Partially Supervised Learning: Second IAPR International Workshop, PSL 2013, Nanjing, China, May 13-14, 2013, Revised Selected Papers PDF
Preview Partially Supervised Learning: Second IAPR International Workshop, PSL 2013, Nanjing, China, May 13-14, 2013, Revised Selected Papers
Zhi-Hua Zhou Friedhelm Schwenker (Eds.) 3 Partially Supervised 8 1 8 AI Learning N L Second IAPR International Workshop, PSL 2013 Nanjing, China, May 2013 Revised Selected Papers 123 Lecture Notes in Artificial Intelligence 8183 Subseries of Lecture Notes in Computer Science LNAI Series Editors Randy Goebel University of Alberta, Edmonton, Canada Yuzuru Tanaka Hokkaido University, Sapporo, Japan Wolfgang Wahlster DFKI and Saarland University, Saarbrücken, Germany LNAI Founding Series Editor Joerg Siekmann DFKI and Saarland University, Saarbrücken, Germany For furthervolumes: http://www.springer.com/series/1244 Zhi-Hua Zhou Friedhelm Schwenker (Eds.) • Partially Supervised Learning Second IAPR International Workshop, PSL 2013 Nanjing, China, May 13-14, 2013 Revised Selected Papers 123 Editors Friedhelm Schwenker Zhi-Hua Zhou Abteilung Neuroinformatik National key Laboratory for Novel Universität Ulm Software Technology Ulm Nanjing University Germany Nanjing China ISSN 0302-9743 ISSN 1611-3349 (electronic) ISBN 978-3-642-40704-8 ISBN 978-3-642-40705-5 (eBook) DOI 10.1007/978-3-642-40705-5 Springer Heidelberg New York Dordrecht London Library of Congress Control Number: 2013950103 CR Subject Classification (1998): I.2.6, I.5, H.2.8, I.2, I.4, H.3 LNCS Sublibrary: SL 7 - Artificial Intelligence (cid:2)Springer-VerlagBerlinHeidelberg2013 Thisworkissubjecttocopyright.AllrightsarereservedbythePublisher,whetherthewholeorpartofthematerial is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproductiononmicrofilmsorinanyotherphysicalway,andtransmissionorinformationstorageandretrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed.Exemptedfromthislegalreservationarebriefexcerptsinconnectionwithreviewsorscholarlyanalysis ormaterialsuppliedspecificallyforthepurposeofbeingenteredandexecutedonacomputersystem,forexclusive use by the purchaser of the work. Duplication of this publication or parts thereof is permitted only under the provisionsoftheCopyrightLawofthePublisher’slocation,initscurrentversion,andpermissionforusemust always be obtained from Springer. Permissions for use may be obtained through RightsLink at the Copyright ClearanceCenter.ViolationsareliabletoprosecutionundertherespectiveCopyrightLaw. Theuseofgeneraldescriptivenames,registerednames,trademarks,servicemarks,etc.inthispublicationdoesnot imply,evenintheabsenceofaspecificstatement,thatsuchnamesareexemptfromtherelevantprotectivelaws andregulationsandthereforefreeforgeneraluse. Whiletheadviceandinformationinthisbookarebelievedtobetrueandaccurateatthedateofpublication,neither theauthorsnortheeditorsnorthepublishercanacceptanylegalresponsibilityforanyerrorsoromissionsthatmay bemade.Thepublishermakesnowarranty,expressorimplied,withrespecttothematerialcontainedherein. Printedonacid-freepaper SpringerispartofSpringerScience+BusinessMedia(www.springer.com) Preface Partially supervised learning (PSL) is a rapidly evolving area of machine learning, data mining and pattern recognition. In many applications unlabeled data may be relatively easy to collect, whereas labeling these data is difficult, expensive, and/or time-consumingasitrequirestheeffortofhumanexperts.PSLisageneralframework for learning with partial supervision, or learning with partially labeled data. In this framework, the supervision information might be a crisp label, or a label plus a confidence value, or it might be an imprecise and/or uncertain soft label defined through a certain type of uncertainty model, or it might be that information about a class label is not available. ThePSLframeworkthusgeneralizesmanykindsoflearningparadigmsincluding supervised and unsupervised learning, semi-supervised learning for classification and regression, transductive learning, semi-supervised clustering, multi-instance learning, weak label learning, policy learning in partially observable environments, etc. Therefore, PSL theories and algorithms are of great interest to various research communities. Research in the field of PSL is still in its early stages and has great potential for further growth, thus leaving plenty of room for further development. PSL2013received26fullsubmissions.TheProgramCommitteeconsistingof24 expertscarefullyreviewedthesubmissions,withthehelpofsomeexternalreviewers. Based on the reviews, 10 papers were selected for presentation at the workshop and inclusion inthepost-workshopproceedings.Authorswere requestedtoimprovetheir manuscriptsbyincorporatingreviewers’commentsandfeedbacksfromtheworkshop audience, leading to the revised selected papers in this volume. The workshop pro- gram was significantly enhanced by the invited talk of Prof. Dale Schuurmans of the University of Alberta, Canada. Thisworkshopwouldnothavebeenpossiblewithoutthehelpofmanyindividuals andorganizations.Firstofall,wewouldliketothanktheProgramCommitteemembers and reviewers for their great efforts in providing insightful comments on the submis- sions.Wealsowishtothankalltheauthorswhohavesubmittedtheirrecentworktothe workshop. The management of the papers, including the preparation of this proceed- ings volume, was done by the EasyChair conference management system. Special thanks go to the local arrangement and publicity chairs, Ming Li, Yang Yu, and Michael Glodek, for their outstanding contributionto the organizationof PSL 2013. VI Preface This workshop was organized by the LAMDA Group of the National Key Lab- oratory for Novel Software Technology, Nanjing University, China, and the Institute of Neural Information Processing, University of Ulm, Germany. We thank the International Association for Pattern Recognition (IAPR), IEEE Computer Society Nanjing Chapter, and the National Science Foundation of China for their support. July 2013 Zhi-Hua Zhou Friedhelm Schwenker Committee Chairs Zhi-Hua Zhou Nanjing University, China Friedhelm Schwenker University of Ulm, Germany Local Arrangement Chair Ming Li Nanjing University, China Publicity Chair Yang Yu Nanjing University, China Michael Glodek University of Ulm, Germany Program Committee Songcan Chen China Xiaofei He China Tom Heskes The Netherlands Steven C.H. Hoi Singapore Rong Jin USA Wee Sun Lee Singapore Hang Li China Yu-Feng Li China Cheng-Lin Liu China Marco Loog The Netherlands Marco Maggini Italy Fabio Roli Italy Dale Schuurmans Canada Masashi Sugiyama Japan Edmondo Trentin Italy Grigorios Tsoumakas Greece Haifeng Wang China Wei Wang China Kai Yu China Shipeng Yu Germany De-Chuan Zhan China VIII Committee Min-Ling Zhang China Jerry Zhu USA Jun Zhu China Additional Reviewers Giorgio Fumera Nan Li Wei Wu Yan-Ming Zhang Tingting Zhao Supported by National Science Foundation of China IAPR (International Association for Pattern Recognition) IEEE Computer Society Nanjing Chapter Contents Partially Supervised Anomaly Detection Using Convex Hulls on a 2D Parameter Space. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 Gabriel B. P. Costa, Moacir Ponti, and Alejandro C. Frery Self-Practice Imitation Learning from Weak Policy . . . . . . . . . . . . . . . . . . 9 Qing Da, Yang Yu, and Zhi-Hua Zhou Semi-Supervised Dictionary Learning of Sparse Representations for Emotion Recognition. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 Markus Kächele and Friedhelm Schwenker Adaptive Graph Constrained NMF for Semi-Supervised Learning . . . . . . . . 36 Qian Li, Liping Jing, and Jian Yu Kernel Parameter Optimization in Stretched Kernel-Based Fuzzy Clustering. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 Chunhong Lu, Zhaomin Zhu, and Xiaofeng Gu Conscientiousness Measurement from Weibo’s Public Information. . . . . . . . 58 Dong Nie, Lin Li, and Tingshao Zhu Meta-Learning of Exploration and Exploitation Parameters with Replacing Eligibility Traces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68 Michel Tokic, Friedhelm Schwenker, and Günther Palm Neighborhood Co-regularized Multi-view Spectral Clustering of Microbiome Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 Evgeni Tsivtsivadze, Hanneke Borgdorff, Janneke van de Wijgert, Frank Schuren, Rita Verhelst, and Tom Heskes A Robust Image Watermarking Scheme Based on BWT and ICA . . . . . . . . 91 Tao Wang, Jin Tang, Bin Luo, and Cheng Zhang A New Weighted Sparse Representation Based on MSLBP and Its Application to Face Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . 104 He-Feng Yin and Xiao-Jun Wu Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 Partially Supervised Anomaly Detection Using Convex Hulls on a 2D Parameter Space Gabriel B. P. Costa1, Moacir Ponti1(B), and Alejandro C. Frery2 1 Instituto de Ciˆencias Matem´aticas e de Computa¸ca˜o, Universidade de S˜ao Paulo, S˜ao Carlos, SP, 13566-590, Brazil {moacir, gpbcosta}@icmc.usp.br 2 Universidade Federal de Alagoas, Macei´o, AL13566-590, Brazil [email protected] Abstract. Anomaly detection is the problem of identifying objects appearing to be inconsistent with the remainder of that set of data. Detecting such samples is useful on various applications such as fault detection, frauddetection anddiagnostic systems. Partially super- vised methods for anomaly detection are interesting because they only need data labeled as one of the classes (normal or abnormal). In this paper, we propose a partially supervised framework for anomaly detec- tionbasedonconvexhullsinaparameterspace,assumingagivenprob- ability distribution. It can be considered a framework since it supports anymodelforthe“normal”samples.Weinvestigateanalgorithmbased onthisframework,assumingtheGaussiandistributionforthenotanom- alous(“normal”)data,andcomparedtheresultswiththeOne-classSVM andNa¨ıveBayesclassifiers,aswellastwostatisticalanomalydetectors. Theproposedmethodshowsaccuracyresultsthatarecomparableorbet- terthanthecompetingmethods.Furthermore,thisapproachcanhandle anyprobabilitydistributionormixtureofdistributions,allowingtheuser tochooseaparameterspacethatmodelsadequatelytheproblemoffind- ing anomalies. Keywords: Anomaly · Outlier · Semi-supervised learning 1 Introduction Anomalydetectionistheproblemoffindingpatterns(samples,individuals)with an unexpected behaviour. Barnett and Lewis [1] defined outliers (anomalies) as observations which appear to be inconsistent with the remainder of that set of data. Due to the nature of the problem, anomalies are often rare and dealing with them can help on applications such as fault detection, fraud detection, network intrusion,diagnosticsystems,andmedicalconditionmonitoring.Anomalydetec- tionmethodstakeasinputasampleorsetofsamples,andidentifywhetherthey are “normal” or “abnormal”, according to what is expected to be found. Note that, in this text, we will refer normal as the data that is consistent with the Z.-H.ZhouandF.Schwenker(Eds.):PSL2013,LNAI8183,pp.1–8,2013. DOI:10.1007/978-3-642-40705-51, (cid:2)c Springer-VerlagBerlinHeidelberg2013
The list of books you might like
