Machine learning, optimization, and big data 2017

LNCS 10710 Giuseppe Nicosia · Panos Pardalos Giovanni Giuffrida · Renato Umeton (Eds.) Machine Learning, Optimization, and Big Data Third International Conference, MOD 2017 Volterra, Italy, September 14–17, 2017 Revised Selected Papers 123 Lecture Notes in Computer Science Commenced Publication in 1973 Founding and Former Series Editors: Gerhard Goos, Juris Hartmanis, and Jan van Leeuwen Editorial Board David Hutchison Lancaster University, Lancaster, UK Takeo Kanade Carnegie Mellon University, Pittsburgh, PA, USA Josef Kittler University of Surrey, Guildford, UK Jon M Kleinberg Cornell University, Ithaca, NY, USA Friedemann Mattern ETH Zurich, Zurich, Switzerland John C Mitchell Stanford University, Stanford, CA, USA Moni Naor Weizmann Institute of Science, Rehovot, Israel C Pandu Rangan Indian Institute of Technology, Madras, India Bernhard Steffen TU Dortmund University, Dortmund, Germany Demetri Terzopoulos University of California, Los Angeles, CA, USA Doug Tygar University of California, Berkeley, CA, USA Gerhard Weikum Max Planck Institute for Informatics, Saarbrücken, Germany 10710 More information about this series at http://www.springer.com/series/7409 Giuseppe Nicosia Panos Pardalos Giovanni Giuffrida Renato Umeton (Eds.) • • Machine Learning, Optimization, and Big Data Third International Conference, MOD 2017 Volterra, Italy, September 14–17, 2017 Revised Selected Papers 123 Editors Giuseppe Nicosia University of Catania Catania Italy Giovanni Giuffrida University of Catania Catania Italy Panos Pardalos University of Florida Gainesville, FL USA Renato Umeton Harvard University Cambridge, MA USA ISSN 0302-9743 ISSN 1611-3349 (electronic) Lecture Notes in Computer Science ISBN 978-3-319-72925-1 ISBN 978-3-319-72926-8 (eBook) https://doi.org/10.1007/978-3-319-72926-8 Library of Congress Control Number: 2017962876 LNCS Sublibrary: SL3 – Information Systems and Applications, incl Internet/Web, and HCI © Springer International Publishing AG 2018 This work is subject to copyright All rights are reserved by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed The use of general descriptive names, registered names, trademarks, service marks, etc in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication Neither the publisher nor the authors or the editors give a warranty, express or implied, with respect to the material contained herein or for any errors or omissions that may have been made The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations Printed on acid-free paper This Springer imprint is published by Springer Nature The registered company is Springer International Publishing AG The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland Preface MOD is an international conference embracing the fields of machine learning, optimization, and data science The third edition, MOD 2017, was organized during September 14–17, 2017 in Volterra (Pisa, Italy), a stunning medieval town dominating the picturesque countryside of Tuscany The key role of machine learning, reinforcement learning, artificial intelligence, large-scale optimization, and big data for developing solutions to some of the greatest challenges we are facing is undeniable MOD 2017 attracted leading experts from the academic world and industry with the aim of strengthening the connection between these institutions The 2017 edition of MOD represented a great opportunity for professors, scientists, industry experts, and postgraduate students to learn about recent developments in their own research areas and to learn about research in contiguous research areas, with the aim of creating an environment to share ideas and trigger new collaborations As chairs, it was an honor to organize a premiere conference in these areas and to have received a large variety of innovative and original scientific contributions During this edition, six plenary lectures were presented: Yi-Ke Guo, Department of Computing, Faculty of Engineering, Imperial College London, UK Founding Director of Data Science Institute Panos Pardalos, Department of Systems Engineering, University of Florida, USA Director of the Center for Applied Optimization Ruslan Salakhutdinov, Machine Learning Department, School of Computer Science at Carnegie Mellon University, USA Director of AI Research at Apple My Thai, Department of Computer and Information Science and Engineering, University of Florida, USA Jun Pei, Hefei University of Technology, China Vincenzo Sciacca, Cloud and Cognitive Division – IBM Rome, Italy There were also two tutorial speakers: Domenico Talia, Dipartimento di Ingegneria Informatica, Modellistica, Elettronica e Sistemistica Università della Calabria, Italy Xin–She Yang, School of Science and Technology – Middlesex University London, UK Moreover, the conference hosted the second edition of the industrial session on “Machine Learning, Optimization and Data Science for Real-World Applications”: Luca Maria Aiello, Nokia Bell Labs, UK Pierpaolo Basile, University of Bari, Italy VI Preface Carlos Castillo, Universitat Pompeu Fabra in Barcelona, Spain Moderator: Aris Anagnostopoulos, Sapienza University of Rome, Italy We received 126 submissions from 46 countries and five continents; each manuscript was independently reviewed by a committee formed by at least five members through a blind review process These proceedings contain 49 research articles written by leading scientists in the fields of machine learning, artificial intelligence, reinforcement learning, computational optimization, and data science presenting a substantial array of ideas, technologies, algorithms, methods, and applications For MOD 2017, Springer generously sponsored the MOD Best Paper Award This year, the paper by Khaled Sayed, Cheryl Telmer, Adam Butchy, and Natasa Miskov-Zivanov titled “Recipes for Translating Big Data Machine Reading to Executable Cellular Signaling Models” received the MOD Best Paper Award This conference could not have been organized without the contributions of these researchers, and so we thank them all for participating A sincere thank you also goes to all the Program Committee, formed by more than 300 scientists from academia and industry, for their valuable work of selecting the scientific contributions Finally, we would like to express our appreciation to the keynote speakers, tutorial speakers, and the industrial panel who accepted our invitation, and to all the authors who submitted their research papers to MOD 2017 September 2017 Giuseppe Nicosia Panos Pardalos Giovanni Giuffrida Renato Umeton Organization General Chair Renato Umeton Harvard University, USA Conference and Technical Program Committee Co-chairs Giuseppe Nicosia Panos Pardalos Giovanni Giuffrida University of Catania, Italy and University of Reading, UK University of Florida, USA University of Catania, Italy Tutorial Chair Giuseppe Narzisi New York University Tandon School of Engineering, USA Industrial Session Chairs Ilaria Bordino Marco Firrincieli Fabio Fumarola Francesco Gullo UniCredit UniCredit UniCredit UniCredit R&D, R&D, R&D, R&D, Italy Italy Italy Italy Organizing Committee Piero Conca Jole Costanza Giorgio Jansen Giuseppe Narzisi Andrea Patane’ Andrea Santoro Renato Umeton CNR, Italy Italian Institute of Technology, Milan, Italy University of Catania, Italy New York University Tandon School of Engineering, USA University of Oxford, UK Queen Mary University London, UK Harvard University, USA Technical Program Committee Agostinho Agra Kerem Akartunali Richard Allmendinger Aris Anagnostopoulos Davide Anguita Universidade de Aveiro, Portugal University of Strathclyde, UK The University of Manchester, UK Università di Roma La Sapienza, Italy University of Genoa, Italy VIII Organization Takaya Arita Jason Atkin Chloe-Agathe Azencott Jaume Bacardit James Bailey Baski Balasundaram Elena Baralis Xabier E Barandiaran Cristobal Barba-Gonzalez Helio J C Barbosa Roberto Battiti Lucia Beccai Aurelien Bellet Gerardo Beni Khaled Benkrid Peter Bentley Katie Bentley Heder Bernardino Daniel Berrar Adam Berry Luc Berthouze Martin Berzins Mauro Birattari Leonidas Bleris Christian Blum Paul Bourgine Anthony Brabazon Paulo Branco Juergen Branke Larry Bull Tadeusz Burczynski Robert Busa-Fekete Sergiy I Butenko Stefano Cagnoni Yizhi Cai Guido Caldarelli Alexandre Campo Angelo Cangelosi Salvador Eugenio Caoili Timoteo Carletti Jonathan Carlson Celso Carneiro Ribeiro Michelangelo Ceci Adelaide Cerveira Uday Chakraborty Nagoya University, Japan The University of Nottingham, UK Institut Curie Research Centre, Paris, France Newcastle University, UK University of Melbourne, Australia Oklahoma State University, USA Politecnico di Torino, Italy University of the Basque Country, Spain University of Malaga, Spain Laboratório Nacional de Computacao Cientifica, Brazil University of Trento, Italy Istituto Italiano di Tecnologia, Italy Inria Lille, France University of California at Riverside, USA The University of Edinburgh, UK University College London, UK Harvard Medical School, USA Universidade Federal de Juiz de Fora, Brazil Tokyo Institute of Technology, Japan CSIRO, Australia University of Sussex, UK SCI Institute, University of Utah, USA IRIDIA, Université Libre de Bruxelles, Belgium University of Texas at Dallas, USA Spanish National Research Council, Spain École Polytechnique Paris, France University College Dublin, Ireland Instituto Superior Tecnico, Portugal University of Warwick, UK University of the West of England, UK Polish Academy of Sciences, Poland Yahoo! Research, NY, USA Texas A&M University, USA University of Parma, Italy University of Edinburgh, UK IMT Lucca, Italy Université Libre de Bruxelles, Belgium University of Plymouth, UK University of the Philippines Manila, Philippines University of Namur, Belgium Microsoft Research, USA Universidade Federal Fluminense, Brazil University of Bari, Italy Universidade de Tras-os-Montes e Alto Douro, Portugal University of Missouri – St Louis, USA Organization Xu Chang W Art Chaovalitwongse Antonio Chella Ying-Ping Chen Haifeng Chen Keke Chen Gregory Chirikjian Silvia Chiusano Miroslav Chlebik Sung-Bae Cho Yonsei Anders Christensen Dominique Chu Philippe Codognet Carlos Coello Coello George Coghill Pietro Colombo David Cornforth Luís Correia Chiara Damiani Thomas Dandekar Ivan Luciano Danesi Christian Darabos Kalyanmoy Deb Nicoletta Del Buono Jordi Delgado Ralf Der Clarisse Dhaenens Barbara Di Camillo Gianni Di Caro Luigi Di Caro Luca Di Gaspero Peter Dittrich Federico Divina Stephan Doerfel Devdatt Dubhashi George Dulikravich Juan J Durillo Omer Dushek Marc Ebner Pascale Ehrenfreund Gusz Eiben Aniko Ekart Talbi El-Ghazali Michael Elberfeld Michael T M Emmerich Andries Engelbrecht IX University of Sydney, Australia University of Washington, USA Università di Palermo, Italy National Chiao Tung University, Taiwan NEC Labs, USA Wright State University, USA Johns Hopkins University, USA Politecnico di Torino, Italy University of Sussex, UK University, South Korea Lisbon University Institute, Portugal University of Kent, UK University Pierre and Marie Curie – Paris 6, France CINVESTAV-IPN, Mexico University of Aberdeen, UK University of Insubria, Italy University of Newcastle, UK University of Lisbon, Portugal University of Milan-Bicocca, Italy University of Würzburg, Germany Unicredit Bank, Italy Dartmouth College, USA Michigan State University, USA University of Bari, Italy Universitat Politecnica de Catalunya, Spain MPG, Germany Université Lille, France University of Padua, Italy IDSIA, Switzerland University of Turin, Italy University of Udine, Italy Friedrich Schiller University of Jena, Germany Pablo de Olavide University of Seville, Spain Kassel University, Germany Chalmers University, Sweden Florida International University, USA University of Innsbruck, Austria University of Oxford, UK Ernst-Moritz-Arndt-Universität Greifswald, Germany The George Washington University, USA VU Amsterdam, The Netherlands Aston University, UK University of Lille, France RWTH Aachen University, Germany Leiden University, The Netherlands University of Pretoria, South Africa Subject Recognition Using Wrist-Worn Triaxial Accelerometer Data 585 Fernandez-Lopez, P., Liu-Jimenez, J., Sanchez-Redondo, C., Sanchez-Reillo, R.: Gait recognition using smartphone In: 2016 IEEE International Carnahan Conference on Security Technology (ICCST), pp 1–7 IEEE (2016) Fulcher, B.D., Jones, N.S.: Highly comparative feature-based time-series classification IEEE Trans Knowl Data Eng 26(12), 3026–3037 (2014) George, S.L.: Research misconduct and data fraud in clinical trials: prevalence and causal factors Int J Clin Oncol 21(1), 15–21 (2016) Iglewicz, B., Hoaglin, D.C.: How to Detect and Handle Outliers, vol 16 ASQ Press, Milwaukee (1993) Kelly, L.A., McMillan, D.G., Anderson, A., Fippinger, M., Fillerup, G., Rider, J.: Validity of actigraphs uniaxial and triaxial accelerometers for assessment of physical activity in adults in laboratory conditions BMC Med Phys 13(1), (2013) Khan, S.S., Madden, M.G.: A survey of recent trends in one class classification In: Coyle, L., Freyne, J (eds.) AICS 2009 LNCS (LNAI), vol 6206, pp 188–197 Springer, Heidelberg (2010) https://doi.org/10.1007/978-3-642-17080-5 21 Liu, F.T., Ting, K.M., Zhou, Z.H.: Isolation forest In: Eighth IEEE International Conference on Data Mining, ICDM 2008, pp 413–422 IEEE (2008) 10 Loughran, R., Agapitos, A., Kattan, A., Brabazon, A., O’Neill, M.: Speaker verification on unbalanced data with genetic programming In: Squillero, G., Burelli, P (eds.) EvoApplications 2016 LNCS, vol 9597, pp 737–753 Springer, Cham (2016) https://doi.org/10.1007/978-3-319-31204-0 47 11 Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., Duchesnay, E.: Scikit-learn: machine learning in Python J Mach Learn Res 12, 2825–2830 (2011) 12 Rajna, P., Szomszed, A.: Actigraphy: a valuable diagnostic tool or a luxury investigation? (Neuropsychiatric aspects) Ideggyogy Sz 62(9–10), 308–316 (2009) 13 Teskey, W.J., Elhabiby, M., El-Sheimy, N.: Inertial sensing to determine movement disorder motion present before and after treatment Sensors 12(3), 3512–3527 (2012) 14 Trost, S.G., McIver, K.L., Pate, R.R.: Conducting accelerometer-based activity assessments in field-based research Med Sci Sports Exerc 37(11), S531 (2005) 15 Wang, L., Ning, H., Tan, T., Hu, W.: Fusion of static and dynamic body biometrics for gait recognition IEEE Trans Circuits Syst Video Technol 14(2), 149–158 (2004) 16 Zhao, W., Chellappa, R., Phillips, P.J., Rosenfeld, A.: Face recognition: a literature survey ACM Comput Surv (CSUR) 35(4), 399–458 (2003) Detection of Age-Related Changes in Networks of B Cells by Multivariate Time-Series Analysis Alberto Castellini(B) and Giuditta Franco Department of Computer Science, Verona University, Strada Le Grazie 15, 37134 Verona, Italy {alberto.castellini,giuditta.franco}@univr.it Abstract Immunosenescence concerns the gradual deterioration of the immune system due to aging Recent advances in cellular phenotyping have enabled key improvements in this context during the last decades In this work we present a novel extensions and integration of data-driven models for describing age-related changes in the network of relationships among cell quantities of eight peripheral B lymphocyte subpopulations Our dataset contains about six thousands samples of patients having an age between one day and ninety-six years, where for each patient, cell quantities of eight peripheral B lymphocyte subpopulations were measured By correlation-based multiple time series segmentation we generate four sets of age-related networks depending on the number of age segments We first analyze a partition in 30 very short segments, then segmentations in 5, and segments Moving from a fine to a large grain segmentation, different aspects of the dataset are highlighted and analyzed Introduction In last decades main mechanisms of the human immune system, one of the most complex and adaptive systems known in nature, were investigated from different perspectives In this context, aging gathered much attention as a complex process which negatively impacts this system [11,27] In fact, it may be assumed that the immune system, as a network of interacting cells, evolves during human life in terms of different aspects Presence/absence, strength and types of interactions, for instance, can change during the life of a person due to the exposure to multiple foreign challenges through childhood, via young and mature adulthood, to the decline of old age [23] The immune network theory, formulated by Jerne [15] in 1974 and subsequently developed by Parelson [22], was a starting point to describe the dynamics of lymphocyte interactions from a quantitative and systemic point of view [21] More recently, advances in cellular phenotyping have enabled to elucidate several functioning mechanisms of the immune system, such as those underling immunosenescence [7,13], having a notable social and economical impact in the design of new therapies and vaccines Latest studies showed that the majority of c Springer International Publishing AG 2018 G Nicosia et al (Eds.): MOD 2017, LNCS 10710, pp 586–597, 2018 https://doi.org/10.1007/978-3-319-72926-8_49 Detection of Age-Related Changes in Networks of B Cells 587 lymphocyte biological variability seems to be age-dependent [26] and immunosenescence seems to be characterized by a decrease in cell-mediated immune functions, where defects in T- and B-cell functions coexist [27] In this work we proposes a pairwise correlation based analysis over a dataset of B-cell quantities, measured in about six thousands patients having an age between one day and ninety-six years For each patient, cell quantities of eight peripheral B lymphocyte subpopulations were measured and related time-series, obtained by ordering patients according to their age, analyzed Pearson correlation was computed to find out age-related relationships between different B-cell subpopulations This kind of analysis allowed us to visualize our data as (correlation based) networks over the available types of B-cells Differences in such age-related networks were investigated in terms of a time-series segmentation problem, where a partition (in intervals) of the time line (i.e., patient age) emerges as an optimal one to maximize a measure of dissimilarity between corresponding B-cell networks (see Fig 1) Fig Age-related changes in immune system network Segmentation of multiple time-series is a complex problem, since different data partitioning may show different aspects of underlying processes and these aspects could have non-synchronous evidence Main methodologies in the field [16,24] take inspiration and extend methods of motif discovery in timeseries [6,19] or are based on clustering [8,25] The choice of the information measure of course has a strong influence on the identification of segments (i.e., time intervals/clusters of time points) and change points A possible measure is represented by the parameters of the (multivariate) mathematical models fitting the data in each segment, since they represent some aspects of the information in the segment itself Comparing these parameters between couples of adjacent segments and maximizing their differences is a way to identify good segmentations If linear regression models are used, then predictor coefficients are compared In [4] a constant (to all age intervals) network was provided by 588 A Castellini and G Franco setting general assumptions, while an age-dependent one was found by restricting statistical thresholds to validate our multivariate linear models A previous model proposed for this dataset [2] describes a possible sequence of (ex-vivo observed) B cell maturation steps in human body It was based on Metabolic P systems [9,10,20], with linear regulation maps generated by regression techniques and genetic algorithms [3,5] The main contribution of this work concerns the generation and analysis of four sets of age-related networks depending on the number of age segments We first analyze a partition in 30 very short segments, then we analyze a segmentation in 5, and segments, where the age-intervals of corresponding segments increase while the number of segments decrease Moving from a fine to a large grain segmentation, different aspects of the dataset are highlighted We used a brute-force algorithm for testing every possible segmentation with specific number of segments, and evaluated these segmentations according to a segmentation performance measure based on the average correlation difference between segments A preliminary comparison between these networks and those computed in [4] is provided The rest of the paper is organized as follows The dataset and algorithm are described in Sect A discussion on initial results is presented in Sect and Sect reports some conclusions and proposals for future work Material and Methods This section describes the immunological dataset and the segmentation method used to generate age-related networks of B cell subpopulations 2.1 Dataset Data were collected at the University Hospital of Verona (Italy) from 2001 to 2012 as measures of amount of B cells exhibiting the combinations of receptor clusters CD27, CD23 and CD5 in 5,954 patients There were 2,910 males and 3,045 females (male/female ratio: 0.95) and the median age of the patients was 37 years (range: 0–95 years) More details on the dataset and the clinical method used to collect it may be found in [2,4,26] The names of population size variables corresponding to each cell phenotype are displayed in Table In other terms, B cell phenotype of subpopulations (indicated by presence and absence of three receptor clusters), may be abstractly described by random variables accounting for quantities of corresponding cell in each patient The dataset is a matrix of 5,954 rows and eight columns, in which rows (i.e., patients) can be sorted by age, obtaining a kind of multivariate time-series where patient age represents time Given the definition of our problem (see previous section), this is a particular case where it is reasonable to reduce cross-sectional data into multivariate time-series In fact, if we sort the data according to the age of patients, we have a screenshot of the human immune system (or, more specifically, of the B-cell network) along the lifetime of a metapatient, who may Detection of Age-Related Changes in Networks of B Cells 589 Table Dataset variables X1 = CD5+ CD23+ CD27- X5 = CD5− CD23− CD27+ X2 = CD5− CD23+ CD27− X6 = CD5+ CD23− CD27+ X3 = CD5− CD23− CD27− X7 = CD5+ CD23+ CD27+ X4 = CD5+ CD23− CD27− X8 = CD5− CD23+ CD27+ be assumed to have a basic functioning system (the number of patients is high enough, to be able to neglect possible known or unknown diseases, defeacts on the system) 2.2 Algorithm The segmentation algorithm here used is a brute-force partitioning method, based on the maximization of differences in correlation between adjacent segments The multiple time-series of patients (sorted by age) is initially split in 30 primitive segments, each containing 200 patients who result in being of very similar age We assume that the time-series is stationary in each primitive segment because the age effect is irrelevant in such small intervals Time-series segmentations are generated by partitioning the vector of primitive segments (1, , 30) in n segments s1 , , sn , such that ≤ n ≤ 30 Segment si = (ki , ki+1 ), ≤ ki < ki+1 ≤ 30, i = 1, , n contains all patients in primitive segments between ki and ki+1 For instance, segment (2, 4) contains all primitive segments from to (namely, patients from 200 to 799 in the age-sorted list of patients) A segmentation having n segments is then represented by an n-uple S n = (k1 , , kn ), where elements ≤ k1 < < kn = 30 are the indexes of the primitive segments that delimit the end of each segment in the segmentation For instance, segmentation (3, 7, 20, 30) contains four segments, namely (1, 3), (4, 7), (8, 20) and (21, 30), hence patients are segmented as (1, 599), (600, 1399), (1400, 1999) and (2000, 5954) Given a segment si , we indicate by ci = (ci,1 , , ci,28 ) the vector containing the correlation of each couple of different cells in the age interval of segment si The number of elements in ci is 28 because there are 28 possible pairs of different cell types in our dataset We compute the vector of absolute differences of correlation between two adjacent segments si and si+1 as di,i+1 = (d1i,i+1 , , d28 i,i+1 ) = |ci+1 − ci | Let us analyze an example from Fig which shows the (unique) segmentation in 30 primitive segments The first element of c1 is the correlation between X1 and X7 in the first age segment (i.e., 0.0–1.0 years), which has a value of 0.49 (see first column of matrix (a)) The first element of c2 is the correlation between X1 and X7 in the second age segment (i.e., 1.0–2.5 years), which has a value of 0.76 (see second column of matrix (a)) The absolute difference of these two values is the value of the first element of vector d1,2 , namely 0.27 (see first column of matrix (b)) 590 A Castellini and G Franco Given a specific number of segments n, our goal is to identify the segmentation Sˆn = (kˆ1 , , kˆn ) which maximizes the overall absolute differences of correlation between adjacent segments To this end we define the performance measure of a generic segmentation Sn with n segments as: m(Sn ) = n−1 i=1 ( 28 j j=1 (di,i+1 (Sn ))) 28 · (n − 1) (1) which represents the average difference in correlation between all adjacent segments in Sn and depends on the specific segmentation points k1 , , kn in which the multiple time-series is partitioned The algorithm described in Table aims at identifying the segmentation Sn that maximizes m(Sn ) Table Brute-force algorithm for generating of the best segmentation with n segments according to the segmentation performance measure in Eq (1) Best segmentation(n) Input: n: number of segments # Initialization bestSegmentation=NULL; bestPerformance=-Inf ; # Search for best segmentation for each segmentation Snh = (k1 , , kn ) | ≤ k1 < < kn = 30 { compute segmentation performance p = m(Snh ); if(p>bestPerformance) { bestSegmentation=Snh ; bestPerformance=p; 10 } 11 } 12 return (bestSegmentation,bestPerformance); The advantage of using this algorithm is that all possible partitioning of the 30 primitive segments in n parts is tested On the other hand, the number of these partitions is the binomial coefficient 29 n which grows very quickly as n increases or decreases to n/2 Table shows the number of possible segmentations depending on n and the time needed by the brute-force algorithm to find the best segmentation The algorithm was implemented in R language and it run on a laptop with processor Intel R QuadCoreTM i7-3537U 2.00 GHz and GB of RAM Detection of Age-Related Changes in Networks of B Cells 591 Table Algorithm performance depending on the number of segments n n # segmentations Time 29 0.15 s 406 1.95 s 3654 16.83 s 23751 1.96 118755 11.52 475020 30 1.15 h 0.01 s Results We started to analyze the basic segmentation in 30 intervals (maximal granularity) having 200 patients each Correlation analysis on this case are reported in Fig 2, where we may see: the pairwise correlation 28 × 30 matrix (a), the absolute value of differences of correlations between two consecutive segments reported in a 28 × 29 matrix (b), and the sorted row-wise average vector (c) Two examples of correlation values (computed on the 30 segments) are graphed in (e): those between X1 and X7 , and between X2 and X5 , where the dotted horizontal line denotes the 0.5 threshold Only absolute values are considered, since we are interested in selecting couple of relatively high correlated variables (rather than in the verse they are correlated) Matrix in (a) is filtered across the threshold value 0.5, so to obtain the bicolor matrix in (d), which corresponds to a sequence of 30 networks over the variables Four of these networks, corresponding to the segments 1, 2, 12, 29, are reported in (f), where we may notice a different presence of edges (i.e., different couples of highly correlated variables) for different age intervals, and a decrease of presence of edges with the age increasing (segment 12 corresponds to the range 25–29 years old and segment 29 to the range 73–78 years, see Fig 2) The 30-partition has the advantage to group patients having very similar age, then detecting actual relationships between variables which cannot be seen macroscopically from the time-series, which are due to data-driven age-independent structural properties of the B-cell network However, we aim at finding a partition with a minor number of segments, having better performance in terms of dissimilarity between corresponding networks, but still keeping the property to exhibit an age-independent network which fits actual relationships among variables over the intervals We run the brute force segmentation algorithm, to analyze performances of first n-partitions with n = 2, , 7, which results are reported in Figs and Performance measure of all possible partitions with 2,3, ,7 segments was computed, and best values are respectively reported in Fig 3(a), where the maximum (m = 0.31) is obtained for a tripartition It generates one B cell network for patients up to one year old, another one for the interval 1–34.8 years, and the last 592 A Castellini and G Franco Fig Initial partition of 30 intervals, each with 200 patients Correlation matrix (a), correlation differences between consecutive segments in absolute value (b), corresponding sorted row average vector (c) having components average 0.129 Pairwise correlations greater than 0.5 (in absolute value) in (d), reported in (f) as networks for a specific sample of segments: 1, 2, 12, 29 Absolute values of pairwise correlations of (X1 , X7 ) and (X2 , X5 ) (e) one for ages over 34.8 One only correlation is kept for the whole life, between X1 and X4 , two phenotypes which differ only for the expression of CD23, as it was also the case for linear models proposed in [4] After one year of age, B cell network has many more edges (that is, it has new variable couples highly correlated), and there are several edges which appear only in the range of age [1–34.8] as it is evident by observing the central column of matrix in (c), which is almost all dark (high correlation) with row neighbours white (scarse correlation) However the network of the second segment is not an enrichment of the previous one, because a couple of correlations, between X4 , X6 and between X2 , X8 , get lost (as they decrease under 0.5 in absolute value) Both Detection of Age-Related Changes in Networks of B Cells 593 Fig Tripartition corresponding to maximal segmentation performance (m = 0.31) Best values for segmentation performance m over partitions with 2, 3, 4, 5, 6, segments in (a), corresponding m values and ages delimiting the segments in (b) On the tripartition with m maximum: pairwise correlations (c) generating networks on the bottom with a filter of 0.5, correlations differences in abs value (d), and corresponding sorted row-wise average (c) these edges connect two cell phenotypes which differ each other for only one receptor expression: CD27 This confirms observations by linear models in [4], where however relationship between X2 , X8 was present until higher age (23 years) 594 A Castellini and G Franco We notice that tripartition is also a natural segmentation we observe if we look at the multivariate time-series (see Fig in [2]), having all a peak around one year, and a descendent tail after about 30 years We presume this granularity keeps track of a sort of macro age-dependent correlation in the three segments, and should be further investigated in its stationarity component Namely, ARIMA models could be considered to eliminate the non-stationarity [14] A good point of the tripartition is to find a couple of change points which are maintained, among the others, in partitions with more segments (see table (b) in Fig 3): 1.0 and 34.8 years The decline of the immune system however happens inside the interval 34.8–96.0, which in the following we analyze within a finer grain Due to these observations, we pass to analyze our data at a major granularity, related to the second best values for m, which are 0.27 and 0.28, corresponding to partitions with and segments respectively, visualized in Fig Fig Bipartition (top) and 5-partition (bottom), corresponding to second best segmentation performance (0.27 and 0.28 respectively) Correlation matrix (a), correlation differences between consecutive segments in absolute value (b), corresponding sorted row average vector (c) and networks with absolute value of pairwise correlations over 0.5 (d) Detection of Age-Related Changes in Networks of B Cells 595 We may notice that the bipartition has an excessively low granularity, since the network corresponding to the second interval of ages (starting with 43.4 years) has no edges - that is, pairwise correlations have all an absolute value smaller than 0.5 On the other hand, the partition with segments turns out to be quite informative and interesting, by finding new stable change points at 69 and 73.2 (see Table (b) of Fig 3) Other less significant change points are found by next partitions in and segments, where respectively 78.2 was first added, and then 65.4 Hence, according to this model, the B cell network alterations, observed as the decline of defence in eldery people, should be investigated around specific age ranges, at 69 and 73 (and of course 78) years In bottom of Fig 4, a good performant 5-partition (with performance measure 0.28) is described in terms of age-related B cell networks (d) It is the model we propose in this paper, with same two networks than in the tripartition until 34.8 years, and three last networks suggesting the dynamics of mature adulthood, to the decline of old age In our data driven correlation networks, B cell networks change dramatically at 69 years with an increasing of all node degrees (correlations of all variables) Namely, in this age range, we notice an irreversible lost of correlation between X1 −X5 (cell phenotypes with all three opposite receptors) and an irreversible recover of both connections X4 −X6 and X1 −X8 , while pairwise correlation between X3 −X7 (cell phenotyes having all or none of the receptors expressed) is the only one which keeps a value greater than 0.5 until the end A biomedical validation of this model will be the next step for future work, in order to test and eventually improve it Conclusion and Ongoing Work A recent broad interest is focused on the lifetime aging of immune system, in terms of changes of immune mechanisms of an individual during his/her infancy, growing/mature age and senescence In particular, efficient and fast computational methods are proposed in the machine learning literature (for data clustering, and feature extraction) to infer new knowledge from given data In this context we are currently considering different types of methodologies for multiple time-series analysis, such as, segmentation [16], change-point detection [1,17,18] Namely, in this paper a simple algorithm allowed us to preliminary analyze some statistically validated partitions of ages where the B cell networks of immune system have change points of interest In our correlation data driven model, years 69 and 73 seems to be critical for the decline studied in immunosenescence The model proposed in this paper may be naturally extended (by improving the partition algorithm, the selected thresholds, by investigating the intermediate zone, with 8–28 intervals, by heuristics from the literature) and improved, with more sophisticated statistical analysis of our specific dataset For instance we are currently considering a recent approach, proposed in [12], where subsequence clustering of multivariate time-series is profitably used for discovering repeated patterns in temporal data Once these patterns have been discovered, the initial dataset can be interpreted as a temporal sequence of only a small 596 A Castellini and G Franco number of states (namely clusters or segments) Patterns are defined by Markov Random Field (MRF) characterizing the interactions between different variables in typical subsequences of specific clusters Based on this graphical representation, a simultaneous segmentation of time-series data may be efficiently realized Acknowledgments Authors would like to thank Antonio Vella (department of pathology and diagnostics, University Hospital of Verona) for providing the dataset used in this work and for interesting discussions on the role of B cells in the immune system References Barnett, I., Onnela, J.-P.: Change point detection in correlation networks Sci Rep 6(18893), 1–11 (2016) Castellini, A., Franco, G., Manca, V., Ortolani, R., Vella, A.: Towards an MP model for B lymphocytes maturation In: Ibarra, O.H., Kari, L., Kopecki, S (eds.) UCNC 2014 LNCS, vol 8553, pp 80–92 Springer, Cham (2014) https://doi.org/ 10.1007/978-3-319-08123-6 Castellini, A., Franco, G., Pagliarini, R.: Data analysis pipeline from laboratory to MP models Nat Comput 10(1), 55–76 (2011) Castellini, A., Franco, G., Vella, A.: Age-related relationships among peripheral B lymphocyte subpopulations In: 2017 IEEE Congress of Evolutionary Computation - CEC, pp 1864–1871 (2017) Springer, Berlin, Germany Castellini, A., Paltrinieri, D., Manca, V.: MP-GeneticSynth: inferring biological network regulations from time series Bioinformatics 31(5), 785–787 (2015) Chiu, B., Keogh, E., Lonardi, S.: Probabilistic discovery of time series motifs In: Proceedings of the 9th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD 2003, pp 493–498 ACM (2003) Davey, F.R., Huntington, S.: Age-related variation in lymphocyte subpopulations Gerontology 23, 381–389 (1977) Duchˆene, F., Garbay, C., Rialle, V.: Learning recurrent behaviors from heterogeneous multivariate time-series Artif Intell Med 39(1), 25–47 (2007) Franco, G., Jonoska, N., Osborn, B., Plaas, A.: Knee joint injury and repair modeled by membrane systems BioSystems 91(3), 473–488 (2008) 10 Franco, G., Manca, V.: A membrane system for the leukocyte selective recruitment In: Mart´ın-Vide, C., Mauri, G., P˘ aun, G., Rozenberg, G., Salomaa, A (eds.) WMC 2003 LNCS, vol 2933, pp 181–190 Springer, Heidelberg (2004) https://doi.org/ 10.1007/978-3-540-24619-0 13 11 Gruver, A.L., Hudson, L.L., Sempowski, G.D.: Immunosenescence of ageing J Pathol 211(2), 144–156 (2007) 12 Hallac, D., Vare, S., Boyd, S., Leskovec, J.: Toeplitz inverse covariance-based clustering of multivariate time series data In: Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD 2017, pp 215–223 ACM, New York (2017) 13 Hicks, M.J., Jones, J.F., Minnich, L.L., Wigle, K.A., Thies, A.C., Layton, J.M.: Age-related changes in T- and B-lymphocyte subpopulations in the peripheral blood Arch Pathol Lab Med 107(10), 518–523 (1983) 14 Hyndman, R.J., Athanasopoulos, G.: Forecasting: Principles and Practice, 2nd edn., O Texts (2014) Detection of Age-Related Changes in Networks of B Cells 597 15 Jerne, N.K.: Towards a network theory of the immune system Annales d’immunologie 125C(1–2), 373–389 (1974) 16 Keogh, E., Chu, S., Hart, D., Pazzani, M.: Segmenting time series: a survey and novel approach In: Data mining in Time Series Databases, pp 1–22 World Scientific, Singapore (1993) 17 Lavielle, M.: Detection of multiple changes in a sequence of dependent variables Stochast Process Appl 83(1), 79–102 (1999) 18 Lavielle, M., Teyssi`ere, G.: Detection of multiple change-points in multivariate time series Lith Math J 46(3), 287–306 (2006) 19 Lin, J., Keogh, E., Lonardi, S., Patel, P.: Finding Motifs in time series In: Proceedings of the Second Workshop on Temporal Data Mining, pp 52–68 ACM (2002) 20 Manca, V., Castellini, A., Franco, G., Marchetti, L., Pagliarini, R.: Metabolic P systems: a discrete model for biological dynamics Chin J Electron 22(4), 717–723 (2013) 21 Menshikov, I., Beduleva, L., Frolov, M., Abisheva, N., Khramova, T., Stolyarova, E., Fomina, K.: The idiotypic network in the regulation of autoimmunity: theoretical and experimental studies J Theor Biol 21(375), 32–9 (2015) 22 Perelson, A.S.: Immune network theory Immunol Rev 110, 5–33 (1989) 23 Simon, A.K., Hollander, G.A., McMichael, A.: Evolution of the immune system in humans from infancy to old age Proc R Soc B Biol Sci 282, 20143085 (2015) 24 Terzi, E., Tsaparas, P.: Efficient algorithms for sequence segmentation In: Proceedings of the 2006 SIAM International Conference on Data Mining, pp 316–327 SIAM (2006) 25 Vahdatpour, A., Amini, N., Sarrafzadeh, M.: Toward unsupervised activity discovery using multi-dimensional motif detection in time series In: Proceedings of the 21st International Joint Conference on Artifical Intelligence, IJCAI 2009, pp 1261–1266 (2009) 26 Veneri, D., Ortolani, R., Franchini, M., Tridente, G., Pizzolo, G., Vella, A.: Expression of CD27 and CD23 on peripheral blood B lymphocytes in humans of different ages Blood Transfus 7, 29–34 (2009) 27 Weiskopf, D., Weinberger, B., Grubeck-Loebenstein, B.: The aging of the immune system Transpl Int 22, 1041–1050 (2009) Author Index Adair, Jason 186 Alves, Maria João 172 Antunes, Carlos Henggeler Aringhieri, Roberto 549 Eftimov, Tome 76 Elezaj, Ogerta 268 172 Babaoglu, Ozalp 449 Bacher, Christopher 562 Bagattini, Francesco 244 Baioletti, Marco 401 Baldanzini, Niccolò 322 Bayá, Gabriel 436 Bayot, Roy Khristopher 337 Blum, Christian 506 Borghesi, Andrea 449 Bridi, Thomas 449 Brownlee, Alexander 186 Butchy, Adam A Cagnoni, Stefano 146 Calimeri, Francesco 540 Campana, Emilio F 121, 198 Cappanera, Paola 244 Caracciolo, Mirco 540 Castellini, Alberto 586 Castro, Natalia 496 Chrétien, Stéphane 234 Cohen, Michael 41 Consoli, S 16 Corrales, Juan Carlos 362 D’Agostino, Danny 121 Daolio, Fabio 186 Darses, Sébastien 234 Dell’Amico, Mauro 158 Dell’Anna, Davide 549 Di Bari, Gabriele 401 Dias, Joana 255, 483 Diez, Matteo 50, 121, 198 DiMaggio, Peter A 88 Carmo Lopes, Maria 255 Duma, Davide 549 Ferreira, Brígida 255 Ferreira, Graciela 496 Fleischner, Herbert 527 Fonlupt, Cyril 100 Fornacciari, Paolo 146 Franco, Giuditta 586 Galleguillos, Cristian 449 Galvani, Adriana 299 Gohari, Amir 376 Gonỗalves, Teresa 337 Griparić, Karlo 309 Gustavsson, Emil 63 Hadjidimitriou, Natalia Selini Hatsugai, Reiji 519 Hendriks, M 16 Horn, Matthias 506 Iemma, Umberto 198 İlker Birbil, Ş 376 Inaba, Mary 519 Jirstrand, Mats 63 Kaan Öztürk, M 376 Kalemi, Edlira 268 Kavaja, Juxhino 146 Kaya, Kamer 376 Kefato, Zekarias T 286 Kiziltan, Zeynep 449 Klocker, Benedikt 527 Koch, Thorsten 158 Kong, Min 414 Korošec, Peter 76 Koroušić Seljak, Barbara 76 Kustra, J 16 158 600 Author Index Landa-Silva, Dario 349 Lazzerini, Beatrice 210 Li, Mingxi 389 Liuzzi, Giampaolo 198 López, Iván Darío 362 Lucia, Angelo 88 Lucidi, Stefano 198 Ma, Chunfeng 414 Marinaki, Magdalene 133 Marinakis, Yannis 133 Marzullo, Aldo 540 Mauceri, Stefano 574 Mauttone, Antonio 436 Mavroeidis, D 16 Mazyad, Ahmad 100 McDermott, James 574 Mironov, S V 109 Miskov-Zivanov, Natasa Montresor, Alberto 286 Mordonini, Monica 146 Musarais, Boris 462 Nakada, Hidemoto 389 Ochoa, Gabriela 186 Pardalos, Panos M 414 Pavone, Mario 222 Pei, Jun 414 Pellegrini, Riccardo 198 Petkovic, Milena 158 Pierini, Marco 322 Pistolesi, Francesco 210 Plebe, Alice 222 Pleshakov, M G 109 Poggi, Agostino 146 Poggioni, Valentina 401 Radspieler, Gerald 309 Raidl, Günther R 506, 527, 562 Rela, Guillermo 426 Rigakis, Manousos 133 Rinaldi, Francesco 198 Robledo, Franco 426, 436, 496 Rocha, Humberto 255, 483 Romero, Pablo 426, 436, 496 Sala, Ramses 322 Salem, Ziad 309 Sayed, Khaled Schmickl, Thomas 309 Schoen, Fabio 244 Senov, Alexander 29 Serani, Andrea 50, 121, 198 Sheikh, Nasrullah 286 Shi, Peng 349 Sidorov, S P 109 Sỵrbu, Alina 449 Smith, Louis 574 Solimeo, Alex 146 Sonnessa, Michele 549 Stamile, Claudio 540 Sweeney, James 574 Tanimura, Yusuke 389 Tarnawski, Radosław 470 Telmer, Cheryl A Teytaud, Fabien 100 Thomas, Edward 88 Tomaiuolo, Michele 146 Trachanatzi, Dimitra 133 Tracolli, Mirco 401 Ulm, Gregor 63 Unold, Olgierd 470 Valencia, Cristian Heidelberg Ventura, Tiago 255 Vos, P 16 Yildirim, Sule 268 Zaleshin, Alexander 299 Zaleshina, Margarita 299 362 ... Giuffrida Renato Umeton (Eds.) • • Machine Learning, Optimization, and Big Data Third International Conference, MOD 2017 Volterra, Italy, September 14–17, 2017 Revised Selected Papers 123 Editors... Switzerland Preface MOD is an international conference embracing the fields of machine learning, optimization, and data science The third edition, MOD 2017, was organized during September 14–17, 2017. .. picturesque countryside of Tuscany The key role of machine learning, reinforcement learning, artificial intelligence, large-scale optimization, and big data for developing solutions to some of the greatest