descriptors
Scalable Coding of Image Collections with Embedded Descriptors N. Adami, A. Boschetti, R. Leonardi, P. Migliorati Department of Electronic for Automation, University of Brescia Via Branze, 38, Brescia,
More informationLeave-One-Out Support Vector Machines
Leave-One-Out Support Vector Machines Jason Weston Department of Computer Science Royal Holloway, University of London, Egham Hill, Egham, Surrey, TW20 OEX, UK. Abstract We present a new learning algorithm
More informationAN IMPROVED K-MEANS CLUSTERING ALGORITHM FOR IMAGE SEGMENTATION
AN IMPROVED K-MEANS CLUSTERING ALGORITHM FOR IMAGE SEGMENTATION WILLIAM ROBSON SCHWARTZ University of Maryland, Department of Computer Science College Park, MD, USA, 20742-327, schwartz@cs.umd.edu RICARDO
More informationEnsemble Combination for Solving the Parameter Selection Problem in Image Segmentation
Ensemble Combination for Solving the Parameter Selection Problem in Image Segmentation Pakaket Wattuya and Xiaoyi Jiang Department of Mathematics and Computer Science University of Münster, Germany {wattuya,xjiang}@math.uni-muenster.de
More informationCost-Conscious Comparison of Supervised Learning Algorithms over Multiple Data Sets
Cost-Conscious Comparison of Supervised Learning Algorithms over Multiple Data Sets Mehmet Aydın Ulaş, Olcay Taner Yıldız, Ethem Alpaydın Technical Report, FBE/CMPE-01/2008-04 Institute of Graduate Studies
More informationEAR RECOGNITION AND OCCLUSION
EAR RECOGNITION AND OCCLUSION B. S. El-Desouky 1, M. El-Kady 2, M. Z. Rashad 3, Mahmoud M. Eid 4 1 Mathematics Department, Faculty of Science, Mansoura University, Egypt b_desouky@yahoo.com 2 Mathematics
More informationBagging and Boosting Algorithms for Support Vector Machine Classifiers
Bagging and Boosting Algorithms for Support Vector Machine Classifiers Noritaka SHIGEI and Hiromi MIYAJIMA Dept. of Electrical and Electronics Engineering, Kagoshima University 1-21-40, Korimoto, Kagoshima
More informationConstraint Projections for Ensemble Learning
Constraint Projections for Ensemble Learning Daoqiang Zhang Songcan Chen Zhi-Hua Zhou 2 Qiang Yang 3 Department of Computer Science & Engineering, Nanjing University of Aeronautics & Astronautics, China
More informationEvaluating the Replicability of Significance Tests for Comparing Learning Algorithms
Evaluating the Replicability of Significance Tests for Comparing Learning Algorithms Remco R. Bouckaert 1,2 and Eibe Frank 2 1 Xtal Mountain Information Technology 215 Three Oaks Drive, Dairy Flat, Auckland,
More informationClass dependent feature weighting and K-nearest neighbor classification
Class dependent feature weighting and K-nearest neighbor classification Elena Marchiori Institute for Computing and Information Sciences, Radboud University Nijmegen, The Netherlands elenam@cs.ru.nl Abstract.
More informationData Complexity in Pattern Recognition
Bell Laboratories Data Complexity in Pattern Recognition Tin Kam Ho With contributions from Mitra Basu, Ester Bernado-Mansilla, Richard Baumgartner, Martin Law, Erinija Pranckeviciene, Albert Orriols-Puig,
More informationNaïve Bayes for text classification
Road Map Basic concepts Decision tree induction Evaluation of classifiers Rule induction Classification using association rules Naïve Bayesian classification Naïve Bayes for text classification Support
More informationI. INTRODUCTION II. RELATED WORK.
ISSN: 2277 128X International Journal of Advanced Research in Computer Science and Software Engineering Research Paper Available online at: A New Hybridized K-Means Clustering Based Outlier Detection Technique
More informationDetector of Image Orientation Based on Borda-Count
Detector of Image Orientation Based on Borda-Count Loris Nanni and Alessandra Lumini DEIS, IEIIT CNR, Università di Bologna Viale Risorgimento 2, 40136 Bologna, Italy lnanni@deis.unibo.it Abstract. Accurately
More informationAn Efficient Learning of Constraints For Semi-Supervised Clustering using Neighbour Clustering Algorithm
An Efficient Learning of Constraints For Semi-Supervised Clustering using Neighbour Clustering Algorithm T.Saranya Research Scholar Snr sons college Coimbatore, Tamilnadu saran2585@gmail.com Dr. K.Maheswari
More informationPrototype Selection for Handwritten Connected Digits Classification
2009 0th International Conference on Document Analysis and Recognition Prototype Selection for Handwritten Connected Digits Classification Cristiano de Santana Pereira and George D. C. Cavalcanti 2 Federal
More informationResearch Article A New Robust Classifier on Noise Domains: Bagging of Credal C4.5 Trees
Hindawi Complexity Volume 2017, Article ID 9023970, 17 pages https://doi.org/10.1155/2017/9023970 Research Article A New Robust Classifier on Noise Domains: Bagging of Credal C4.5 Trees Joaquín Abellán,
More informationOptimization Model of K-Means Clustering Using Artificial Neural Networks to Handle Class Imbalance Problem
IOP Conference Series: Materials Science and Engineering PAPER OPEN ACCESS Optimization Model of K-Means Clustering Using Artificial Neural Networks to Handle Class Imbalance Problem To cite this article:
More informationEvaluating Ranking Composition Methods for Multi-Objective Optimization of Knowledge Rules
Evaluating Ranking Composition Methods for Multi-Objective Optimization of Knowledge Rules Rafael Giusti Gustavo E. A. P. A. Batista Ronaldo C. Prati Institute of Mathematics and Computer Science ICMC
More informationFeature Selection Technique to Improve Performance Prediction in a Wafer Fabrication Process
Feature Selection Technique to Improve Performance Prediction in a Wafer Fabrication Process KITTISAK KERDPRASOP and NITTAYA KERDPRASOP Data Engineering Research Unit, School of Computer Engineering, Suranaree
More informationEvolving SQL Queries for Data Mining
Evolving SQL Queries for Data Mining Majid Salim and Xin Yao School of Computer Science, The University of Birmingham Edgbaston, Birmingham B15 2TT, UK {msc30mms,x.yao}@cs.bham.ac.uk Abstract. This paper
More informationOutlier Detection Using Unsupervised and Semi-Supervised Technique on High Dimensional Data
Outlier Detection Using Unsupervised and Semi-Supervised Technique on High Dimensional Data Ms. Gayatri Attarde 1, Prof. Aarti Deshpande 2 M. E Student, Department of Computer Engineering, GHRCCEM, University
More informationEnhancing K-means Clustering Algorithm with Improved Initial Center
Enhancing K-means Clustering Algorithm with Improved Initial Center Madhu Yedla #1, Srinivasa Rao Pathakota #2, T M Srinivasa #3 # Department of Computer Science and Engineering, National Institute of
More informationFuzzy-Citation-KNN: a fuzzy nearest neighbor approach for multi-instance classification
Fuzzy-Citation-KNN: a fuzzy nearest neighbor approach for multi-instance classification Pedro Villar, Rosana Montes, Ana María Sánchez and Francisco Herrera Department of Software Engineering University
More informationOptimal Extension of Error Correcting Output Codes
Book Title Book Editors IOS Press, 2003 1 Optimal Extension of Error Correcting Output Codes Sergio Escalera a, Oriol Pujol b, and Petia Radeva a a Centre de Visió per Computador, Campus UAB, 08193 Bellaterra
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 informationAn automatic extraction method of the domains of competence for learning classifiers using data complexity measures
Knowl Inf Syst (2015) 42:147 180 DOI 10.1007/s10115-013-0700-4 REGULAR PAPER An automatic extraction method of the domains of competence for learning classifiers using data complexity measures Julián Luengo
More informationSimplifying OCR Neural Networks with Oracle Learning
SCIMA 2003 - International Workshop on Soft Computing Techniques in Instrumentation, Measurement and Related Applications Provo, Utah, USA, 17 May 2003 Simplifying OCR Neural Networks with Oracle Learning
More informationSeismic regionalization based on an artificial neural network
Seismic regionalization based on an artificial neural network *Jaime García-Pérez 1) and René Riaño 2) 1), 2) Instituto de Ingeniería, UNAM, CU, Coyoacán, México D.F., 014510, Mexico 1) jgap@pumas.ii.unam.mx
More informationCluster homogeneity as a semi-supervised principle for feature selection using mutual information
Cluster homogeneity as a semi-supervised principle for feature selection using mutual information Frederico Coelho 1 and Antonio Padua Braga 1 andmichelverleysen 2 1- Universidade Federal de Minas Gerais
More informationDynamic Human Fatigue Detection Using Feature-Level Fusion
Dynamic Human Fatigue Detection Using Feature-Level Fusion Xiao Fan, Bao-Cai Yin, and Yan-Feng Sun Beijing Key Laboratory of Multimedia and Intelligent Software, College of Computer Science and Technology,
More informationAnnealing Based Approach to Optimize Classification Systems
Proceedings of International Joint Conference on Neural Networks, Orlando, Florida, USA, August 12-17, 2007 Annealing Based Approach to Optimize Classification Systems Paulo V. W. Radtke, Robert Sabourin,
More informationorganizing maps generating abilities of as possible the accuracy of for practical classifier propagation and C4.5) [7]. systems.
Hirotaka Inoue, Kyoshiroo Sugiyama / International Journal of Computing, (), -6 computing@computingonline.net www.computingonline.net ISSN 77-69 International Journal of Computing SELF-ORGANIZING NEURAL
More informationOff-line Signature Verification Using Writer-Independent Approach
Off-line Signature Verification Using Writer-Independent Approach Luiz S. Oliveira, Edson Justino, and Robert Sabourin Abstract In this work we present a strategy for off-line signature verification. It
More informationImage retrieval based on bag of images
University of Wollongong Research Online Faculty of Informatics - Papers (Archive) Faculty of Engineering and Information Sciences 2009 Image retrieval based on bag of images Jun Zhang University of Wollongong
More informationThe Interpersonal and Intrapersonal Variability Influences on Off- Line Signature Verification Using HMM
The Interpersonal and Intrapersonal Variability Influences on Off- Line Signature Verification Using HMM EDSON J. R. JUSTINO 1 FLÁVIO BORTOLOZZI 1 ROBERT SABOURIN 2 1 PUCPR - Pontifícia Universidade Católica
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 informationApplying Supervised Learning
Applying Supervised Learning When to Consider Supervised Learning A supervised learning algorithm takes a known set of input data (the training set) and known responses to the data (output), and trains
More informationOn Error Correlation and Accuracy of Nearest Neighbor Ensemble Classifiers
On Error Correlation and Accuracy of Nearest Neighbor Ensemble Classifiers Carlotta Domeniconi and Bojun Yan Information and Software Engineering Department George Mason University carlotta@ise.gmu.edu
More informationClassification with Class Overlapping: A Systematic Study
Classification with Class Overlapping: A Systematic Study Haitao Xiong 1 Junjie Wu 1 Lu Liu 1 1 School of Economics and Management, Beihang University, Beijing 100191, China Abstract Class overlapping has
More informationEffect of the PSO Topologies on the Performance of the PSO-ELM
2012 Brazilian Symposium on Neural Networks Effect of the PSO Topologies on the Performance of the PSO-ELM Elliackin M. N. Figueiredo and Teresa B. Ludermir Center of Informatics Federal University of
More informationA Neural Network for Real-Time Signal Processing
248 MalkofT A Neural Network for Real-Time Signal Processing Donald B. Malkoff General Electric / Advanced Technology Laboratories Moorestown Corporate Center Building 145-2, Route 38 Moorestown, NJ 08057
More informationToward Part-based Document Image Decoding
2012 10th IAPR International Workshop on Document Analysis Systems Toward Part-based Document Image Decoding Wang Song, Seiichi Uchida Kyushu University, Fukuoka, Japan wangsong@human.ait.kyushu-u.ac.jp,
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 informationKeywords: clustering algorithms, unsupervised learning, cluster validity
Volume 6, Issue 1, January 2016 ISSN: 2277 128X International Journal of Advanced Research in Computer Science and Software Engineering Research Paper Available online at: www.ijarcsse.com Clustering Based
More informationIntroducing Partial Matching Approach in Association Rules for Better Treatment of Missing Values
Introducing Partial Matching Approach in Association Rules for Better Treatment of Missing Values SHARIQ BASHIR, SAAD RAZZAQ, UMER MAQBOOL, SONYA TAHIR, A. RAUF BAIG Department of Computer Science (Machine
More informationModel Based Sampling Fitting an ensemble of models into a single model
2015 International Conference on Computational Science and Computational Intelligence Fitting an ensemble of models into a single model Tony Lindgren Department of Computer and Systems Sciences Stockholm
More informationComparison Study of Different Pattern Classifiers
Comparison Study of Different Pattern s Comparison Study of Different Pattern s Ameet Joshi, Shweta Bapna, Sravanya Chunduri Abstract This paper presents a comparison study of the different parametric
More informationWeighting and selection of features.
Intelligent Information Systems VIII Proceedings of the Workshop held in Ustroń, Poland, June 14-18, 1999 Weighting and selection of features. Włodzisław Duch and Karol Grudziński Department of Computer
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 informationThe Un-normalized Graph p-laplacian based Semi-supervised Learning Method and Speech Recognition Problem
Int. J. Advance Soft Compu. Appl, Vol. 9, No. 1, March 2017 ISSN 2074-8523 The Un-normalized Graph p-laplacian based Semi-supervised Learning Method and Speech Recognition Problem Loc Tran 1 and Linh Tran
More informationEffect of Principle Component Analysis and Support Vector Machine in Software Fault Prediction
International Journal of Computer Trends and Technology (IJCTT) volume 7 number 3 Jan 2014 Effect of Principle Component Analysis and Support Vector Machine in Software Fault Prediction A. Shanthini 1,
More informationInduction of Multivariate Decision Trees by Using Dipolar Criteria
Induction of Multivariate Decision Trees by Using Dipolar Criteria Leon Bobrowski 1,2 and Marek Krȩtowski 1 1 Institute of Computer Science, Technical University of Bia lystok, Poland 2 Institute of Biocybernetics
More informationImage Classification Using Wavelet Coefficients in Low-pass Bands
Proceedings of International Joint Conference on Neural Networks, Orlando, Florida, USA, August -7, 007 Image Classification Using Wavelet Coefficients in Low-pass Bands Weibao Zou, Member, IEEE, and Yan
More informationSYMBOLIC FEATURES IN NEURAL NETWORKS
SYMBOLIC FEATURES IN NEURAL NETWORKS Włodzisław Duch, Karol Grudziński and Grzegorz Stawski 1 Department of Computer Methods, Nicolaus Copernicus University ul. Grudziadzka 5, 87-100 Toruń, Poland Abstract:
More informationStatistical Analysis of Metabolomics Data. Xiuxia Du Department of Bioinformatics & Genomics University of North Carolina at Charlotte
Statistical Analysis of Metabolomics Data Xiuxia Du Department of Bioinformatics & Genomics University of North Carolina at Charlotte Outline Introduction Data pre-treatment 1. Normalization 2. Centering,
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 informationCreating ensembles of classifiers via fuzzy clustering and deflection
Fuzzy Sets and Systems ( ) www.elsevier.com/locate/fss Creating ensembles of classifiers via fuzzy clustering and deflection Huaxiang Zhang a,, Jing Lu b a Department of Computer Science, Shandong Normal
More informationCalibrating Random Forests
Calibrating Random Forests Henrik Boström Informatics Research Centre University of Skövde 541 28 Skövde, Sweden henrik.bostrom@his.se Abstract When using the output of classifiers to calculate the expected
More informationA Bagging Method using Decision Trees in the Role of Base Classifiers
A Bagging Method using Decision Trees in the Role of Base Classifiers Kristína Machová 1, František Barčák 2, Peter Bednár 3 1 Department of Cybernetics and Artificial Intelligence, Technical University,
More informationFeature-weighted k-nearest Neighbor Classifier
Proceedings of the 27 IEEE Symposium on Foundations of Computational Intelligence (FOCI 27) Feature-weighted k-nearest Neighbor Classifier Diego P. Vivencio vivencio@comp.uf scar.br Estevam R. Hruschka
More informationSlant normalization of handwritten numeral strings
Slant normalization of handwritten numeral strings Alceu de S. Britto Jr 1,4, Robert Sabourin 2, Edouard Lethelier 1, Flávio Bortolozzi 1, Ching Y. Suen 3 adesouza, sabourin@livia.etsmtl.ca suen@cenparmi.concordia.ca
More informationTime Series Prediction as a Problem of Missing Values: Application to ESTSP2007 and NN3 Competition Benchmarks
Series Prediction as a Problem of Missing Values: Application to ESTSP7 and NN3 Competition Benchmarks Antti Sorjamaa and Amaury Lendasse Abstract In this paper, time series prediction is considered as
More information