A Multi-path Strategy for Hierarchical Ensemble Classification

Alshdaifat, Esra’a; Coenen, Frans; Dures, Keith

doi:10.1007/978-3-319-08979-9_16

Cited by 3 publications

(5 citation statements)

References 15 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…FastText was used for NLP to create a machine learning model to tag paragraphs. FastText is often on par with deep learning classifiers in terms of accuracy, and is many orders of magnitude faster for training and evaluation [ 36 , 37 ].…”

Section: Methodsmentioning

confidence: 99%

A Virtual Community for Disability Advocacy: Development of a Searchable Artificial Intelligence–Supported Platform

Morr¹,

Maret²,

Muhlenbach³

et al. 2021

JMIR Form Res

View full text Add to dashboard Cite

Background The lack of availability of disability data has been identified as a major challenge hindering continuous disability equity monitoring. It is important to develop a platform that enables searching for disability data to expose systemic discrimination and social exclusion, which increase vulnerability to inequitable social conditions. Objective Our project aims to create an accessible and multilingual pilot disability website that structures and integrates data about people with disabilities and provides data for national and international disability advocacy communities. The platform will be endowed with a document upload function with hybrid (automated and manual) paragraph tagging, while the querying function will involve an intelligent natural language search in the supported languages. Methods We have designed and implemented a virtual community platform using Wikibase, Semantic Web, machine learning, and web programming tools to enable disability communities to upload and search for disability documents. The platform data model is based on an ontology we have designed following the United Nations Convention on the Rights of Persons with Disabilities (CRPD). The virtual community facilitates the uploading and sharing of validated information, and supports disability rights advocacy by enabling dissemination of knowledge. Results Using health informatics and artificial intelligence techniques (namely Semantic Web, machine learning, and natural language processing techniques), we were able to develop a pilot virtual community that supports disability rights advocacy by facilitating uploading, sharing, and accessing disability data. The system consists of a website on top of a Wikibase (a Semantic Web–based datastore). The virtual community accepts 4 types of users: information producers, information consumers, validators, and administrators. The virtual community enables the uploading of documents, semiautomatic tagging of their paragraphs with meaningful keywords, and validation of the process before uploading the data to the disability Wikibase. Once uploaded, public users (information consumers) can perform a semantic search using an intelligent and multilingual search engine (QAnswer). Further enhancements of the platform are planned. Conclusions The platform ontology is flexible and can accommodate advocacy reports and disability policy and legislation from specific jurisdictions, which can be accessed in relation to the CRPD articles. The platform ontology can be expanded to fit international contexts. The virtual community supports information upload and search. Semiautomatic tagging and intelligent multilingual semantic search using natural language are enabled using artificial intelligence techniques, namely Semantic Web, machine learning, and natural language processing.

show abstract

Section: Methodsmentioning

confidence: 99%

A Virtual Community for Disability Advocacy: Development of a Searchable Artificial Intelligence–Supported Platform

Morr¹,

Maret²,

Muhlenbach³

et al. 2021

JMIR Form Res

View full text Add to dashboard Cite

show abstract

“…When considering hierarchical classification models the necessary class partitioning can be conducted using a variety of methods such as data splitting or clustering. The performance of the binary tree approach, the most commonly used hierarchical ensemble model, is significantly influenced by the adopted class partitioning method; inappropriate choices can result in poor performance (Alshdaifat, Coenen, & Dures, 2013a, 2013b, 2014. Other than the nature of the grouping method to be adopted, a second drawback of the binary tree based hierarchical ensemble model is that if a record is misclassified early on in the classification process (near the root of the hierarchy) it will continue to be misclassified at deeper levels; the so called "successive misclassification" problem.…”

Section: Literature Reviewmentioning

confidence: 99%

“…Several mechanisms can be adopted, such as: (i) applying some voting scheme and selecting the candidate class associated with the highest vote, or (iii) generating an accumulated weight for each candidate class and selecting the class associated with the highest accumulated weight. According to previous work conducted by the authors (Alshdaifat, Coenen, & Dures, 2013b, 2014the last strategy is likely to produce the best classification performance, thus it is adopted with respect to the work presented in this paper. Using this strategy we take into consideration all probability values in a followed path to produce an accumulated value.…”

Section: Algorithm 1 Rooted Dag Generationmentioning

confidence: 99%

“…Thissectionprovidesagenericoverviewofthehierarchicalensemblemethodologyforsolvingthe multi-classclassificationproblem.Thehierarchicalensemblemethodologyisarelativelyrecently proposedapproachtoaddressthemulti-classclassificationproblemwhichinvolvesthegeneration ofahierarchical"meta-algorithm" (Kumaretal.,2002;Madzarovetal.,2008).Acommonstructure adoptedforhierarchicalclassification,asnotedintheprevioussection,isabinarytreestructure constructed in either a bottom-up or top-down manner (Beygelzimer et al, 2007;Kumar et al, 2002).Inthetop-downapproach,therootnodecontainsthecompletesetofclasslabels{c 1 , c 2 , …, c n }.Startingfromtheroot,thesetofclasslabelsateachnodeisrecursivelysplit,andaclassifieris trainedtodistinguishbetweenthetwosubsets.Usingthebottom-upapproachamergingprocessis adoptedsimilartoagglomerativehierarchicalclustering.Thetwonodeswiththeclosestdistanceare mergedtoformanodedescribinganewmeta-class (Beygelzimeretal.,2007).Anexamplebinary treehierarchyispresentedinFigure1. Attheroot,wediscriminatebetweentwogroupsofclass labels{a, b, c}and{d, e}.Atthenextlevel,wedistinguishbetweensmallergroups,andsoon,till wereachnodeswithclassifiersthatcanassignasingleclasslabeltoagivenrecord. When considering hierarchical classification models the necessary class partitioning can be conductedusingavarietyofmethodssuchasdatasplittingorclustering.Theperformanceofthe binarytreeapproach,themostcommonlyusedhierarchicalensemblemodel,issignificantlyinfluenced by the adopted class partitioning method; inappropriate choices can result in poor performance (Alshdaifat,Coenen,&Dures,2013a,2013b,2014.Otherthanthenatureofthegroupingmethod tobeadopted,aseconddrawbackofthebinarytreebasedhierarchicalensemblemodelisthatifa recordismisclassifiedearlyonintheclassificationprocess(neartherootofthehierarchy)itwill continuetobemisclassifiedatdeeperlevels;thesocalled"successivemisclassification"problem. Inpreviousworktheauthorshavesuggestedamultiple-pathstrategy,whichallowsformorethan onepathtobefollowedwithinthebinarytreeduringtheclassificationstage.Thismultiple-path strategywasfacilitatedbytheuseofclassifiers,suchasNaiveBayesorClassificationAssociation RuleMining(CARM),whichfeatureprobabilityorconfidencevaluesthatcanbeusedtodetermine whereonepathshouldbefollowedandwheretwopathsshouldbefollowed.However,themulti-path strategyonlypartiallyresolvesthesuccessivemisclassificationproblem,fundamentallythebinary treestructureisnotsufficientlyexpressivetocapturethenatureofmulti-classclassification.…”

Section: Literature Reviewmentioning

confidence: 99%

“…TheMultiple-Pathstrategyisdesignedtoaddressthesuccessivemisclassificationissue,discussed earlier,thatisassociatedwithhierarchicalclassification.Inthemultiple-pathstrategymorethanone pathcanbefollowedwithintheDAGclassificationmodel.Morespecifically,theBayesianprobability Passociatedwithindividualclassgroupswillbeusedtodictatewhetheroneormorepathswillbe followed,ateachnode,accordingtoapredefinedthresholdsigmaσ(0≤σ<1).Althoughmanypaths canbefollowedateachDAGnode,onlytwopathsaresuggestedasamaximum,ateachDAGnode, sothatcomparisonscanbemadewiththebinarytreehierarchicalensemblemodel(whereonlya maximumoftwopathscanbefollowedateachtreenode).Asecondreasonistolimitthecomplexity oftheproposedDAGmodel,theneedforthiswillbecomeclearlaterinthispaperintheevaluation sectionwheretheclassificationtimeisreportedforsingleandmultiplepathstrategies.Anissue associatedwiththesuggestedmultiple-pathstrategyishowtodecidethefinalclasslabelfromthe collectionof"candidateclasses"resultingfromfollowingmultiplepaths.Severalmechanismscan beadopted,suchas:(i)applyingsomevotingschemeandselectingthecandidateclassassociated withthehighestvote,or(iii)generatinganaccumulatedweightforeachcandidateclassandselecting theclassassociatedwiththehighestaccumulatedweight.Accordingtopreviousworkconducted bytheauthors (Alshdaifat,Coenen,&Dures,2013b,2014…”

Section: Multiple Paths Strategymentioning

confidence: 99%

See 2 more Smart Citations

A Directed Acyclic Graph (DAG) Ensemble Classification Model

Alshdaifat

Coenen

Dures

2017

International Journal of Data Warehousing and Mining

Self Cite

View full text Add to dashboard Cite

In this paper a hierarchical ensemble classification approach, that utilizes a Directed Acyclic Graph (DAG) structure, is proposed as a solution to the multi-class classification problem. Two main DAG structures are considered: (i) rooted DAG, and (ii) non-rooted DAG. The main challenges that are considered in this paper are: (i) the successive misclassification issue associated with hierarchical classification, and (i) identification of the starting node within the non-rooted DAG approach. To address these issues the idea is to utilize Bayesian probability values to: select the best starting DAG node, and to dictate whether single or multiple paths should be followed within the DAG structure. The reported experimental results indicated that the proposed DAG structure is more effective than when using a simple binary tree structure for generating a hierarchical classification model.

show abstract