SUFFUSE: Simultaneous Fuzzy-Rough FeatureSample Selection

Transcription

1 Journal o Advances in Inormation Technology Vol. 6, No. 3, August 205 SUFFUSE: Simultaneous Fuzzy-Rough FeatureSample Selection Javad Rahimipour Anaraki Department o Computer Science, Memorial University o Newoundland, St. John s, NL, Canada jra066@mun.ca Saeed Samet Faculty o Medicine, Memorial University o Newoundland, St. John s, NL, Canada saeed.samet@med.mun.ca Jeon-Hyun Lee and Chang-Wook Ahn Department o Computer Engineering, Sungkyunkwan University, Suwon, South Korea {ljh08375, cwan}@skku.edu datasets can also be decreased in terms o samples, Sample Selection (SS) methods have emerged to reduce size o datasets by removing irrelevant samples. Thereore, by employing FS and SS methods, datasets dimensions can be lowered and urther processing can be done more eiciently. Raman and Ioerger [], proposed a eature selection, and sample selection method. The ormer eliminates irrelevant eatures using a sequential search on eature space to maintain a balance between local hypotheses and prediction accuracy. The latter, uses Hamming distance to ilter out samples, and naive bayes classiier to predict class labels based on the selected samples. Then each method has been applied on a same dataset to perorm two dimensional selection. Rozsypal and Kubat [2] have introduced simultaneous eature-sample selection based on genetic algorithm with the aim o increasing classiication accuracy and decreasing the number o selected eatures and samples. Chromosome designation has been established to accommodate two subsets o integers, each representing selected eatures and samples. The itness unction has been designed based on the number o retained eatures and samples, and also the number o misclassiied examples. Rough Set Theory (RST) [3] is one o the most successul mathematical tools in FS [4] which nowadays receives much o attention in SS. This theory has been applied to many real-world applications [5] since it allows minimal representation o the data while sustaining semantic o data with no human provided inormation. However, RST is only decent to deal with crisp and discrete data; thereore, a combination o RST and Fuzzy Set has been proposed in [6] to overcome this inadequacy. Stand on Fuzzy-Rough set (FR), some research has been conducted in FS [5], [7] and SS [8], and very ew works have been done in simultaneous Fuzzy-Rough eature-sample selection [9]. Genetic Programming (GP) is capable o inding hidden relations in data and presenting them in terms o Abstract One o the most successul tools or modeling and dealing with uncertainty is Rough Set Theory. Based on this theory several Feature Selection methods have been proposed. As an extension, Fuzzy-Rough set has been introduced to deal with vagueness o both discrete and continuous data in Feature and Sample Selection methods. However, both Fuzzy-Rough Sample Selection and Simultaneous Fuzzy-Rough Feature-Sample Selection are investigated by ew. This paper proposes a novel Simultaneous Fuzzy-Rough Feature-Sample Selection method based on Shuled Frog Leaping Algorithm. The eectiveness o proposed method demonstrated and compared through its perormance resulting rom nine conventional as well as an improved mgp classiiers over iteen UCI datasets. This work is also applied to a real world classiication problem o noisy Functional Near-Inrared Spectroscopy neural signals. Experimental results show meaningul increase in classiication accuracy, and decrease in dataset size according to non-parametric statistical analysis. Index Terms uzzy-rough sets, simultaneous uzzy-rough eature-sample selection, eature selection, sample selection I. INTRODUCTION The amount o raw data produced daily is much higher than the inormation extracted rom them. Thereore, more cost and time are needed to process, save and maintain those data or later processing. Many problems in machine learning, data mining and pattern recognition involve big datasets. A high dimensional data in terms o number o eatures and samples needs huge eort to be processed. Thereore, Feature Selection (FS) methods can eectively reduce the size o datasets in one direction by selecting signiicant columns. These methods select mostinormative eatures which are highly correlated to the outcome and loosely depended on other eatures in avor o minimizing urther processing. Since the size o Manuscript received May 29, 205; revised July 20, 205. doi: /jait

2 mathematical unctions [0]. This method has been widely used in tough classiication problems and investigated by many researchers to develop classiiers or two- and multi-class problems. In [], An et al. designed a new multi-tree GP (mgp) classiier by modiying crossover and mutation operators. In this paper we have proposed a Simultaneous Fuzzy- Rough Feature-Sample Selection method (SUFFUSE) based on Shuled Frog Leaping Algorithm (SFLA) [2], as well as an improved mgp. The rest o the paper is organized as ollows: Section II describes preliminaries o FR, SFLA and mgp. Section III presents the proposed methods, SUFFUSE, and improved mgp. In Section IV, experimental results are shown. Application to noisy Functional Near-Inra-red Spectroscopy (NIRS) neural signals dataset and conclusion are placed in Sections V and VI, respectively. II. PRELIMINARIES Two undamental components o eature, sample and eature-sample selections are Evaluation Metric and Search Method. In this work the ormer is based on Fuzzy-Rough Positive Region (FRPR), and the latter uses SFLA. Finally, an improved mgp classiier analyzes and builds data models to igure out capabilities o proposed methods. All basics are categorized as ollows: A. Evaluation Metric: Fuzzy-Rough Positive Region (FRPR) In RST, data are organized in decision table. Let be the universe o discourse, R be the equivalence relation on, so approximation space is shown by (, R ). Let X be a subset o and P be a subset o A, which is a non-empty set o attributes. Approximating X using RST is done by means o lower and upper approximations. Objects in lower approximation ( PX ) are the ones which are surely classiied in X regarding the attributes in P. Upper approximation o X with regards to ( PX ) contains objects which are possibly classiied in X regarding the attributes in P. Based on these approximations, three dierent regions are deined as positive, negative and boundary that are shown by Equations, 2, and 3, respectively [3]. POS ( ) P Q PX () X / Q NEG ( ) P Q PX (2) X / Q BND ( ) P Q PX PX (3) X / Q X / Q B. Search Method: Shuled Frog Leaping Algorithm (SFLA) SFLA is a meta-heuristic search algorithm which is inspired by real rogs. The search starts by generating population over the search space. Then the population is divided into subpopulations called memeplexes which are able to evolve separately. In each memeplex, rogs participate in meme evolution due to inection by other rogs. By meme evolution, each rog s perormance is increased reerring to the best rog in each memeplex and poor ideas evolve toward new ideas. The rogs are inected both by best rogs in their memeplex and the entire population. Ater speciied number o evolutions, memeplexes are mixed together and new memeplexes are emerged by shuling the population. This process migrates rogs to dierent regions o the swamp. Thereore they can share their experiences with other rogs. A modiied binary orm o SFLA has been applied to the problem o simultaneous selection. C. Multi-tree Genetic Programming Classiier In [], individuals o a c-class problem are generated randomly with c trees. Then all the individuals are evaluated using itness unction and top N individuals are selected based on -wise tournament selection. The classiier continues by applying crossover and mutation or generating new individuals. Then, the worst individuals are substituted with the newly generated best ones and the classiier continues until the stopping criterion is satisied. III. PROPOSED METHODS A. Evaluation Metric: Fuzzy-Rough Positive Region (FRPR) The FRSS [4] is based on FRPR as an evaluation measure, and SFLA as a search method. The length o each rog in population is equal to the number o samples in the dataset where their presence and absence are depicted by one and zero, respectively. As SFLA generates initial population, related dataset ormations are constructed reerring to each individual rog. Then, itness o all rogs is calculated using FRPR as shown in Equation. Each rog s ormation is shown in Fig., where s j {0,} and j is number o samples o dataset. Figure. Each rog's ormation in FRSS Table I represents a dataset with two eatures and seven samples. Based on the table, a possible rog s ormation and related dataset is presented in Fig. 2 and Table II, respectively. TABLE I. Samples A DECISION TABLE Features 2 Class s Yes s No 2 s No 3 s Yes 4 s No 5 s Yes 6 s No 7 04

3 Table III demonstrates the inal dataset ormation based on the original dataset in Table I and by reerring to presence and absence o both eatures and samples in Fig. 4. Figure 2. A possible rog s ormation in FRSS TABLE II. RESULTING DATASET REFERRING TO POSSIBLE FROG S FORMATION Samples Features 2 Class s Yes s No 3 s Yes 4 s Yes 6 The SFLA continues until the stopping criterion, which is either maximum iteration or gaining the highest FRPR value, is satisied. Feature and sample selections can be done either in order or simultaneously. Applying either eature or sample selection beorehand might have a huge eect on the inal perormance. Even i the irst operation has a great eiciency, the outcome would be less desirable since each method acts independently. Thus, simultaneous selection would increase the quality o the outcome by considering ongoing two dimensional selections together. At the starting point, a population consists o rogs with the length proportional to the number o eatures and samples is generated. Fig. 3 depicts each rog s ormation. In this ormation, each bit s value and position show the presence or absence o either a eature or sample that speciies the inal structure o the extracted dataset rom the original one, where i, sj {0,}, and i and j are the number o eatures and samples in each dataset, respectively. Figure 3. Each rog s ormation with eatures and samples individuals Fig. 4 shows the ormation o each rog or aorementioned table. Since the irst position is equal to one, thereore the proportional eature should participate in the new dataset. Similarly, those samples which corresponding bits are equal to one will orm the output dataset. Figure 4. Possible rog s ormation with eatures and samples individuals TABLE III. RESULTING DATASET OF POSSIBLE FROG S FORMATION WITH FEATURE AND SAMPLES INDIVIDUALS Samples Features Class s 0.65 Yes s 0.48 No 3 s 0.70 Yes 4 s 0.05 Yes 6 Since Rough Set could not deal with continuous values, the original Fuzzy-Rough set has been proposed by Dubois and Prade [6] to elude this lack. Later, a new deinition was introduced by Radzikowska and Kerre [5] and then Shen and Jensen [5] modiied the original deinitions. In [7], inal deinitions o X -lower and X - upper approximations based on Fuzzy-Rough sets are presented as in Equations 4 and 5, where I is Łukasiewicz Fuzzy implicator, which is deined by min( x y,) and T is Łukasiewicz Fuzzy t -norm, which is shown by max( x y,0). ( x) in I{ ( x, y), ( y)} (4) RPX RP X y ( x) sup T{ ( x, y), ( y)} (5) RP X y RP ( x, y) { ( x, y)} (6) RP ap In Equation 6, R P is Fuzzy similarity relation and ( xy, ) is the degree o similarity between objects x Ra and y, considering eature a [7]. A uzzy similarity relation is shown in Equation 7, where a the variance o eature a. Positive region in RST is deined as a union o lower approximations. Reerring to extension principle [7], the membership o object x to a FRPR is deined in Equation 8. Ra ( a( y) ( a( x) a )), a R ( x, y) max min,0, P (( a( x) a ) a( y)) a POSP ( Q) PX X / Q X (7) ( x) sup ( x). (8) I the equivalence class o which x belongs to, does not belong to the positive region, obviously x will not be a part o the positive region. Equation 8 is the itness unction o the search algorithm which measures the signiicance o the selected eatures-samples subset [8]. Finally, SFLA evaluates each inal dataset corresponding to each rog by calculating FRPR. The best rog in each memeplex inects other rogs, and as a result the whole 05

4 population moves toward the inal goal, which is inding the lowest number o eatures and samples with the highest itness value. In the very irst point, dataset is loaded and the number o its eatures and samples, speciies all parameters o SFLA. In SUFFUSE, SFLA and FRPR collaborate to ind the best eature-sample subsets. Then the classiication methods, which involve conventional classiiers as well as improved mgp, classiy the datasets. The value o division o classiication accuracies mean by summation o the number o selected eatures and samples is calculated and compared with the results o the FRFS and FRSS. Fig. 5 shows the overall worklow o SUFFUSE. Figure 5. Simultaneous uzzy-rough eature-sample selection worklow B. Improved Multi-tree GP Classiier This method is robust to noise since the voting system is inspired by honey bee migration that is less sensitive to noise. Fig. 6 describes the method. Fig. 7 shows the representation o each individual with its equation reerring to the number o trees (m), which is speciied by user and number o classes. For instance a three-class dataset would have two classiiers. In the proposed classiier our main parts have been modiied as ollows: ) Fitness unction The new multi-modal itness unction is based on classiication accuracy and variance. The goal is to maximize the classiication margin, while decreasing intraclass similarities using Equation 9. Equation 0 calculates the centroid o each class to be used in Equation 9. Thereore itness unction is determined by the summation o Classiication Accuracy (CA) and distance unction as shown in Equation. Distance () i Tmx, CentroidA (9) Max( classa) Min( classa) iclassa () i Tmx, iclassa CentroidA classa (0) Fitness CA Distance () 2) Selection strategy The selection process has three stages. At irst top 3% o previous generation is selected to construct new generation, and i there were more than 3% individuals with highest ranking, top 0% will be selected. However, i two or more classiiers have the same itness value, all o them will be used in the next generation. Then 65% o the new generation is selected based on pair-wise tournament selection. Finally the rest o the individuals will be randomly generated. 3) Mutation The mutation process contains three policies or the internal mutation and one policy or the external one. In the internal mutation, a node can add, remove or exchange children. Thus the whole tree is reconstructed in the external mutation as Fig. 8 shows. 4) Crossover The crossover is divided into the internal and external crossovers. In the ormer, trees are selected in each individual based on the internal crossover probability parameter. The latter is based on one-point crossover and it takes place among any trees by considering external crossover probability. Fig. 9 describes the crossover strategy. IV. EXPERIMENTAL RESULTS Fiteen UCI datasets [6] have been selected to measure the perormance o the proposed methods. Parameter selection or SFLA has been ormulated based on the number o eatures F, samples S and eaturesamples FS using trial and error method. The results are mentioned in Table IV, in which m is the number o memeplexes, n is the number o rogs in each memeplex, N is the number o evolution processes, q is the number o rogs which are selected randomly rom n rogs to orm a memeplex and Smax is the maximum step size allowed to be adopted ater inection. TABLE IV. SFLA PARAMETERS FOR FRFS, FRSS AND SUFFUSE Method m n N q S max FRFS F 2.20 F 0.70 F x0.50 F 0.45 F 0.50 FRSS S 0.02 S 0.0 S 0.0 S 0.50 S 0.50 SUFFUSE FS 0.02 FS 0.0 FS 0.0 FS 0.50 FS 0.50 Each algorithm runs ten times over the datasets and inormation-rich eatures, samples, and eatures-samples are selected by FRFS, FRSS and SUFFUSE, respectively. The best results over all iterations are chosen and presented in Table V in terms o the number o selected eatures and samples and overall model size. The number o samples are ix in the results o FRFS as it only selects eatures, whereas, the number o eatures are constant or FRSS since it just aects samples. The mean o ranking or each method is calculated and shown in Table VI, in which SUFFUSE perorms 5% and 3% better than FRSS and FRFS, respectively. 06

5 Figure 6. Improved Multi-Tree GP Figure 7. An individual with m (n - ) trees Figure 8. Proposed mutation operator 07

6 Figure 9. Proposed crossover operator TABLE V. RESULTING REDUCTION AND MODEL SIZE BY FRFS, FRSS, & SUFFUSE Dataset Unreduced FRFS FRSS SUFFUSE S F Size S F Size S F Size S F Size Blood Trans Breast Cancer Breast Tissue Cleveland Glass Heart Ionosphere Lung Cancer Olitos Parkinson Pima Indian Dia Sonar Soybean SPECTF Heart Wine TABLE VI. RANKING OF FRFS, FRSS AND SUFFUSE BASED ON MODEL SIZE Dataset FRFS FRSS SUFFUSE Blood Trans. 3 2 Breast Cancer 3 2 Breast Tissue 3 2 Cleveland 2 3 Glass 2 3 Heart 2 3 Ionosphere 2 3 Lung Cancer 3 2 Olitos 3 2 Parkinson 2 3 Pima Indian Dia. 3 2 Sonar 3 2 Soybean 3 2 SPECTF Heart 3 2 Wine 2 3 Mean Table VII shows mean o the classiication results or conventional classiiers (such as PART, JRip, Naive Bayes, Bayes Net, J48, BFTree, FT, NBTree and RBFNetwrok, which are implemented in WEKA [7]) as well as improved mgp, and Fig. 0 presents the classiication worklow process. The mean o accuracies o conventional classiiers or our proposed method shows 3.55% increase comparing to unreduced datasets, as well as 2.55% and.58% improvement comparing with FRFS and FRSS, respectively. Whereas, the result o improved mgp or SUFFUSE shows 5.58%, 4.0% and.23% increase comparing to the results o improved GP or unreduced datasets, FRFS and FRSS. As the initial experiment results show, the usion o SUFFUSE with improved mgp produces the simplest model which leads to the higher classiication accuracies. Figure 0. Classiication worklow 08

7 TABLE VII. AVERAGE CLASSIFICATION ACCURACIES (%) OF CONVENTIONAL CLASSIFIERS (PART, JRIP, NAIVE BAYES, BAYES NET, J48, BFTREE, FT, NBTREE AND RBFNETWORK) AND IMPROVED MGP BASED ON FRFS, FRSS AND SUFFUSE RESULTS Dataset Unreduced FRFS FRSS SUFFUSE Conv. mgp Conv. mgp Conv. mgp Conv. mgp Blood Trans Breast Cancer Breast Tissue Cleveland Glass Heart Ionosphere Lung Cancer Olitos Parkinson Pima Indian Dia. Sonar Soybean SPECTF Heart Wine Mean A Non-parametric statistical analysis [8] is employed to compare the overall perormance o each method based on the results o improved mgp in Table VII. The average ranks obtained by each method in the Friedman test are presented in Table VIII. As shown, SUFFUSE has gained the lowest ranking, which proves the eectiveness o the proposed method. Friedman statistic (distributed according to chi-square with 3 degrees o reedom) is 27.56, and p-value computed by Friedman test is 4e 6. By reerring to the post hoc comparison results in Table IX, the probability o FRFS and Unreduced to perorm better than SUFFUSE is less (5e 3)% and (8e 3)%, respectively. Also, the probability o FRSS to outrun SUFFUSE is less than 48%. TABLE VIII. AVERAGE RANKINGS OF THE ALGORITHMS (FRIEDMAN) Algorithm Ranking SUFFUSE.4667 FRSS.8000 Unreduced FRFS TABLE IX. POST HOC COMPARISON TABLE FOR = 0:05 (FRIEDMAN) i algorithm z ( r R ) / SE p 3 FRFS Unreduced FRSS V. APPLICATION TO FUNCTIONAL NEAR-INFRARED SPECTROSCOPY (FNIRS) NEURAL SIGNALS To show the appropriateness o the proposed methods, a real world dataset called Neural Signal is used as a benchmark dataset. The neural signal acquisition has been done by a multi-channel optical brain imaging system (NIR-300) and the levels o oxy-, deoxy- and total-haemoglobin have been speciied using 6 signal channels at 2 Hz sampling rate. i Figure. Experimental scenarios or acquiring NIRS neural signals The signals are collected through the optical ibers, which are attached to the pre-rontal cortex. As Fig. shows, two cognitive activities o rest right imagery movement and rest let imagery movement have been sampled in a dataset with three classes, rest, right and let. The dataset has 280 samples and 45 eatures. Table X shows the average classiication accuracies o applying FRFS, FRSS and SUFFUSE. It can be seen that SUFFUSE ends to higher classiication accuracy comparing to unreduced, FRFS and FRSS, both by using conventional and improved mgp. The proposed classiication system results 5.83% higher than the other classiiers. TABLE X. AVERAGE CLASSIFICATION ACCURACIES (%) OF CONVENTIONAL CLASSIFIERS (PART, JRIP, NAIVE BAYES, BAYES NET, J48, BFTREE, FT, NBTREE AND RBFNETWORK) & IMPROVED MGP FOR UNREDUCED & REDUCED NEURAL SIGNAL DATASET USING FRFS, FRSS & SUFFUSE Dataset Neural Signal Unreduced FRFS FRSS SUFFUSE Conv. mgp Conv. mgp Conv. mgp Conv. mgp VI. CONCLUSION This paper proposes a novel Simultaneous Fuzzy- Rough Feature-Sample Selection (SUFFUSE), and an improved multi-tree GP (mgp). The SUFFUSE selects eatures and samples simultaneously by coding both in a single rog o SFLA, and use Fuzzy-Rough Positive Region (FRPR) as itness unction to evaluate selected subsets. An improved mgp classiier, classiies the results o proposed methods based on the new selection strategy, itness unction, mutation and crossover operators. Finally, the experimental results o SUFFUSE, Fuzzy-Rough Feature Selection (FRFS) and Fuzzy- Rough Feature Selection (FRSS) on iteen UCI datasets show the eectiveness o the proposed methods, both in terms o classiication accuracy and models size. As a real-world application, the proposed methods handle NIRS neural signal dataset. It can be seen rom the 09

8 results that SUFFUSE and mgp have a great impact on classiication accuracy comparing to independent eature and sample selections. As a uture work, we are so excited to apply improved version o SFLA, and perorm broad comparisons among dierent evolutionary algorithms. ACKNOWLEDGMENT This work has been partially supported by the Natural Sciences and Engineering Research Council o Canada (NSERC) and the Research & Development Corporation o Newoundland and Labrador (RDC). REFERENCES [] B. Raman and T. R. Ioerger, Enhancing learning using eature and example selection, Texas A&M University, College Station, TX, USA, [2] A. Rozsypal and M. Kubat, Selecting representative examples and attributes by a genetic algorithm, Intelligent Data Analysis, vol. 7, no. 4, pp , [3] Z. Pawlak, Rough sets, International Journal o Computer & Inormation Sciences, vol., no. 5, pp , 982. [4] J. R. Anaraki and M. Etekhari, Rough set based eature selection: A review, in Proc th Conerence on Inormation and Knowledge Technology, May 203, pp [5] Q. Shen and R. Jensen, Selecting inormative eatures with uzzyrough sets and its application or complex systems monitoring, Pattern Recognition, vol. 37, no. 7, pp , [6] D. Dubois and H. Prade, Putting rough sets and uzzy sets together, in Intelligent Decision Support, ser. Theory and Decision Library, R. Slowinski, Ed. Springer Netherlands, vol., 992, pp [7] R. Jensen and Q. Shen, New approaches to uzzy-rough eature selection, IEEE Transactions on Fuzzy Systems, vol. 7, no. 4, pp , Aug [8] R. Jensen and C. Cornelis, Fuzzy-rough instance selection, in Proc. 200 IEEE International Conerence on Fuzzy Systems (FUZZ), July 200, pp. 7. [9] N. Mac Parthal ain and R. Jensen, Simultaneous eature and instance selection using uzzy-rough bireducts, in Proc. 203 IEEE International Conerence on Fuzzy Systems (FUZZ), July 203, pp. 8. [0] D. P. Muni, N. R. Pal, and J. Das, A novel approach to design classiiers using genetic programming, IEEE Transactions on Evolutionary Computation, vol. 8, no. 2, pp , April [] J. An, J. Lee, and C. Ahn, An eicient gp approach to recognizing cognitive tasks rom nirs neural signals, Science China Inormation Sciences, vol. 56, no. 0, pp. 7, 203. [2] M. Eusu, K. Lansey, and F. Pasha, Shuled rog-leaping algorithm: a memetic meta-heuristic or discrete optimization, Engineering Optimization, vol. 38, no. 2, pp , [3] J. Komorowski, Z. Pawlak, L. Polkowski, and A. Skowron, Rough sets: A tutorial, in Rough-Fuzzy Hybridization: A New Trend in Decision Making, S. K. Pal and A. Skowron, Eds. Secaucus, NJ, USA: Springer- Verlag New York, Inc., 998, pp [4] J. R. Anaraki and C. Ahn, Fuzzy-rough sample selection, in Proc th International Workshop on Computer Science and Engineering (WCSE), August 204, pp [5] A. M. Radzikowska and E. E. Kerre, A comparative study o uzzy rough sets, Fuzzy Sets and Systems, vol. 26, no. 2, pp , [6] M. Lichman. (203). UCI machine learning repository. [Online]. Available: [7] M. Hall, E. Frank, G. Holmes, B. Pahringer, P. Reutemann, and I. H. Witten, The weka data mining sotware: An update, SIGKDD Explor. Newsl., vol., no., pp. 0 8, Nov [8] J. Alcala-Fdez, A. Fernandez, J. Luengo, J. Derrac, and S. Garcia, Keel data-mining sotware tool: Data set repository, integration o algorithms and experimental analysis ramework, Multiple- Valued Logic and Sot Computing, vol. 7, no. 2-3, pp , 20. Javad Rahimipour Anaraki is a member o e-health Research Unit. He received the B.E. degree in Computer Sotware rom the Shahid Ashrai Esahani University, Esahan, Iran, and the M.E. degrees in Computer Engineering rom the Shahid Bahonar University o Kerman, Kerman, Iran, in 2007 and 202, respectively. In 204, he joined the Department o Computer Science at Memorial University o Newoundland as Ph.D. student. His current research interests include uzzy-rough sets, eature and sample selection, and privacy-preserving methods. Saeed Samet is a member o the Faculty o Medicine as an Assistant Proessor at the e-health Research Unit. Prior to that, he was a Post- Doctoral Fellow at the e-health Inormation Laboratory at the Children Hospital o Eastern Ontario Research Institute (CHEO-RI) rom 200 to 202. During his post-doc ellowship he designed and developed several secure protocols or various health applications. He received his Ph.D. in Computer Science rom the University o Ottawa in 200 and his thesis was Privacy-Preserving Data Mining, in which he has proposed and designed dierent protocols on privacy-preserving methods or standard data mining and machine learning techniques. Jeon-Hyun Lee is a member o Sungkyunkwan Evolutionary Algorithm Lab (SEAL) at Sungkyunkwan University (SKKU), Republic o Korea. In 200, he joined the Department o Computer Science and Engineering at Sungkyunkwan University as Ph.D. student. His current research interests include genetic algorithm and swarm robotics. Chang Wook Ahn received Ph.D. degree at the Department o Inormation and Communications, Gwangju Institute o Science and Technology (GIST), Republic o Korea, in He was a Visiting Scholar at the Illinois Genetic Algorithms Laboratory (IlliGAL), University o Illinois at Urbana Champaign (UIUC), in From 2005 to 2007, he was a research sta member at the 4G Communication Research Group, Samsung Advanced Institute o Technology. From 2007 to 2008, he was a Research Proessor in GIST. He is currently an Associate Proessor in Sungkyunkwan University (SKKU), Republic o Korea. His research interests include genetic algorithms, multiobjective optimization, and the applications o evolutionary techniques to wireless networks. 0