Advances in artificial intelligence IBERAMIA SBIA 2006, jaime simao sichman, helder coelho, solange oliveira rezende, 2006 851

Lecture Notes in Artificial Intelligence Edited by J G Carbonell and J Siekmann Subseries of Lecture Notes in Computer Science 4140 Jaime Simão Sichman Helder Coelho Solange Oliveira Rezende (Eds.) Advances in Artificial Intelligence – IBERAMIA-SBIA 2006 2nd International Joint Conference: 10th Ibero-American Conference on AI 18th Brazilian AI Symposium Ribeirão Preto, Brazil, October 23-27, 2006 Proceedings 13 Series Editors Jaime G Carbonell, Carnegie Mellon University, Pittsburgh, PA, USA Jörg Siekmann, University of Saarland, Saarbrücken, Germany Volume Editors Jaime Simão Sichman University of São Paulo, Computer Engineering Department Av Prof Luciano Gualberto, tv 3, 158, 05508-970 São Paulo, SP, Brazil E-mail: jaime.sichman@poli.usp.br Helder Coelho University of Lisbon, Department of Informatics Edifício C6, Campo Grande, 1749-016 Lisboa, Portugal E-mail: hcoelho@di.fc.ul.pt Solange Oliveira Rezende University of São Paulo, Department of Computer Science Av Trabalhador São-Carlense, 400, 13560-970 São Carlos, SP, Brazil E-mail: solange@icmc.usp.br Library of Congress Control Number: 2006933826 CR Subject Classification (1998): I.2.6-7, I.2, F.1.1, F.4.2-3, F.2 LNCS Sublibrary: SL – Artificial Intelligence ISSN ISBN-10 ISBN-13 0302-9743 3-540-45462-4 Springer Berlin Heidelberg New York 978-3-540-45462-5 Springer Berlin Heidelberg New York This work is subject to copyright All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting, reproduction on microfilms or in any other way, and storage in data banks Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer Violations are liable to prosecution under the German Copyright Law Springer is a part of Springer Science+Business Media springer.com © Springer-Verlag Berlin Heidelberg 2006 Printed in Germany Typesetting: Camera-ready by author, data conversion by Scientific Publishing Services, Chennai, India Printed on acid-free paper SPIN: 11874850 06/3142 543210 Preface The Brazilian Artificial Intelligence (AI) community decided in 2004 to organize an International Joint Conference, joining IBERAMIA 2006 (the 10th Ibero-American Artificial Intelligence Conference), SBIA 2006 (the 18th Brazilian Artificial Intelligence Symposium), and SBRN 2006 (the 9th Brazilian Neural Networks Symposium) This decision was a consequence of the successful event organized in 2000, when the First International Joint Conference IBERAMIA/ SBIA 2000 (7th IberoAmerican Artificial Intelligence Conference and 15th Brazilian Artificial Intelligence Symposium) occurred in Brazil Moreover, in 2006 the artificial intelligence community celebrated the golden anniversary of the 1956 Dartmouth Conference that marked the beginning of artificial intelligence as a research field SBIA 2006 was the 18th conference of the SBIA conference series, which is the leading Brazilian conference for the presentation of AI research and applications Since 1995, SBIA has become an international conference, with papers written in English, an international Program Committee, and proceedings published in Springer’s Lecture Notes in Artificial Intelligence (LNAI) series IBERAMIA 2006 was the 10th conference of the IBERAMIA conference series, which has been one of the most suitable forums for Ibero-American AI researchers (from South and Central American countries, Mexico, Spain and Portugal) to present their results Following the SBIA and EPIA (Portuguese Conference on AI) experiences, from IBERAMIA 1998 on, it has also become an international conference with proceedings published in Springer’s LNAI series The International Joint Conference was held in Ribeir˜ ao Preto, Brazil, during October 23-27, 2006 The call for papers was very successful In the registration phase, when the authors were asked to submit just the title and abstract of their papers, 281 submissions were received In the second phase, 246 authors uploaded their full paper to be reviewed These submissions came from 23 different countries1 , as shown in Table After the revision process, 62 papers were accepted to be published in these proceedings In order to improve the overall quality of this joint conference, a double-blind reviewing procedure was adopted, and the acceptance rate was fixed around 25% of the submitted papers A large group of reviewers from all content areas was set up to carry out the difficult and challenging work of selecting the papers Every paper was evaluated by at least three reviewers, and ranked by the JEMS system Whenever conflicts were found, open discussions among the reviewers were triggered and analyzed later to support our final decision The conference included keynote speeches, introductory and advanced tutorials by world-leading researchers, several workshops covering specific topics and a thesis and dissertation contest In the case of several authors from different countries, we have considered the first author affiliation VI Preface Table IBERAMIA/SBIA 2006 Paper Submission Summary Country Registered Submitted Accepted Argentina 15 14 Australia 1 Brazil 149 128 33 Chile 8 China 3 – Colombia – – Cuba 3 Ecuador 2 – France 9 Germany 2 – Italy 1 Korea 5 – Malaysia 1 – Mexico 44 37 Poland 2 – Portugal Spain 18 17 Taiwan 1 – Turkey 2 – UK 1 – Uruguay 1 USA – – Venezuela 3 Total 281 246 62 We would like to thank all the authors for submitting their papers, as well as all members of the Program Committee and the additional referees for their hard work The high quality of the papers included in this volume would not be possible without their participation and diligence We also gratefully acknowledge the help of our colleagues whose support in the organization of this conference was invaluable Finally, we would like to express our gratitude to Alfred Hofmann and his staff at Springer for giving us again the opportunity to publish the proceedings in the LNAI series October 2006 Jaime Sim˜ ao Sichman Helder Coelho Solange Oliveira Rezende Organization The International Joint Conference IBERAMIA/SBIA/SBRN 2006 was organized by several AI research groups that belong to the University of S˜ ao Paulo (USP), Brazil: Instituto de Ciˆencias Matem´aticas e de Computa¸c˜ao (ICMC) Departamento de Ciˆencias de Computa¸c˜ao (SCE) Laborat´orio de Inteligˆencia Computacional (LABIC) Campus S˜ ao Carlos Faculdade de Filosofia, Ciˆencias e Letras de Ribeir˜ ao Preto (FFCLRP) Departamento de F´ısica e Matem´atica (DFM) Campus Ribeir˜ ao Preto Laborat´ orio de T´ecnicas Inteligentes (LTI) Departamento de Engenharia de Computa¸c˜ao e Sistemas Digitais (PCS) Escola Polit´ecnica (EP) Campus S˜ ao Paulo Instituto de Matem´ atica e Estat´ıstica (IME) Departamento de Ciˆencia da Computa¸c˜ao (MAC) Laborat´orio de L´ogica, Inteligˆencia Artificial e M´etodos Formais (LIAMF) Campus S˜ ao Paulo Escola de Artes, Ciˆencias e Humanidades (EACH) Curso de Sistemas de Informa¸c˜ao (SI) Campus S˜ ao Paulo Organizing Committee General Chair Solange Oliveira Rezende (LABIC/SCE/ICMC) General Co-chair Antonio Carlos Roque da Silva Filho (DFM/FFCLRP) Program Chairs Jaime Sim˜ ao Sichman (LTI/PCS/EP, Brazil) Helder Coelho (LABMAG/DI/FC/UL, Portugal) VIII Organization Workshop Chair Maria das Gra¸cas Volpe Nunes (LABIC/SCE/ICMC) Tutorial Chair Maria Carolina Monard (LABIC/SCE/ICMC) Local Organization Committee Alneu de Andrade Lopes (LABIC/SCE/ICMC) Flavio Soares Correa da Silva (LIAMF/MAC/IME) Ivandr´e Paraboni (SI/EACH) Jos´e de Jesus Perez Alcazar (SI/EACH) Leliane Nunes de Barros (LIAMF/MAC/IME) Marcelo Finger (LIAMF/MAC/IME) Renato Tinos (DFM/FFCLRP) Roseli Aparecida Francelim Romero (LABIC/SCE/ICMC) Sandra Maria Aluisio (LABIC/SCE/ICMC) Thiago Alexandre Salgueiro Pardo (LABIC/SCE/ICMC) Zhao Linag (LABIC/SCE/ICMC) IBERAMIA Steering Committee Alvaro Albornoz (ITESM, Mexico) Arlindo Oliveira (INESC, Portugal) Carlos Alberto Reyes (INAOE, Mexico) Christian Lemaitre (UAM, Mexico) Federico Barber (UPV, Spain) Francisco Garijo (TI+D, Spain) Helder Coelho (UL, Portugal) Jaime Sim˜ ao Sichman (USP, Brazil) Miguel Toro (US, Spain) SBIA Steering Committee Ana Bazzan (UFRGS, Brazil) Geber Ramalho (UFPE, Brazil) Gerson Zaverucha (UFRJ, Brazil) Guilherme Bittencourt (UFSC, Brazil) Maria Carolina Monard (USP, Brazil) Organization IX Supporting Scientific Societies SBC AEPIA APPIA SMIA Sociedade Brasileira de Computa¸c˜ao Associaci´on Espa˜ nola para Inteligencia Artificial Associa¸c˜ao Portuguesa para Inteligˆencia Artificial Sociedad Mexicana de Inteligencia Artificial Sponsoring Institutions The International Joint Conference 2006 had the financial support of the following Brazilian institutions: FAPESP CNPq CAPES FINEP Funda¸c˜ao de Amparo a` Pesquisa Estado de S˜ ao Paulo Conselho Nacional de Pesquisa Coordena¸c˜ao de Aperfei¸coamento Pessoal de N´ıvel Superior Financiadora de Estudos e Projetos Program Committee Adolfo Guzm´an-Arenas Adina Florea Alberto Oliart Aldo Dragoni Alejandro Zunino Alexander Gelbukh Alexis Drogoul Alipio Jorge Alvaro Albornoz Amal El Fallah Seghrouchni Ana Bazzan Ana Paiva Ana Serrano Ana Cristina Garcia Ana Teresa Martins Analia Amandi Instituto Polit´ecnico Nacional (Mexico) University Politehnica of Bucharest (Romania) Instituto Tecnol´ ogico y de Estudios Superiores de Monterrey (Mexico) Universit` a Politecnica delle Marche (Italy) Universidad Nac del Centro de la Pcia de Buenos Aires (Argentina) Instituto Polit´ecnico Nacional (Mexico) Institut de Recherche pour le D´eveloppement (France) Universidade Porto (Portugal) Instituto Tecnol´ ogico y de Estudios Superiores de Monterrey (Mexico) Universit´e Paris VI (France) Universidade Federal Rio Grande Sul (Brazil) Universidade T´ecnica de Lisboa (Portugal) Universidad Nacional EAD (Spain) Universidade Federal Fluminense (Brazil) Universidade Federal Cear´ a (Brazil) Universidad Nac del Centro de la Pcia de Buenos Aires (Argentina) X Organization Andr´e Carvalho Andre Valente Andrea Omicini Angel Kuri Anna Helena Reali Costa Antˆ onio Rocha Costa Ariadne Carvalho Arlindo Oliveira Arturo Hernandez Aguirre Augustin Lux Augusto Loureiro da Costa Brian Mayoh Carles Sierra Carlos Alberto Reyes Carlos Brizuela Carlos Castro Carlos Coello Carlos Ribeiro Catherine Tessier Catholijn Jonker Celso Kaestner Christian Lemaitre Cora Toledo Cristiano Castelfranchi David Hales Universidade de S˜ ao Paulo (Brazil) Knowledge Systems Ventures LLC (USA) Universit` a di Bologna (Italy) Instituto Tecnol´ ogico Aut´ onomo de M´exico (Mexico) Universidade de S˜ ao Paulo (Brazil) Universidade Cat´olica de Pelotas (Brazil) Universidade de Campinas (Brazil) Instituto de Engenharia de Sistemas e Computadores (Portugal) Centro de Investigaci´ on en Matem´ aticas (Mexico) Institut National de Recherche en Informatique et en Automatique (France) Universidade Federal da Bahia (Brazil) Aarhus University (Denmark) Institut d’Investigaci´ on en Intellig`encia Artificial (Spain) ´ Instituto Nacional de Astrof´ısica Optica y Electr´ onica (Mexico) Centro de Investigaci´on Cient´ıfica y de Educaci´on Superior de Enseada (Mexico) Universidad T´ecnica Federico Santa Mar´ıa (Chile) Instituto Polit´ecnico Nacional (Mexico) Instituto Tecnol´ ogico da Aeron´ autica (Brazil) Office National d’Etudes et de Recherches A´erospatiales (France) Nijmegen Institute for Cognition and Information (The Netherlands) Pontif´ıcia Universidade Cat´ olica-PR (Brazil) Universidad Aut´ onoma Metropolitana (Mexico) Laboratorio Nacional de Inform´ atica Avanzada (Mexico) Istituto di Scienze e Tecnologie della Cognizione (Italy) Universit´a di Bologna (Italy) 620 F.G Cozman, C.P de Campos, and J.C.F da Rocha Hailperin, T.: Best possible inequalities for the probability of a logical function of events American Mathematical Monthly 72 (1965) 343–359 Nilsson, N.J.: Probabilistic logic Artificial Intelligence 28 (1986) 71–87 de Finetti, B.: Theory of Probability, vol 1-2 Wiley, New York (1974) 10 Bruno, G., Gilio, A.: Applicazione del metodo del simplesso al teorema fondamentale per le probabilita nella concezione soggettivistica Statistica 40 (1980) 337–344 11 Coletti, G., Scozzafava, R.: Stochastic independence in a coherent setting Annals of Mathematics and Artificial Intelligence 35 (2002) 151–176 12 Biazzo, V., Gilio, A., Lukasiewicz, T., Sanfilippo, G.: Probabilistic logic under coherence: Complexity and algorithms In: International Symposium on Imprecise Probabilities and Their Applications, Shaker (2001) 51–61 13 Andersen, K.A., Hooker, J.N.: Bayesian logic Decision Support Systems 11 (1994) 191–210 14 Coletti, G., Scozzafava, R.: Zero probabilities in stochastic independence Information, Uncertainty, Fusion (2000) 185–196 15 Vantaggi, B.: Conditional Independence and Graphical Models in a Coherent Setting PhD thesis, Universit`a di Perugia, Italy (1999) 16 Lakshmanan, L.V.S., Sadri, F.: Probabilistic deductive databases In: Symposium on Logic Programming (1994) 254–268 17 Ng, R., Subrahmanian, V.S.: Probabilistic logic programming Information and Computation 101(2) (1992) 150–201 18 Lukasiewicz, T.: Probabilistic logic programming In: European Conference on Artificial Intelligence (1998) 388–392 19 Bacchus, F.: Representing and Reasoning with Probabilistic Knowledge MIT Press, Cambridge, Mass (1990) 20 Cozman, F.G.: Calculation of posterior bounds given convex sets of prior probability measures and likelihood functions Journal of Computational and Graphical Statistics 8(4) (1999) 824–838 21 Couso, I., Moral, S., Walley, P.: A survey of concepts of independence for imprecise probabilities Risk, Decision and Policy (2000) 165–181 22 de Campos, L., Moral, S.: Independence concepts for convex sets of probabilities In: Conference on Uncertainty in Artificial Intelligence, San Francisco, Morgan Kaufmann (1995) 108–115 23 Walley, P.: Statistical Reasoning with Imprecise Probabilities Chapman and Hall, London (1991) 24 Drenick, R.F.: Multilinear programming: Duality theories Journal of Optimization Theory and Applications 72(3) (1992) 459–486 25 Gochet, W., Smeers, Y.: A branch-and-bound method for reversed geometric programming Operations Research 27(5) (1979) 983–996 26 Horst, R., Tuy, H.: Global Optimization: Deterministic Approaches Springer Verlag, Heidelberg (1996) 27 Maranas, C.D., Floudas, C.A.: Global optimization in generalized geometric programming Computers and Chemical Engineering 21(4) (1997) 351–370 28 Sherali, H.D., Adams, W.P.: A Reformulation-Linearization Technique for Solving Discrete and Continuous Nonconvex Problems Kluwer Academic Publishers, Dordrecht/Boston/London (1999) 29 Tuy, H.: Convex Analysis and Global Optimization Volume 22 of Nonconvex Optimization and Its Applications Kluwer Academic Publishers, Dordrecht (1998) 30 de Campos, C.P., Cozman, F.G.: The inferential complexity of Bayesian and credal networks In: International Joint Conference on Artificial Intelligence (2005) 1313–1318 31 Pearl, J.: Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference Morgan Kaufmann, San Mateo, California (1988) Probabilistic Logic with Strong Independence 621 32 Dechter, R., Mateescu, R.: Mixtures of deterministic-probabilistic networks and their AND/OR search space In: Conference on Uncertainty in Artificial Intelligence, Arlington, Virginia, AUAI Press (2004) 120–129 33 Cowell, R., Dawid, P., Lauritzen, S., Spiegelhalter, D.: Probabilistic Networks and Experts Systems Springer-Verlag, New York (1999) 34 de Campos, C.P., Cozman, F.G.: Belief updating and learning in semi-qualitative probabilistic networks In: Conference on Uncertainty in Artificial Intelligence (2005) 153–160 35 Zhang, N.L., Poole, D.: A simple approach to bayesian network computations In: Canadian Conference on Artificial Intelligence (1994) 171–178 36 Dechter, R.: Bucket elimination: A unifying framework for probabilistic inference In Horvitz, E., Jensen, F., eds.: Conference on Uncertainty in Artificial Intelligence, San Francisco, Morgan Kaufmann Publishers (1996) 211–219 37 de Campos, C.P., Cozman, F.G.: Inference in credal networks using multilinear programming In: Second Starting AI Researcher Symposium, Valencia, IOS Press (2004) 50–61 38 Cozman, F.G., de Campos, C.P., Ide, J.S., da Rocha, J.C.F.: Propositional and relational Bayesian networks associated with imprecise and qualitative probabilistic assessments In: Conference on Uncertainty in Artificial Intelligence, Banff, AUAI Press (2004) 104–111 39 Cozman, F.G.: Credal networks Artificial Intelligence 120 (2000) 199–233 40 da Rocha, J.C.F., Cozman, F.G.: Inference in credal networks with branch-and-bound algorithms In: International Symposium on Imprecise Probability and Their Applications, Lugano, Switzerland, Carleton Scientific (2003) 41 da Rocha, J.C.F., Cozman, F.G., de Campos, C.P.: Inference in polytrees with sets of probabilities In: Conference on Uncertainty in Artificial Intelligence, New York, Morgan Kaufmann (2003) 217–224 42 Ide, J.S., Cozman, F.G.: Approximate inference in credal networks by variational mean field methods In: International Symposium on Imprecise Probability and Their Applications (2005) 203–212 43 Cozman, F.G., de Campos, C.P.: Local computation in credal networks In: Workshop on Local Computation for Logics and Uncertainty, Valencia, IOS Press (2004) 5–11 44 Ladkin, P.B., Maddux, R.: On binary constraint problems Journal of the ACM 41(3) (1994) 435–469 45 van Beek, P., Dechter, R.: On the minimality and global consistency of row-convex constraint networks Journal of the ACM 42(3) (1995) 543–561 46 Gettor, L., Friedman, N., Koller, D., Taskar, B.: Learning probabilistic models of relational structure In: International Conference on Machine Learning (2001) 170–177 47 Glesner, S., Koller, D.: Constructing flexible dynamic belief networks from first-order probabilistic knowledge bases Symbolic and Quantitative Approaches to Reasoning with Uncertainty (1995) 217–226 48 Haddawy, P., Suwandi, M.: Decision-theoretic refinement planning using inheritance abstraction In: AAAI Spring Symposium Series: Decision-Theoretic Planning (1994) 49 Jaeger, M.: Relational Bayesian networks In: Conference on Uncertainty in Artificial Intelligence, San Francisco, Morgan Kaufmann (1997) 266–273 50 Ngo, L., Haddawy, P.: Answering queries from context-sensitive probabilistic knowledge bases Theoretical Computer Science 171(1–2) (1997) 147–177 51 Poole, D.: Probabilistic Horn abduction and Bayesian networks Artificial Intelligence 64 (1993) 81–129 Bayesian Model Combination and Its Application to Cervical Cancer Detection Miriam Mart´ınez1 , Luis Enrique Sucar2 , Hector Gabriel Acosta3 , and Nicandro Cruz3 Tecnol´ ogico de Acapulco, Acapulco, Guerrero, M´exico miriamma ds@hotmail.com INAOE, Tonantzintla, Puebla, M´exico esucar@inaoep.mx Universidad Veracruzana, Xalapa, Veracruz, M´exico {heacosta, ncruz}@uv.mx Abstract We have developed a novel methodology to combine several models using a Bayesian approach The method selects the most relevant attributes from several models, and produces a Bayesian classifier which has a higher classification rate than any of them, and at the same time is very efficient Based on conditional information measures, the method eliminates irrelevant variables, and joins or eliminates dependent variables; until an optimal Bayesian classifier is obtained We have applied this method for diagnosis of precursor lesions of cervical cancer The temporal evolution of the color changes in a sequence of colposcopy images is analyzed, and the resulting curve is fit to an approximate model In previous work we develop different mathematical models to describe the temporal evolution of each image region, and based on each model to detect regions that could have cancer In this paper we combine the three models using our methodology and show very high accurracy for cancer detection, superior to any of the original models Introduction In this work we propose a method to combine several models using a Bayesian approach The method selects the most relevant attributes from several models, and produces a Bayesian classifier that has a high accuracy (improving all the original models), and at the same time is very efficient The method includes two phases, discretization and structural improvement Discretization is based on the minimum description length (MDL) principle, so the number of intervals that minimizes the MDL is obtained per attribute To deal with dependent and irrelevant attributes, we apply a structural improvement method that eliminates and/or joins attributes, based on mutual and conditional information measures We have applied this method for the analysis of colposcopy images for diagnosis of cervical cancer For this we used the parameters from three mathematical models that characterize the temporal evolution of each pixel in the image Each model has different number of parameters, in total 11 attributes, all continuous There are three classes Our method produces a very simple and efficient classifier with an accuracy of 95% J.S Sichman et al (Eds.): IBERAMIA-SBIA 2006, LNAI 4140, pp 622–631, 2006 c Springer-Verlag Berlin Heidelberg 2006 Bayesian Model Combination and Its Application 623 The rest of the document is organized as follows Section describes the Bayesian classifier and related work in structural improvement Section presents our method for model combination Section introduces the cervical cancer detection problem, and Section the image processing and modeling stages In the next section we present the experimental results We conclude with a summary and directions for future work Related Work There are several ways to combine several models for classification purposes One is to build several classifiers, one for each model, and then combine the outputs of each classifier by, for example, majority voting [15] Another is to build an ensemble or cascade of classifiers, as in Boosting [7] We propose a different approach We combine the parameters of all the models into a single classifier, and then use this for classification We use is a simple, naive Bayes classifier For building this classifier, we eliminate or join irrelevant and dependent attributes, through a structural improvement stage Next, we briefly review the naive Bayes classifier and related work in structural improvement A Bayesian classifier obtains the posterior probability of each class, Ci, using Bayes rule, as the product of the prior probability of the class by the conditional probability of the attributes given the class (likelihood), divided by the probability of the attributes The naive Bayes classifier (NBC) makes the simplifying assumption that the attributes are independent given the class, so the likelihood can be obtained by the product of the individual conditional probabilities of each attribute given the class In this way, the number of parameters increases linearly with the number of attributes, instead of exponentially Graphically, a NBC can be represented as star-structured Bayesian network [12], with a root node, C, that corresponds to the class variable, which is connected to the attributes, E1, , En The attributes are assumed conditionally independent given the class, so there are not arcs between them Thus, the posterior probability, P (Ci | E1, , En), of class Ci is given by: P (Ci | E1, , En) = P (Ci)P (E1 | Ci) P (En | Ci)/P (E) 2.1 (1) Structural Improvement The naive Bayes classifier assumes that the attribute are independent given the class If this is not true, there are two basic alternatives One is to transform the structure of the classifier to a Bayesian network, by introducing directed arcs between the dependent attributes There are several variants of this approach which are described in [8] The disadvantage is that the simplicity of the NBC is lost, so learning the model and then classifying new instances becomes more complex The other alternative is to transform the structure but maintaining a star or tree-structured network For this, [13] introduces basic operations: (i) eliminate an attribute, (ii) join two attributes into a new combined variable, 624 M Mart´ınez et al Fig Structural improvement: (a) original structure, (b) one attribute is eliminated, and (c) two attributes are joined into one variable (iii) introduce a new attribute that makes two dependent attributes independent (hidden node) Figure illustrates the first two operators, which are the ones we use in this work These operations are based on statistical tests to measure the correlation of pairs of attributes given the class variable Later, [11] proposes an alternative algorithm for variable elimination and merging (that correspond to the first two operators) The algorithm is based on two search procedures: (i) forward sequential selection and joining (FSSJ) and (ii) backward sequential elimination and joining (BSEJ) This procedures start from a full (empty) structure, and the they select attribute for elimination (addition) or for combination, testing the classification accuracy after each operation The advantage of these approaches is that they preserve the simplicity and efficiency of the NBC In contrast to previous work, we combine elimination of irrelevant attributes with elimination or combination of dependent attributes Also, our method is based in simple information measures, that avoid testing the classifier for each possible structure, and makes learning more efficient Finally, our method also integrates a discretization stage to deal with continuous attributes Model Combination The proposed method obtains an efficient Bayesian classifier for combining several models for classification It considers (i) discretization of continuous variables, (ii) selection of relevant attributes, and (iii) elimination or combination of dependent attributes The method obtains a naive Bayes classifier with the minimum number of attributes obtained form the different models The general algorithm is the following: (1) Initialization, (2) Discretization, (3) Structural Improvement, and (4) Evaluation Next we describe each stage in detail 3.1 Initialization This is step is done only once to build the initial classifier It considers all the attributes from each model (full structure) and an initial partition for the Bayesian Model Combination and Its Application 625 continuous attributes with two equal size intervals Based on this initial structure, the parameters are learned from training data 3.2 Discretization Given the current structure, in this stage the discretization for each continuous attribute is optimized The method is based on the minimum description length (MDL) principle The MDL measure makes a compromise between the accuracy and complexity of the discretization The measure we use, estimates the accuracy by measuring the mutual information between each attribute and the class; and the complexity by counting the number of parameters This is done for all the continuous attributes This stage is described in detail in [10] 3.3 Structural Improvement Given the current discretization (the one with the best MDL from the previous stage), in this phase the structure is improved to eliminate superfluous attributes and eliminate or combine dependent attributes This phase considers the following stages: The mutual information between each attribute and the class are obtained, and those attributes that not provide information (below a threshold) are eliminated The reminding attributes are tested using conditional mutual information (CMI) for each pair of attributes given the class If this value is high for a pair of attributes it is an indication that the attributes are not independent, so these are candidates for elimination or combination Each pair of attributes with high CMI are considered for elimination or combination One is deleted or both are combined into a single attribute The option with better classification rate is selected This process is repeated until there are no more superfluous or dependent attributes At this stage, the method can go back the discretization stage to optimize again the partitions for each continuous attribute, and then repeat structural improvement 3.4 Evaluation The final stage consists on evaluating the accuracy of the final classifier with test data Cervical Cancer Detection Cervical cancer is the first cause of death in Mexican women If it is detected early, the probability of cure is very high After a Pap smear test, colposcopy is the most common technique to diagnose this disease because, although it 626 M Mart´ınez et al Fig An example of a colposcopic image, where five regions (A, B, C, D, E) were selected from the image sequence (left) The corresponding Acetowhite response functions are plotted against time (the X-axis is time and the Y-axis is the relative intensity), including the corresponding adjusted curve using model (right) is more expensive, it has a higher sensitivity and specificity [5] Basically, the colposcopic test consists of the evaluation of the level of “white” color intensity that the cervical tissue reaches after acetic acid application An example of a colposcopic image is shown in figure There are two problems to develop a colposcopic test Firstly, the visual analysis has to be done by a well-trained gynecologist Secondly, the evaluation of the images is subjective in the sense that there are not standard criterion to correlate the tissue color (whitening) with the lesion degree [6] Many approaches have been proposed to automatically characterize cervical lesions from colposcopic images, most of them use color, texture, or shading, but none of these approaches have shown to be robust enough to be used as a diagnostic tool [14,9] More recently, some researchers have suggested to use the temporal patterns intrinsic to the color changes, but there is not a complete understanding of how to represent the dynamics of the whitening that occurs after acetic acid application, and how to use it to make an automatic diagnosis adviser [2,4,3] In previous work [1], we developed a technique to model the temporal evolution of the light changes in the tissue, and to parameterize the dynamics of these temporal changes using different mathematical models In the following section, this technique is explained, as well as the different mathematical models developed Then we present how the models are combined based on our methodology 5.1 Image Processing and Modeling Data Acquisition Images were acquired using a colposcope (dfv Vasconsellos model CP-M7, with magnification 16 X without any optical filter) and a color camera (Sony SSCDC50A), with a viewing distance of 20 cm During the first minute of image acquisition, 60 images (640x480) were taken as base line reference (1 frame/second) Bayesian Model Combination and Its Application 627 Then, after acetic acid application, five hundred and forty images were taken in minutes using the same sampling frequency In order to simplify image analysis, the images were processed in gray scale at a resolution of 74x99 5.2 Image Analysis The methodology used to analyze colposcopic images involves main stages: (i) image Registration, (ii) time series construction, and (iii) modeling The acquisition of colposcopic images spans in average minutes and even though that the patient is fixed, some small random movements are unavoidable, which are often local (patients breathing, movements due to the muscle tonus, etc.) To be able to analyze the sequence of images, i.e compare and evaluate corresponding structures, the objects in the images should be brought into the same position by removing the differences due to the patient movements - the time series of colposcopic image has to be registered In our case, we supposed the main source of the misalignments can be modeled by simple translation The analysis of preliminary registration results showed that this assumption was correct The method can transform the whole data using the same parameters, or can be local, depending on the local variations It can be based directly on the image intensity values (area-based methods) or can be done using some features computed from the images (feature-based methods) Because colposcopic images not contain many distinctive details, an area-based method was chosen The classical representative of area-based methods is the normalized crosscorrelation This method exploits matching directly image intensities The measure of similarity is computed for window pairs from the input and reference images and its maximum is searched [16] Because the tissue appearance changes over time due to the effect of the acetic acid effect, the searched pattern defined in the window can look different at a different times, so it is not appropriate to define a static reference image The input and the reference images are updated continuously starting with the first and second images of the sequence respectively, then the input and the reference images are redefined by the second and the third images and so on The starting points to initialize the search are updated by the last position in which the pattern window was found This registration strategy allows not only to contend with the fact that the searched pattern changes over time, but also to reduce the spatial space over which to develop the search After the registration process, the signal to noise ratio was increased using a spatial low pass filter implemented using a kernel window (3x3) The intensity value of each pixel over time was used to construct a time series, we called this, the Aceto-white response functions (AwRFs) On a similar way, each AwRF was smoothed using a polynomial filter with a polynomial of order and frames size 60 The filtered time series can be plotted against time (in the image sequence), some examples are shown in figure In previous work [1], three different mathematical models were used to model the evolution of each pixel and based on this to diagnose the probability of cancer 628 M Mart´ınez et al Model - Important Points This model considers important points in the curve: T s: the time to reach its maximum value T b: the maximum value T c: the time from the maximum value until the signal returns to its original value Model - Polynomial Polynomial of degree that approximates the signal, with parameters, P 1, P 2, P 3, P 4, P 5: AwRF (t) = P + P 2t + P 3/t + P 4/t2 + P 5/t3 (2) For < t < 1, where t(i) = samplei /samples The model is a function of time, where the start point of the dynamic function is at time = 60 seconds (the time at which the acetic acid was applied) So if 600 images were acquired on the image sequence, t(1) is equal to 0.1 (60/600) For analytical purposes the AwRFs can be divided in two sets: the upwards part and the downwards part The first part represents that the signal reaches the maximum level at some time, and the second one represents the dynamics of how the signal goes back from the maximum to the base line The peak time can be easily computed by solving AwRF (t) = for t In the same way, the maximum percentage of change reached can be known substituting for this t Model - Parabola Parabola approximation with parameters, P h, P k, P p After an analysis of the AwRFs we found that the first part of the signal can be approximated using a parabola The standard form of the equation of a parabola with vertex (h, k) and directrix y = k−p is expressed as (x−h)2 = 4p(y −k) with respect to the vertical axis Where the focus lies on the axis, p units from the vertex Under this representation the percentage of change and the time peak can be seen as the vertex (h, k) of the parabola, and p as an estimator of the speed of change An example of the parabolic approximation for different curves is depicted in figure Experimental Results We evaluated our methodology in the classification of different regions of colposcopy videos Firstly, some image sequences are classified with the help of an expert Secondly, based on the training cases, the time series are obtained and described using the models presented in the previous section Finally, the model combination methodology is applied and a Bayesian classifier is generated that combines the models This combined classifier was tested with other image sequences, different to the ones for training Bayesian Model Combination and Its Application 629 Fig Some examples of parabolic approximations to typical AwRFs Fig An example of a colposcopic image with the interface used by the expert to label the training cases 6.1 Knowledge Acquisition The knowledge extraction process was implemented asking the colposcopist to make an analysis of the image sequence Assisted by a graphical interface in which a representative image of the cervix was shown, the colposcopist was asked to define over the image the different regions with regards to three categories usually reported in a colposcopic description: typical transformation zone, low-grade lesion, and high-grade lesion The main idea is to ask the expert to make a colposcopic evaluation of one original image and to define regions with an associated label which correspond to a colposcopic feature This knowledge extraction procedure is made through the use of a graphical interface depicted in figure 6.2 Results We applied our methodology by considering the models, so all the parameters are the attributes for the classifier, 11 continuous attributes There are three classes: normal (T), low degree (A) and high degree (B) For the experiments 630 M Mart´ınez et al we used 1055 sample data labeled by an expert We used a holdout testing procedure, with approx 2/3 of the data for training and 1/3 for testing (800 for training and 255 for testing) After obtaining the best discretization for each attribute, the results of the structural improvement stage are summarized in table We can observe that the algorithm eliminates initially a number of redundant attributes, then it combines two dependent attributes and then it eliminates two more attributes, until it arrives at the final structure with attributes: T s − P k, T c, P 2, P 4, where one of them is the combination of two attributes of the original models This final classifier obtains a 95% accuracy In comparison, the results with the best single model were of aprox 90% accuracy, using the same data and testing procedure We also compared it with other classifiers using the same data (using all the attributes): (i) naive Bayes, 89%, (ii) tree augmented naive Bayes (TAN), 94%, and (iii) decision tree (C4.5), 94% The accuracy of our best model is clearly superior to the Naive Bayes and similar to TAN and C4.5 However, the model obtained is much simpler than TAN and C4.5, and thus more efficient Conclusions and Future Work We have developed a methodology to combine several models using a Bayesian approach The method selects the most relevant attributes from several models, and produces a Bayesian classifier that has a high accuracy and at the same time is very efficient Based on conditional information measures, the method eliminates irrelevant variables, and joins or eliminates dependent variables; until an optimal Bayesian classifier is obtained We have applied this method for diagnosis of precursor lesions of cervical cancer Based on the dynamics of a video sequence of colposcopy images, different mathematical models were generated to describe the Aceto-white response functions over time By combining these models with our approach, we generated a Bayesian classifier with attributes (one is a combination of of the original parameters) that produces a 95% accuracy for the test cases We are working on extending this method for dynamic models, based on dynamic Bayesian networks In the future we will like to apply the methodology Table Structural improvement stage for cancer diagnosis Operation No attr attributes 11 TsTbTcP1P2P3P4P5PhPkPp Join Ts-Pk 10 Ts-PkTbTcP1P2P3P4P5PhPp Elim P1 Ts-PkTbTcP2P3P4P5PhPp Elim Pp Ts-PkTbTcP2P3P4P5Ph Elim Ph Ts-PkTbTcP2P3P4P5 Elim Tc Ts-PkTbTcP2P4P5 Elim P3 Ts-PkTcP2P4P5 Elim P5 Ts-PkTcP2P4P5 Acc 88 89 89 90 90 91 92 95 Bayesian Model Combination and Its Application 631 in other domains, considering more models and parameters, to test the scalability of this approach References H Acosta-Mesa, B Zitov´ a, H R´ıos-Figueroa, N Cruz-Ram´ırez, A Mar´ınHern´ andez, R Hern´ andez-Jim´enez, B Cocotle-Ronz´ on, and E Hern´ andez-Galicia Cervical cancer detection using colposcopic images: a temporal approach In Proc Sixth International Conference on Computer Science (ENC’05), pages 158–164 IEEE, 2005 M Anderson, A M Jordan J., and F Sharp A Text and Atlas of Integrated Colposcopy Mosby, 1993 W Pogue B, A Zelenchuk, W Harper, G C Burke, E E Burke, and D M Harper Analysis of acetic acid-induced whitening of high-grade squamous intraepithelial lesions Journal of Biomedical Optics, pages 397–403, 2001 C Balas A novel optical imaging method for the early detection, quantitative grading, and mapping of cancerous and precancerous lesions of cervix IEEE Transactions on medical imaging, pages 96–104, 2001 E Burghardt, P Hellmuth, and F Girardi Primary care colposcopy Thieme, 2004 B L Craine, C J O’Toole, and Q Ji Digital imaging colposcopy: Corrected area measurements using shape-from-shading IEEE Transactions on medical imaging, pages 1003–1010, 1998 R Freund and E Schapire A short introduction to boosting Journal of the Japanese Society for Artificial Intelligence, 14:771–780, 1999 N Friedman, D Geiger, and M Goldszmidt Bayesian network classifiers Machine Learning, 29:131–163, 1997 Q Ji and Eric Craine Texture analysis for classification of cervix lesions IEEE Transactions on medical imaging, pages 1144–1149, 2000 10 M Mart´ınez and L.E Sucar Learning an optimal naive bayes classifier In Proc Congreso de Investigaci´ on y Extensi´ on ITESM, 2006 11 Michael J Pazzani Searching for atribute dependencies in bayesian classifiers In preliminary Papers of the Intelligence and Statistics, pages 424–429, 1996 12 J Pearl Probabilistic reasoning in intelligent systems Morgan Kaufmann, Palo Alto, Calif., U.S.A., 1988 13 L.E Sucar, D.F Gillies, and D.A Gillies Objective probabilities in expert systems Artificial Intelligence, 61:187–208, 1993 14 A T Vlastos, Andres Zuluaga, and Michele Follen New approaches to cervical cancer screening Contemporany Ob/Gyn, pages 87–103, 2002 15 X Yi, Z Kuo, and C Zhang Classifier combination based on active learning In Proc of the 17th International Conference on Pattern Recognition (ICPR’04), pages 184–187 IEEE, 2004 16 B Zitova Image registration methods: a survey Image and Vision Computing, pages 977–1000, 2003 Author Index Acosta, Hector Gabriel 622 Aguilera, Gabriel 602 Almeida, Vˆ ania 370 Anacleto, Junia 370 Anaya-S´ anchez, Henry 472 Aquino, Ronaldo R.B de 228 Asteasuain, Fernando 188 ´ Avila, Br´ aulio Coelho 289 Cruz, Nicandro 622 Cuautle, Rodrigo 542 Dantas Kaio A.A 108 Dantas, Maria Jos´e P 258 De Baets, Bernard 238 De Lima, Tiago 409 Delgado, Karina Valdivia Dignum, Frank Domingos, Pedro Drummond, Isabela 319 Duarte, Julio C 309 Bajo, Javier 58 Barros, Leliane Nunes de 7, 502 Bello, Rafael 238, 400 Berlanga-Llavori, Rafael 472 Bhatt, Mehul 430 Biajoli, Fabr´ıcio Lacerda 208 Borrajo, Daniel 128 Brena, Ramon F 522 Brito, Leonardo da C 258 Britos, Paola 128 Bruno, Odemir M 159 Burrieza, Alfredo 602 Cabalar, Pedro 419, 592 Campiteli, Mˆ onica G 159 Campo, Marcelo 390 Campos, Andr´e M.C 108 Campos, Cassio P de 612 Canuto, Anne M.P 108 Carballido, Jessica A 188 Carde˜ noso-Payo, Valent´ın 360 Carvalho, Aparecido de 370 Carvalho Jr., Manoel A 228 Carvalho, Paulo Henrique P de Castelfranchi, Cristiano 98 Castillo, Luis Fernando 58 Coelho, Andr´e L.V 118 Coelho, Helder 37 Colton, Simon 349 Corchado, Juan M 58 Cordero, Pablo 602 Costa, Anna Helena Reali 552 Costa, Vitor Santos 441 Cozman, F´ abio G 502, 612 Crasso, Marco 390 Enembreck, Fabr´ıcio 268, 289 Escolano, Francisco 149 Escudero-Mancebo, David 360 Espinosa, Jose 370 Far´ıas-Z´ arate, Cl´ audia J 198 Ferreira, Aida A 228 Ferreira, Anita 27 Ferreira, Francicleber Martins 582 Finger, Marcelo 482 Flach, Peter A Flores, Cecilia Dias 138 Freitas, Eduardo Noronha de Andrade 68 Freitas, Maria Claudia 309 Furtado, Vasco 118 Furuta, Alexandre H 380 258 Garat, Diego 492 Garcia, Maria M 238 Garc´ıa-Mart´ınez, Ram´ on 128 Garza Casta˜ n´ on, Luis E 168 Gluz, Jo˜ ao Carlos 138 Godoi, Muriel de S 370, 380 Gomes, Eduardo Rodrigues 17 Grau, Ricardo 400 Guerra-Hern´ andez, Alejandro 572 G´ uzman, Inmaculada P de 602 Heinen, Milton Roberto 562 Hernandez, Hector 249 634 Author Index Herrera, Myriam 299 Herzig, Andreas 409 Hirata, Celso Massaki 78 Os´ orio, Fernando Santos Osorio, Maria 542 Jaffe, Klaus 88 Jonsson, Anders 37 Kaestner, Celso A.A 339 Kepler, F´ abio N 482 Koerich, Alessandro L 268, 339 ´ Kuri-Morales, Angel 329 Le Lann, Marie-V´eronique 249 Lee, Huei Diana 278 Leite, Daniel Saraiva 462 Lieberman, Henry 370 Lima, Wagner da Silva 68 Lira, Milde M.S 228 Lopes, Carlos Roberto 512 Lorena, Luiz Antonio Nogueira 208, 218 Maceri, Pablo 128 Magee, Derek 349 Martinez, Alexandre S 159 Martinez, Emmanuel 522 Mart´ınez, Miriam 622 Martins, Ana Teresa 582 Mateos, Cristian 390 Mej´ıa-Guevara, Iv´ an 329 Melo, Adriano 118 Melo, Jony Teixeira de 512 Menezes, Talita 48 Milidi´ u, Ruy L 309 Monard, Maria Carolina 278 Montes de Oca, Sa´ ul 168 Mora-Bas´ an ˜ez, Carlos Rub´en de la 572 Morales-Men´endez, Rub´en 168 Morell, Carlos 238, 400 Moura, Eliane C.M 108 Mu˜ noz, Emilio 602 N´eris, Vˆ ania 370 Nolazco-Flores, Juan A 178 N´ obrega Neto, Otoni 228 Odakura, Valguima 552 Odintsov, Sergei 592 Oliveira, Alexandre C.M 218 562 Paes, Aline 441 Pearce, David 592 Pe˜ nalver, Antonio 149 Penteado, Raqueline R de M Pfeiffer, Carlos F 178 Pons-Porrata, Aurora 472 Ponzoni, Ignacio 188 Quental, Violeta 380 309 Rahayu, Wenny 430 Ramalho, Geber 48 Ram´ırez-Ruiz, Jos´e A 178 Reis, Danilo 118 Renter´ıa, Ra´ ul P 309 Revoredo, Kate 441 Ribeiro, Richardson 268 Rino, Lucia Helena Machado 462 Rocha, Jos´e Carlos F da 612 Rodr´ıguez, Yanet 238, 400 Ross, Peter 198 Rueda, Luis 299 S´ aez, Juan M 149 S´ anchez, Abraham 542 Sandri, Sandra 319 Santos, Araken M 108 Santos, C´ıcero N 309 Santos, Emanuel B 108 Santos, Paulo 349, 419 Seixas, Louise 138 Silla Jr., Carlos N 339 Silva, Fl´ avio S Corrˆea da 451 Silva, Geane B 228 Silva, Jos´e Reinaldo 532 Silva, S´ergio R.P Da 380 Silveira, Ricardo Azambuja 17 Soares, Rodrigo G 108 Steels, Luc 572 Sterling, Gerald 430 Sucar, Luis Enrique 622 Talarico, Am´erico 370 Tedesco, Patr´ıcia 48 Terashima-Mar´ın, Hugo 198, 522 Author Index Tonidandel, Flavio 532 Trevizan, Felipe W 502 Trigo, Paulo 37 Tsutsumi, Marie 370 Uruguay, Andr´e Luiz Pimentel Vazquez, Gustavo E Viccari, Rosa Maria 188 17, 138 Wang, DeLiang Wu, Feng Chung 278 78 Valenzuela-Rend´ on, Manuel 198 Vaquero, Tiago Stegun 532 Vasconcelos, Wamberto W 451 Zapata, Ren´e 542 Zaverucha, Gerson 441 Zem-Mascarenhas, Silvia Zunino, Alejandro 390 370 635 ... Sublibrary: SL – Artificial Intelligence ISSN ISBN-10 ISBN-13 030 2-9 743 3-5 4 0-4 546 2-4 Springer Berlin Heidelberg New York 97 8-3 -5 4 0-4 546 2-5 Springer Berlin Heidelberg New York This work is subject... http://www.cs.washington.edu/ai/alchemy J.S Sichman et al (Eds.): IBERAMIA-SBIA 2006, LNAI 4140, p 3, 2006 c Springer-Verlag Berlin Heidelberg 2006 Reinventing Machine Learning with ROC Analysis... Science+Business Media springer.com © Springer-Verlag Berlin Heidelberg 2006 Printed in Germany Typesetting: Camera-ready by author, data conversion by Scientific Publishing Services, Chennai, India Printed