UsingCompactCoevolutionaryAlgorithmforMatching...

Research ArticleUsing Compact Coevolutionary Algorithm for MatchingBiomedical Ontologies

Xingsi Xue 1 Jie Chen1 Junfeng Chen 2 and Dongxu Chen3

1College of Information Science and Engineering Fujian University of Technology Fuzhou 350118 China2College of IOT Engineering Hohai University Nanjing 213022 China3Fujian Medical University Union Hospital Fuzhou 350001 China

Correspondence should be addressed to Xingsi Xue jack8375gmailcom

Received 8 May 2018 Accepted 30 August 2018 Published 8 October 2018

Academic Editor Carmen De Maio

Copyright copy 2018 Xingsi Xue et al )is is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Over the recent years ontologies are widely used in various domains such as medical records annotation medical knowledgerepresentation and sharing clinical guideline management andmedical decision-making To implement the cooperation betweenintelligent applications based on biomedical ontologies it is crucial to establish correspondences between the heterogeneousbiomedical concepts in different ontologies which is so-called biomedical ontology matching Although Evolutionary algorithms(EAs) are one of the state-of-the-art methodologies to match the heterogeneous ontologies huge memory consumption longruntime and the bias improvement of the solutions hamper them from efficiently matching biomedical ontologies To overcomethese shortcomings we propose a compact CoEvolutionary Algorithm to efficiently match the biomedical ontologies Particularlya compact EA with local search strategy is able to save the memory consumption and runtime and three subswarms with differentoptimal objectives can help one another to avoid the solutionrsquos bias improvement In the experiment two famous testing casesprovided by Ontology Alignment Evaluation Initiative (OAEI 2017) ie anatomy track and large biomed track are utilized to testour approachrsquos performance )e experimental results show the effectiveness of our proposal

1 Introduction

Ontologies provide a shared and common vocabulary forrepresenting a domain of knowledge [1] Over the recentyears ontologies are widely used in various domains such asmedical records annotation [2] medical knowledge repre-sentation and sharing clinical guidelines management [3]and medical decision-making [4] However most biomedicalontologies are developed independently by different expertswho might define one entity with different names or indifferent ways causing the problem of ontology heteroge-neity For example to describe the muscles that surround andpower the human heart the National Cancer Institutersquosthesaurus and ontology (NCI) [5] use the name ldquoMyocar-diumrdquo whereas the Foundation Model of Anatomy (FMA)[6] uses ldquoCardiac Muscle Tissuerdquo To implement the co-operation between intelligent applications based on bio-medical ontologies it is crucial to establish correspondences

between the heterogeneous biomedical concepts in differentontologies which is so-called biomedical ontology matching

Recently Evolutionary Algorithms (EAs) are one of thestate-of-the-art methodologies to match the heterogeneousontologies [7] However huge memory consumption longruntime and the bias improvement of the solutions hamperEA-based ontology matching techniques from efficientlymatching biomedical ontologies )us besides the quality ofalignments main memory consumption and runtimeneeded by the ontology matcher are of prime importancewhen matching the biomedical ontologies In this paper wepropose to use the compact EA [8] which utilizes a prob-abilistic representation of the population to save thememory consumption of classic EA )en we introduce thelocal search strategy into its evolving process to balance theexploration and exploitation and reduce the runtimeneeded On this basis we further propose a compact Co-evolutionary Algorithm which utilizes three subswarms

HindawiComputational Intelligence and NeuroscienceVolume 2018 Article ID 2309587 8 pageshttpsdoiorg10115520182309587

with different objectives to help one another to avoid thesolutionrsquos bias improvement caused by traditional metric f-measure [9]

)e rest of the paper is organized as follows Section 2describes the related works Section 3 gives some basicconcepts of ontology ontology alignment and the similaritymeasures Section 4 presents the optimal model problem andthe details of the compact Coevolutionary Algorithm formatching biomedical ontologies Section 5 gives the ex-perimental results and relevant analysis finally Section 6draws the conclusions

2 Related Work

21 Evolutionary Algorithm-Based Ontology MatchingTechnique Due to the complex and time-consuming natureof the ontology matching process EA-based methods couldpresent a good methodology for obtaining ontology align-ments and indeed have already been applied to solve theontology alignment problem by reaching acceptable results[10] Different from other EA based approaches [11ndash13]which models the ontology alignment process as a meta-matching problem ie how to determine the best appro-priate weight configuration in ontology matching processin order to obtain a satisfactory alignment in this workontology matching problem is considered as a global entitymatching problem Genetic Algorithm-Based OntologyMatching (GAOM) [14] is the representative system whichutilized Genetic Algorithm (GA) to determine the optimalontology alignment Particularly GAOM utilizes thechromosomes to describe the potential alignments betweentwo ontologies and utilizes GAs to determine the optimalsolution Besides MapPSO and MapEVO [15] whichexploited the Particle SwarmOptimization Algorithm (PSO)[16] and Evolutionary Programming (EP) [17] respectivelyalso adopted this idea Acampora et al [18] designeda Memetic Algorithm (MA) which introduced a local searchprocess to improve the performance of EA More recentlyXue et al [19 20] respectively used the compact EA andcompact Population-Based Incremental Learning Algorithm(PBIL) to save the memory consumption without sacrificingthe solutionrsquos quality Compact EA and compact PBILrepresented the population as a probability vector (PV) overthe set of solutions and are operationally equivalent to theorder-one behaviour of the simple EA with uniformcrossover In this way a much smaller number of solutionsmust be stored in the memory thus significantly reducingthe memory consumption

22CoevolutionaryAlgorithm )eCoevolutionary Algorithm[21] makes multiple swarms simultaneously evolve andcommunicate with one another to improve the searchperformance Currently distributed coevolution is the mostpopular coevolving process which shares the search in-formation among multiple swarms through the populationmigration strategy During the searching process differentswarms have evolving strategies and configurations Tanet al [22] proposed to decompose the problemrsquos solution

vector into multiple swarms to evolve simultaneously Muand Liu [23] presented anM-elite Coevolutionary Algorithmthat applied different elite strategies in the coevolvingprocess )e elite centered swarm has the highest priorityand other swarms implemented the cooperative coevolvingprocess In [24] a parallel evolving mechanism was designedby dividing the population into three swarms that evolvedindependently However all the swarms use the sameevolving strategy and the swarmrsquos evolving process swarmwas relatively independent which decreased the algorithmrsquosexploration and exploitation ability More recently Wanget al [25] proposed a two-elite strategy which makes use ofthe differences between two elites to guide the wholeevolving process

Different from all the techniques mentioned above inthis work we propose a compact coevolutionary Algorithmto match the biomedical ontologies which combines theadvantages of the compact EA and coEvolutionary Algo-rithm to save the memory consumption and runtime andovercome the bias improvement of solutions

23 Preliminaries

231 Ontology Ontology Alignment and Ontology MatchingProcess In this work an ontology is defined as a quadrupleO (C P I A) where

(i) C is the class set ie the set of concepts thatpopulate the domain of interest

(ii) P is the property set ie the set of relations betweenthe concepts of domain

(iii) I is the instance set ie the set of objects in the realworld representing the instances of a concept and

(iv) A is the axiom set ie the statements that say whatis true about the modeled domain

An alignment A between two ontologies O1 and O2 isdefined as a set of correspondences and each correspondenceis a triple (e1 e2 n) where e1 and e2 are the entities in O1 andO2 respectively and n isin [0 1] is a confidence value holdingfor the correspondence between them In this work the re-lation existing between two ontology entities is the equiva-lence () )e ontology matching process can be defined asa function θ(O1 O2 p r) [26] where p is the parameter setand r is the resource set Ontology matching process returnsa new alignment AN between ontologies O1 and O2

232 Concept Similarity Measure Concept similaritymeasure is the foundation of biomedical ontology matching[27] In this work we utilize an asymmetrical conceptsimilarity measure to calculate the biomedical conceptsrsquosimilarity values First for each biomedical concept weconstruct a profile for it by collecting the label commentand property information such as label domain and rangefrom itself and all its direct descendants )en the similarityof two biomedical concepts c1 and c2 is measured based onthe similarity of their profiles p1 and p2 which can becalculated by the following two asymmetrical measures

2 Computational Intelligence and Neuroscience

sim1 p1 p2( 1113857 p1 capp2

11138681113868111386811138681113868111386811138681113868

p1

sim2 p1 p2( 1113857 p1 capp2

11138681113868111386811138681113868111386811138681113868

p2

(1)

where |p1| and |p2| are the cardinalities of the profile p1and p2 respectively |p1 capp2| is the number of identicalelements in p1 and p2)e similarity value of e1 and e2 is equalto (sim1(p1 p2) + sim2(p1 p2))2 when |sim1(p1 p2)minussim2(p1 p2)|le δ and otherwise 0

In this work δ is the threshold to measure the extent ofthe semantic equivalence between sim1(p1 p2) andsim2(p1 p2) When the similarity value between two profileelements is above the threshold they are identified as se-mantically similar Generally δ should be set relatively smallto reflect sim1(e1 e2) and sim2(e1 e2) have little differencewhen the entity e1 and e2 are semantically equivalentHowever if δ is too small we would miss many semanticallyequivalent terms )erefore the suggested domain of δ is[001 010] In this work to obtain a suitable we conducteda pre-experiment on the benchmark by varying the value of δin its suggested domain and found the semantic equivalenceperforms well when δ is assigned to 006

Moreover the similarity value of two profile elements iscalculated by N-gram distance [28] which is the most per-forming string-based similarity measure for the biologicalontology matching problem and a linguistic measure whichcalculate a synonymy-based distance through the UnifiedMedical Language System (UMLS) [29] Given two words w1and w2 their similarity sim2(w1 w2) is equal to 1 when twowords are synonymous and otherwise Nminus gram(w1 w2)

24 Compact Coevolutionary Algorithm

241 Rough Alignment Evaluations In this work we sup-pose that in the golden alignment one concept in theontology is matched with only one concept in the otherontologies and vice versa Two rough alignment evaluationsieMatchCoverage andMatchRatio are utilized to measurethe alignments quality In particular MatchCoverage isutilized to approximate recall [9] which calculates thefraction of concepts which exist in at least one correspon-dence in the resulting alignment in comparison to the totalnumber of concepts in the ontology )e formula of it ispresented as follows

MatchCoverage CO1minusMatch

11138681113868111386811138681113868

11138681113868111386811138681113868 + CO2minusMatch

11138681113868111386811138681113868

11138681113868111386811138681113868

CO1

11138681113868111386811138681113868

11138681113868111386811138681113868 + CO2

11138681113868111386811138681113868

11138681113868111386811138681113868ϵ[0 1] (2)

where

(i) CO1minusMatch andCO2minusMatch are thematched concept setsof ontology O1 and O2 respectively and

(ii) CO1and CO2

are the concept sets of ontology O1 andO2 respectively

And MatchRatio is used to approximate precision [9]which calculates the ratio between the number of found

correspondences and the number of matched concepts )eformula of it is presented as follows

MatchRatio CO1minusMatch

11138681113868111386811138681113868

11138681113868111386811138681113868 + CO2minusMatch

11138681113868111386811138681113868

11138681113868111386811138681113868

2 middot CorrO1minusO2

11138681113868111386811138681113868

11138681113868111386811138681113868ϵ[0 1] (3)

where

(i) CorrO1minusO2is the correspondence set in the align-

ment and(ii) CO1minusMatch andCO2minusMatch are thematched concept sets

of ontology O1 and O2 respectively

In most instances it requires considering bothMatchCoverage and MatchRatio to measure the align-mentrsquos quality By referring to the most common combiningfunction f-measure [9] we define MatchFmeasure asfollows

MatchFmeasure 2 timesMatchCoverage middot MatchRatio

MatchCoverage + MatchRatio

(4)

242 2e Optimal Model for Ontology Entity MatchingProblem Given two biomedical ontologies O1 and O2 wetake maximizing MatchFmeasure as the goal and theoptimal model for ontology entity matching problem can bedefined as follows

max MatchFmeasure(X)

st X x1 x2 x O1| | x O1| |+11113874 1113875T

xi 1 2 x O2| |

x O1| |+1 isin [0 1]

(6)

where the decision variable X represents an alignment be-tween O1 and O2 xi represents the ith correspondencebetween ith concept in O1 and xith concept in O2 |O1| and|O2| are the cardinalities of the concept set in O1 and O2respectively and x|O1|+1 isin [0 1] is the threshold to filter thefinal alignment

One of the shortcomings of MatchFmeasure is that theimprovement of it does not say anything about whether bothMatchCoverage andMatchRatio are simultaneously improvedor not In other words no matter how large a measuredimprovement in MatchFmeasure is it can still be extremelydependent on the improvement on one of the individualmetrics [30] To overcome this bias improvement we proposea compact coevolutionary Algorithm which has three PVsthat characterize subswarms that aim at maximizingMatchCoverage MatchRatio and MatchFmeasure re-spectively)rough the cooperation of three PVs we dedicateto ensure the simultaneous improvement on MatchCoverageand MatchRatio during the evolving process

243 Compact Evolutionary Algorithm Model-based op-timization using probabilistic modeling of the search space is

Computational Intelligence and Neuroscience 3

one of the areas where research on Compact EvolutionaryAlgorithm (CEA) has considerably advanced in recent yearsIn each generation CEA updates the probability vector(PV) which is a probabilistic model describing the uni-variate statistics of the best solutions and then uses it togenerate new candidate solutions By employing the PVinstead of a population of solutions to simulate the behaviorof classic EA a much smaller number of individuals isneeded to be stored in the memory )us CEA can sig-nificantly reduce the memory consumption [31] In order tofurther improve CEA performance we introduce the localsearch strategy into CEArsquos evolving process )is marriagebetween global search and local search is helpful in reducingthe possibility of the premature convergence and increasingthe convergence speed

In the next three main components of CEA iechromosome-encoding mechanism probability vector andlocal search strategy are respectively presented

(1) Chromosome-Encoding Mechanism in this work thegenes are encoded through the binary codingmechanism and can be divided into two parts )efirst part stands for the correspondences in thealignment and the other one stands for a thresholdGiven the total number n1 and n2 of two biomedicalconcepts in ontologies the first part of a chromo-some (or PV) consists of n1 gene segments and thebinary code length (BCL) of each gene segment isequal to log2(n2) + 05 which ensures each genesegment could present any target ontology classrsquosindex while the second part of a chromosome (orPV) has only one gene segment whose BCL is equalto log2(1numAccuracy) + 05 which can ensurethis gene segment could present any threshold valueunder the numerical accuracy numAccuracy )usthe total length of the chromosome (or PV) is equalto n1 times log2(n2) + 05 + log2(1numAccuracy) + 05

Given a gene segment geneSeg geneBit1 geneBit2 middot middot middot1113864

geneBitn where geneBiti is the ith gene bit value of thegene segment we decode to obtain a decimal number whosevalue is equal to 1113936

ni12geneBiti In particular with respect to

the first part decoding results the decimal numbers obtainedrepresent the indexes of the target classes where 0 means thesource instance is not mapped to any target ontologyrsquos classWith regard to the second part of decoding result thedecimal number obtained should multiply the thresholdrsquosnumerical accuracy Last but not least if a decimal numberd obtained is larger than u we will replace it with ud

(2) Probability Vector in general CEA aims at generatinga PVwhich represents a population of high evaluationsolutions and its operations take place directly on thePV In this work the number of elements in PV isequal to the number of individualrsquos gene bits and eachelementrsquos value is in [01] and here is an example onhow to use PV (05 09 03 08)T to generate a newsolution First generate four random numbers suchas 06 05 08 and 09 )en compare the numberswith the elements in PV accordingly to determine the

new generated individualrsquos gene values For examplesince 06gt 05 the first gene bitrsquos value of the newsolution is 0 and similarly the remaining gene bitsrsquovalues are 1 0 and 0 respectively In this way the newsolution we obtain is 0100 By repeating this pro-cedure we can obtain various individuals In additionif 0100 is the elite solution in the current generationPV should be updated according to its informationGiven PVrsquos update rate say 01 if the gene value of theelite is 0 the corresponding element of PV will minus01 otherwise add 01 In this way the updated PV is(04 10 02 07)T

(3) Local Search Strategy local search process tries toimprove the elite solution by searching in theneighborhood of it In this work we utilize a cross-over operator to implement the local search processwhich randomly copies a sequential fragment ofindnewrsquos genes into the corresponding positions ofindneighbor to generate a new solution For the sake ofclarity given the length of the chromosome len andthe crossover probability pc the pseudocode of thebinary crossover operator is shown in Algorithm 1

)is procedure is similar with the two-point crossoverwhere the first cut point is randomly selected from1 2 middot middot middot len and the second point is determined such thatL consecutive genes (counted in a circular manner) are takenfrom indnew Since indnew and indelite are both generatedthrough the PV most of their gene bit values are the same)erefore even when pc is large indneighbor only mutatesa few gene bit values of indelite In this sense this variationoperator can be considered fairly exploitative

244 Pseudocode of Compact Coevolutionary AlgorithmIn this work we use three PVs to represent the subswarmsfor maximizing MatchRatio MatchCoverage andMatchFmeasure respectively In particular the PV hererepresents the population that consists of the solutions of itscorresponding representative subproblem and this prob-lemrsquos neighbor subproblems Finally these PVs help eachother in the process of determining three representa-tive solutions which are given in the following Here wemark three representative subproblems of maximizingMatchRatio maximizing MatchCoverage and maximizingMatchFmeasure with the symbols Pmr Pmc and Pmf

(1) indneighbor indelitecopy()(2) generate i round(rand(0 len))(3) indneighbor[i] indnew[i](4) while(random(0 1)ltpc)

(5) i i + 1(6) if (i len)

(7) i 0(8) end if

(9) indneighbor[i] indnew[i](10) end while

ALGORITHM 1


Input (i) O1 and O2 two biomedical ontologies(ii) len the length of PV(iii) maxGen maximum number of generations(iv) UR PVrsquos update rate(v) pc crossover probability(vi) pm mutation probability(vii) MR mutation rate

Output the solution with best MatchFmeasureStep 1 Initialization

Step 11 Set the generation gen 0Step 12 Set the neighbor subproblem of Pmr and Pmc as Pmf and the neighbor subproblems of Pmf as Pmr and PmcStep 13 Initialize PVmr PVmc and PVmf by setting all the probabilities inside as 05Step 14 UsingPVmr PVmc and PVmf to generate the elites which are marked with symbols elitemr elitemc and elitemf for PmrPmc and Pmf respectively

Step 2 Evolving process

Step 21 Update PVmr PVmc and PVmf respectivelyTake updating PVmr for instance the procedures of updating PVmc and PVmf is similar to it

Step 211 Crossover

(1) Generate a new individual indnew through PVmr(2) [winner loser] compete(elitemr indnew)(3) if(winner indnew)(4) elitemr indnew(5) for i 0 iltPVmrlength i++(6) if(winner[i] 1)(7) PVmr[i] PVmr[i] + UR(8) if(PVmr[i]gt 1)(9) PVmr[i] 1(10) else(11) PVmr[i] PVmr[i]minusUR(12) if (PVmr[i]lt 0)(13) PVmr[i] 0

Step 212 Mutation

(14) for(i 0 i ltlen i++)(15) if((random(0 1)ltPm)(16) PVmr[i] PVmr[i]times (1minusMR) + random(0or1) times MR

Step 213 Local search

(17) Generate an individual indnew through PVmr(18) indneighbor elitemrcopy()(19) Generate i round(random(0 len))(20) indneighbor[i] indnew[i](21) while((random(0 1) lt Pc))(22) i i+ 1(23) if((i indneighborlength))(24) i 0(25) indneighbor[i] indnew[i](26) end While(27) [winner loser] compete(elitemr indneighbor)(28) if((winner indnew))(29) elitemr indnew

Step 22 Update PVmr PVmc and PVmf mutuallyFor Pmr (or Pmc) PVtemp is generated by applying the pc-based uniform crossover operator [32] on the PVmr (or Pmc) and itsneighbor subproblemrsquos probability vector PVmf )en generate an individual a through PVtemp and try to update the PVmr

and PVmf through the competition with elitemr (or elitemc) and elitemf

Continued


respectively and three PVs for solving Pmr Pmc and Pmf

with the symbols PVmr PVmc and PVmf respectively Wepresent the pseudocode of compact Coevolutionary Al-gorithm in Algorithm 2

25 Experimental Results and Analysis In this work weexploit the Anatomy (httpoaeiontologymatchingorg2017anatomyindexhtml) and Large Biomed (httpwwwcsoxacukisgprojectsSEALSoaei2017) track to study the ef-fectiveness of our approach which are provided by the On-tologyAlignment Evaluation Initiative (OAEI 2017) (httpoaeiontologymatchingorg2017) )e Anatomy track includes twoontologies (1 task) ie the Adult Mouse Anatomy (AMA)ontology (2744 classes) and a part of NCI describing thehuman anatomy (3304 classes) Large Biomed track (3tasks) aims at finding alignments between FMA SNOMEDCT and NCI which respectively contains 78989 122464and 66724 classes Particularly )e large Biomedic track issplit into three matching problems FMA-NCI FMA-SNOMED and SNOMED-NCI and each matching prob-lem in these tasks involving different fragments of the inputontologies

)e Compact Coevolutionary Algorithm uses the fol-lowing parameters which represent a trade-off setting ob-tained in an empirical way to achieve the highest averagealignment quality on all exploited testing datasets

(i) Numerical accuracy 001(ii) Update rate 01(iii) Crossover probability 06(iv) Mutation probability 003(v) Mutation rate 005(vi) Maximum generation 3000

3 Results and Analysis

In order to compare the quality of our proposal with theparticipants of OAEI 2017 (httpoaeiontologymatchingorg

2017resultsindexhtml) and Population-Based IncrementalLearning Algorithm (PBIL) [20] which is a state-of-the-artcompact EA-based ontology matching technique we evaluatethe obtained alignments with traditional recall precision andf-measure PBIL and our approachrsquos results in Table 1 andTable 2 are the mean values in thirty time independent exe-cutions )e symbols P R and F in tables stand for precisionrecall and f-measure respectively

As can be seen from Table 1 our approachrsquos f-measureoutperforms all the competitors and our approachrsquos run-time is ranked the 4th place In Table 2 our approachrsquos f-measure is the highest in task1 task2 and task3 For therunning time in task1 and task 2 our approach is in the 3rdplace and 4th place in task3 In both tracks our approachoutperforms AML which is the top ontology matcher anddeveloped primarily for the biomedical ontology matchingin all tasks in terms of f-measure and the runtime in ourapproach is also very close to or less than AML )e ex-perimental results show that the cooperation among threeswarms with different objectives can effectively overcomethe bias improvements and improve the quality of bio-medical ontology alignments

In particular PBIL works with one PV but our ap-proach utilizes three PVs to cooperate with each otherduring the evolving process to improve the solutionrsquosquality As can be seen from the experimental results al-though our approach takes only a little more runtime thanPBIL the qualities of our results are much better than PBILin terms of both recall and precision which shows that ourapproach can effectively overcome the bias improvement ofsolutions in PBIL

4 Conclusion

In this work in order to overcome the drawbacks in tra-ditional E-based ontology matching techniques we for thefirst time propose a compact Coevolutionary Algorithm toefficiently match the biomedical ontologies In our approachthree PVs are utilized to characterize three subswarms that

For PVmf PVtemp is generated through applying the uniform crossover operator between PVmr and Pmc which are itsneighbor subproblemsrsquo PVs )en generate an individual a through PVtemp and try to update the Pmf through thecompetition with elitemf

Step 3 Stopping Criteria

(30) if (maxGen is reached)(31) stop and the elite with best MatchFmeasure(32) else(33) gen gen+1(34) go to Step 2(35) end if

In the evolving process we first update PVmr PVmc and PVmf respectively (Step 21) which is equivalent to the process ofupdating the solutions of Pmr Pmc and Pmf )en we update PVmr PVmc and PVmf mutually (Step 22) which is equal toupdating the solutions of Pmr Pmc and Pmf through their shared neighbor subproblemsrsquo solutions ie using the information ofa PV to help its neighbor PVs

ALGORITHM 2


take as objectives maximizing MatchCoverage MatchRatioand MatchFmeasure respectively and in each generationPVs are first updated with CEA paradigm and then help eachother to search for better solutions in the search space In theexperiment OAEI 2017rsquos Anatomy track and Large Biomedtrack are utilized to test our approachrsquos performance and theresults show that our approach can efficiently determinebetter ontology alignments than state-of-the-art biomedicalontology matching techniques

Data Availability

)e data used to support the findings of this study have notbeen made available because of the protection of technicalprivacy and confidentiality

Conflicts of Interest

)e authors declare that they have no conflicts of interest

Acknowledgments

)is work is supported by the National Natural ScienceFoundation of China (Nos 61503082 and 61403121) NaturalScience Foundation of Fujian Province (No 2016J05145)Scientific Research Startup Foundation of Fujian Universityof Technology (No GY-Z15007) Scientific Research De-velopment Foundation of Fujian University of Technology(No GY-Z17162) and Fujian Province Outstanding YoungScientific Researcher Training Project (No GY-Z160149)

References

[1] S M Falconer Cognitive support for semi-automatic ontologymapping PhD dissertation University of Victoria VictoriaCanada 2009

[2] H Lacuteopez-Fernacuteandez M Reboiro-Jato D Glez-Pentildea et alldquoBioannote a software platform for annotating biomedicaldocuments with application in medical learning environ-mentsrdquo Computer Methods and Programs in Biomedicinevol 111 no 1 pp 139ndash147 2013

[3] D Isern D Sacuteanchez and A Moreno ldquoOntology-drivenexecution of clinical guidelinesrdquo Computer Methods andPrograms in Biomedicine vol 107 no 2 pp 122ndash139 2012

[4] P De Potter H Cools K Depraetere et al ldquoSemantic patientinformation aggregation and medicinal decision supportrdquoComputer Methods and Programs in Biomedicine vol 108no 2 pp 724ndash735 2012

[5] J Golbeck G Fragoso F Hartel et al ldquo)e National CancerInstitutersquos thesaurus and ontologyrdquo Web Semantics ScienceServices and Agents on the World Wide Web vol 1 no 1pp 75ndash80 2011

[6] C Rosse and J L Mejino Jr ldquoA reference ontology forbiomedical informatics the foundational model of anatomyrdquoJournal of Biomedical Informatics vol 36 no 6 pp 478ndash5002003

[7] X Xue and J Liu ldquoCollaborative ontology matching based oncompact interactive evolutionary algorithmrdquo Knowledge-Based Systems vol 137 pp 94ndash103 2017

[8] S Baluja ldquoPopulation-based incremental learning a methodfor integrating genetic search based function optimizationand competitive learningrdquo Department Of Computer ScienceCarnegie Mellon University Pittsburgh PA USA Tech Rep1994

[9] C J V Rijsberge Information Retrieval University ofGlasgow London UK 1975

[10] X Xue and Y Wang ldquoUsing memetic algorithm for instancecoreference resolutionrdquo IEEE Transactions on Knowledge andData Engineering vol 28 no 2 pp 580ndash591 2016

[11] J Martinez-Gil E Alba and J Montes ldquoOptimizing ontologyalignments by using genetic algorithmsrdquo in Proceedings ofFirst International Workshop on Nature based reasoning for

Table 1 Comparison of our approach with the participants inOAEI 2017 on anatomy track

System R P F Runtime (second)AML 093 095 094 37YAM-BIO 092 094 093 70POMap 090 094 093 808LogMapBio 089 088 089 820XMap 086 092 089 37LogMap 084 091 088 22KEPLER 074 095 083 234LogMapLite 072 096 082 19SANOM 077 089 082 295Wiki2 073 088 080 2204ALIN 033 099 050 836EA 076 088 078 22Our approach 094 097 095 34

Table 2 Comparison of our approach with the participants inOAEI 2017 on the Large Biomed track

System R P F Runtime (second)Task1 whole FMA and NCI ontologiesXMap 085 088 087 130AML 087 084 086 77YAM-BIO 089 082 085 279LogMap 081 086 083 92LogMapBio 083 082 083 1552LogMapLite 082 067 074 10Tooll 074 069 071 1650PBIL 076 088 078 22Our approach 087 089 088 72Task2 whole FMA and SNOMED ontologiesXMap 084 077 081 625YAM-BIO 073 089 080 468AML 069 088 077 177LogMap 065 084 073 477LogMapBio 065 081 072 2951LogMapLite 021 085 034 18Tooll 013 087 023 2140PBIL 072 074 072 147Our approach 081 084 082 183Task3 whole SNOMED and NIC ontologiesAML 067 090 077 312YAM-BIO 070 083 076 490LogMapBio 064 084 073 4728LogMap 060 087 071 652LogMapLite 057 080 066 22XMap 055 082 066 563Tooll 022 081 034 1105PBIL 064 081 071 304Our approach 073 088 079 326


the semantic Web vol 419 pp 31ndash45 Karlsruhe GermanyNovember 2008

[12] J M V Naya M M Romero and J P Loureiro ImprovingOntology Alignment through Genetic Algorithms InformationScience Reference Hershey New York 2010 pp 240ndash259

[13] A-L Ginsca and A Iftene ldquoUsing a genetic algorithm foroptimizing the similarity aggregation step in the process ofontology alignmentrdquo in Proceedings of 9th Roedunet In-ternational Conference pp 118ndash122 Sibiu Romania June 2010

[14] J Wang Z Ding and C Jiang ldquoGAOM genetic algorithmbased ontology matchingrdquo in Proceedings of IEEE AsiaCPa-cific Conference on Services Computing pp 617ndash620Guangzhou China December 2006

[15] J Bock and J Hettenhausen ldquoDiscrete particle swarm opti-misation for ontology alignmentrdquo Information Sciencesvol 192 pp 152ndash173 2012

[16] J Kennedy Particle Swarm Optimization Springer BostonMA USA 2011

[17] L J Fogel A J Owens and M J Walsh Artificial Intelligencethrough Simulated Evolution John Wiley Chichester UK1966

[18] G Acampora V Loia S Salerno et al ldquoA hybrid evolutionaryapproach for solving the ontology alignment problemrdquo In-ternational Journal of Intelligent Systems vol 27 pp 189ndash2162012

[19] X Xue J Liu P Tsai et al ldquoOptimizing ontology alignmentby using compact genetic algorithmrdquo in Proceedings of 11thInternational Conference on Computational Intelligence andSecurity pp 231ndash234 Guangzhou China December 2016

[20] X Xue and J Chen ldquoOptimizing ontology alignment throughhybrid population-based incremental learning algo-rithmrdquoMemetic Computing pp 1ndash9 2018 In press

[21] P R Ehrlich and P H Raven ldquoButterflies and plants a studyin coevolutionrdquo Evolution vol 18 no 4 pp 586ndash608 1964

[22] Y J Y K C Tan and C K Goh ldquoA distributed cooperativecoevolutionary algorithm for multiobjective optimizationrdquoIEEE Transactions on Evolutionary Computation vol 10no 5 pp 527ndash549 2006

[23] L C J C H Mu and Y Liu ldquoM-elite coevolutionary algo-rithm for numerical optimizationrdquo Journal of Softwarevol 20 no 11 pp 2925ndash2938 2009

[24] Q Zhou and W J Luo ldquoA novel multi-population geneticalgorithm for multiple-choice multidimensional knapsackproblemrdquo in Proceedings of 5th International Symposium onAdvances in Computation and Intelligence Berlin GermanySpringer-Verlag October 2010 pp 148ndash157

[25] X Wang Q Liu Q Fu and L Zhang ldquoDouble elite co-evolutionary genetic algorithmrdquo Journal of Software vol 23no 4 pp 765ndash775 2012

[26] J Euzenat and P Valtchev ldquoSimilarity-based ontologyalignment in owlliterdquo in Proceedings of 16th European Con-ference on Artificial Intelligence Proceeding pp 333ndash337Valencia Spain August 2004

[27] A Maedche and S Staab ldquoMeasuring similarity betweenontologiesrdquo in Proceedings of 14th International Conferenceon Knowledge Engineering and Knowledge Managementpp 251ndash263 Ischia Island Italy July 2002

[28] G Kondrak ldquoN-gram similarity and distancerdquo in In-ternational Symposium on String Processing and InformationRetrieval pp 115ndash126 Springer Glasgow UK October 2005

[29] O Bodenreider ldquo)e unified medical language system (umls)integrating biomedical terminologyrdquo Nucleic Acids Researchvol 32 no 90001 pp D267ndashD270 2004

[30] X Xue and Y Wang ldquoOptimizing ontology alignmentsthrough a memetic algorithm using both match f-measureand unanimous improvement ratiordquo Artificial Intelligencevol 223 pp 65ndash81 2015

[31] F Neri G Iacca and E Mininno Compact OptimizationHandbook of Optimization Springer Berlin Germany 2013

[32] G Syswerda ldquoUniform crossover in genetic algorithmsrdquo inProceedings of 2ird International Conference on GeneticAlgorithms pp 2ndash9 San Francisco USA June 1989


Computer Games Technology

International Journal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Journal ofEngineeringVolume 2018

Advances in

FuzzySystems


Volume 2018


ReconfigurableComputing



Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

thinspArtificial Intelligence

Hindawiwwwhindawicom Volumethinsp2018


Civil EngineeringAdvances in


Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications


Hindawi

wwwhindawicom Volume 2018

Advances in

Multimedia


Biomedical Imaging



Engineering Mathematics


RoboticsJournal of



Computational Intelligence and Neuroscience


Mathematical Problems in Engineering

Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018


Human-ComputerInteraction

Advances in


Scientic Programming

Submit your manuscripts atwwwhindawicom

with different objectives to help one another to avoid thesolutionrsquos bias improvement caused by traditional metric f-measure [9]

)e rest of the paper is organized as follows Section 2describes the related works Section 3 gives some basicconcepts of ontology ontology alignment and the similaritymeasures Section 4 presents the optimal model problem andthe details of the compact Coevolutionary Algorithm formatching biomedical ontologies Section 5 gives the ex-perimental results and relevant analysis finally Section 6draws the conclusions

2 Related Work

21 Evolutionary Algorithm-Based Ontology MatchingTechnique Due to the complex and time-consuming natureof the ontology matching process EA-based methods couldpresent a good methodology for obtaining ontology align-ments and indeed have already been applied to solve theontology alignment problem by reaching acceptable results[10] Different from other EA based approaches [11ndash13]which models the ontology alignment process as a meta-matching problem ie how to determine the best appro-priate weight configuration in ontology matching processin order to obtain a satisfactory alignment in this workontology matching problem is considered as a global entitymatching problem Genetic Algorithm-Based OntologyMatching (GAOM) [14] is the representative system whichutilized Genetic Algorithm (GA) to determine the optimalontology alignment Particularly GAOM utilizes thechromosomes to describe the potential alignments betweentwo ontologies and utilizes GAs to determine the optimalsolution Besides MapPSO and MapEVO [15] whichexploited the Particle SwarmOptimization Algorithm (PSO)[16] and Evolutionary Programming (EP) [17] respectivelyalso adopted this idea Acampora et al [18] designeda Memetic Algorithm (MA) which introduced a local searchprocess to improve the performance of EA More recentlyXue et al [19 20] respectively used the compact EA andcompact Population-Based Incremental Learning Algorithm(PBIL) to save the memory consumption without sacrificingthe solutionrsquos quality Compact EA and compact PBILrepresented the population as a probability vector (PV) overthe set of solutions and are operationally equivalent to theorder-one behaviour of the simple EA with uniformcrossover In this way a much smaller number of solutionsmust be stored in the memory thus significantly reducingthe memory consumption

22CoevolutionaryAlgorithm )eCoevolutionary Algorithm[21] makes multiple swarms simultaneously evolve andcommunicate with one another to improve the searchperformance Currently distributed coevolution is the mostpopular coevolving process which shares the search in-formation among multiple swarms through the populationmigration strategy During the searching process differentswarms have evolving strategies and configurations Tanet al [22] proposed to decompose the problemrsquos solution

vector into multiple swarms to evolve simultaneously Muand Liu [23] presented anM-elite Coevolutionary Algorithmthat applied different elite strategies in the coevolvingprocess )e elite centered swarm has the highest priorityand other swarms implemented the cooperative coevolvingprocess In [24] a parallel evolving mechanism was designedby dividing the population into three swarms that evolvedindependently However all the swarms use the sameevolving strategy and the swarmrsquos evolving process swarmwas relatively independent which decreased the algorithmrsquosexploration and exploitation ability More recently Wanget al [25] proposed a two-elite strategy which makes use ofthe differences between two elites to guide the wholeevolving process

Different from all the techniques mentioned above inthis work we propose a compact coevolutionary Algorithmto match the biomedical ontologies which combines theadvantages of the compact EA and coEvolutionary Algo-rithm to save the memory consumption and runtime andovercome the bias improvement of solutions

23 Preliminaries

231 Ontology Ontology Alignment and Ontology MatchingProcess In this work an ontology is defined as a quadrupleO (C P I A) where

(i) C is the class set ie the set of concepts thatpopulate the domain of interest

(ii) P is the property set ie the set of relations betweenthe concepts of domain

(iii) I is the instance set ie the set of objects in the realworld representing the instances of a concept and

(iv) A is the axiom set ie the statements that say whatis true about the modeled domain

An alignment A between two ontologies O1 and O2 isdefined as a set of correspondences and each correspondenceis a triple (e1 e2 n) where e1 and e2 are the entities in O1 andO2 respectively and n isin [0 1] is a confidence value holdingfor the correspondence between them In this work the re-lation existing between two ontology entities is the equiva-lence () )e ontology matching process can be defined asa function θ(O1 O2 p r) [26] where p is the parameter setand r is the resource set Ontology matching process returnsa new alignment AN between ontologies O1 and O2

232 Concept Similarity Measure Concept similaritymeasure is the foundation of biomedical ontology matching[27] In this work we utilize an asymmetrical conceptsimilarity measure to calculate the biomedical conceptsrsquosimilarity values First for each biomedical concept weconstruct a profile for it by collecting the label commentand property information such as label domain and rangefrom itself and all its direct descendants )en the similarityof two biomedical concepts c1 and c2 is measured based onthe similarity of their profiles p1 and p2 which can becalculated by the following two asymmetrical measures


sim1 p1 p2( 1113857 p1 capp2

11138681113868111386811138681113868111386811138681113868

p1

sim2 p1 p2( 1113857 p1 capp2

11138681113868111386811138681113868111386811138681113868

p2

(1)







11138681113868111386811138681113868

11138681113868111386811138681113868 + CO2minusMatch

11138681113868111386811138681113868

11138681113868111386811138681113868

CO1

11138681113868111386811138681113868

11138681113868111386811138681113868 + CO2

11138681113868111386811138681113868

11138681113868111386811138681113868ϵ[0 1] (2)

where


(ii) CO1and CO2





11138681113868111386811138681113868

11138681113868111386811138681113868 + CO2minusMatch

11138681113868111386811138681113868

11138681113868111386811138681113868


11138681113868111386811138681113868

11138681113868111386811138681113868ϵ[0 1] (3)

where







(4)



st X x1 x2 x O1| | x O1| |+11113874 1113875T

xi 1 2 x O2| |

x O1| |+1 isin [0 1]

(6)


















(5) i i + 1(6) if (i len)

(7) i 0(8) end if


ALGORITHM 1







Step 211 Crossover


Step 212 Mutation






Continued












4 Conclusion






ALGORITHM 2



Data Availability




Acknowledgments


References













































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in




sim1 p1 p2( 1113857 p1 capp2

11138681113868111386811138681113868111386811138681113868

p1

sim2 p1 p2( 1113857 p1 capp2

11138681113868111386811138681113868111386811138681113868

p2

(1)







11138681113868111386811138681113868

11138681113868111386811138681113868 + CO2minusMatch

11138681113868111386811138681113868

11138681113868111386811138681113868

CO1

11138681113868111386811138681113868

11138681113868111386811138681113868 + CO2

11138681113868111386811138681113868

11138681113868111386811138681113868ϵ[0 1] (2)

where


(ii) CO1and CO2





11138681113868111386811138681113868

11138681113868111386811138681113868 + CO2minusMatch

11138681113868111386811138681113868

11138681113868111386811138681113868


11138681113868111386811138681113868

11138681113868111386811138681113868ϵ[0 1] (3)

where







(4)



st X x1 x2 x O1| | x O1| |+11113874 1113875T

xi 1 2 x O2| |

x O1| |+1 isin [0 1]

(6)


















(5) i i + 1(6) if (i len)

(7) i 0(8) end if


ALGORITHM 1







Step 211 Crossover


Step 212 Mutation






Continued












4 Conclusion






ALGORITHM 2



Data Availability




Acknowledgments


References













































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in

















(5) i i + 1(6) if (i len)

(7) i 0(8) end if


ALGORITHM 1







Step 211 Crossover


Step 212 Mutation






Continued












4 Conclusion






ALGORITHM 2



Data Availability




Acknowledgments


References













































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in









Step 211 Crossover


Step 212 Mutation






Continued












4 Conclusion






ALGORITHM 2



Data Availability




Acknowledgments


References













































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in














4 Conclusion






ALGORITHM 2



Data Availability




Acknowledgments


References













































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in





Data Availability




Acknowledgments


References













































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in
































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in




UsingCompactCoevolutionaryAlgorithmforMatching...

Documents

Transcript of UsingCompactCoevolutionaryAlgorithmforMatching...