Coverage-based resampling: Building robust consolidated decision trees

Ibarguren, Igor; Pérez, Jesús M.; Muguerza, Javier; Gurrutxaga, Ibai; Arbelaitz, Olatz

doi:10.1016/j.knosys.2014.12.023

Cited by 37 publications

(38 citation statements)

References 58 publications

(80 reference statements)

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…The use of the evolutionary approach enabled them to choose the most criticalsamples for the classifier learning step.Extensive volume of datasets were conveyed for examinations, through this high efficiency had achieved.Complexity of the classification increases and difficult to implement in real time platform are the disadvantages in this model. Jesus M.et al [22] presented a resampling approach for classificationalgorithms that use numerous subsamples.This procedure had beenapplied to CTC (Consolidated Tree Construction) algorithm over various classification contexts. A robust classification algorithm should not justbe able to rank in the top positions for certain classification problems.…”

Section: Related Workmentioning

confidence: 99%

Efficient DANNLO Classifier for Multi-class Imbalanced Data on Hadoop

2017

IJMTER

View full text Add to dashboard Cite

Abstract--In recent years, multi-class imbalance data classification is a major problem in big data. In such situations, we focused on developing a new Deep Artificial Neural Network Learning Optimization (DANNLO) Classifier for large collection of imbalanced data.In our proposed work, first the dataset reduction using principal component analysis for dimensionality reduction and initial centroid is computed.Then, parallel hierarchical Pillar k-means clustering algorithm based on MapReduceis usedto partitioning of an imbalanced data set into similar subset, which can improve the computational cost. The resultant clusters are given as input to the deep ANN for learning. In the next stage, deep neural network has been trained using the back propagation algorithm. In order to optimize the n-dimensional weight space, firefly optimization algorithm is used. Attractiveness and distance of each firefly is computed. Hadoop is used to handle these large volumes of variable size data. Imbalanced datasets is taken from ECDC (European Centre for Disease Prevention and Control) repository.The experimental results illustrated that the proposed method can significantly improve the effectiveness in classifying imbalanced data based on TP rate, F-measure, G-mean measures, confusion matrix, precision, recall, and ROC. The experimental results suggests that DANNLO classifier exceed other ordinary classifiers such as SVM and Random forest classifier on tested imbalanced data sets.

show abstract

Section: Related Workmentioning

confidence: 99%

Efficient DANNLO Classifier for Multi-class Imbalanced Data on Hadoop

2017

IJMTER

View full text Add to dashboard Cite

show abstract

“…Berriki, estrategia bat proposatu dugu eragozpen honi aurre egiteko, estalduran oinarritutako laginketa [16], hain zuzen. Estrategia honek, sortutako lagin multzoan dagoen jatorrizko laginaren kasuen ordezkaritza edo estaldura gutxieneko bat bermatzen duen lagin kopurua estimatzen du.…”

Section: Sarreraunclassified

“…Klase anitzeko datu-basetan (2 baino gehiagokoetan), klaseak orekatzea izango da onartzen den aukera bakarra. Mota honetako azpilaginak hainbat lanetan erabili dira zuhaitz kontsolidatuak sortzeko; kasu batzuetan, gutxiengoaren klasearen proportzio berri bat erabilita (desberdina datu-basearen arabera [13] edo beti %50a, hau da, lagin orekatuak, berriki argitaratutako lanean [16]) eta, beste batzuetan, klase-banaketa balio desberdin zabal bat erabilita [18,19], gutxiengoaren %2-tik %98-ra, honek duen eragina emaitzetan aztertzeko.…”

Section: Inplementatutako Laginketakunclassified

“…Sarreran jada esan den bezala, estalduran (ingelesez coveragean) oinarritutako laginketa berriki proposatua izan zen [16] kalkulatzeko zenbat lagin erabili behar ziren sortutako lagin multzoan estatistikoki bermatzeko jatorrizko laginaren kasuen portzentaje minimo bat nolabait jasota geratzen zela. Estimazio hau, jakina, laginketa mota bakoitzerako desberdina izango da eta lortu nahi den estaldura mailaren araberakoa.…”

Section: Estalduran Oinarritutako Laginketaunclassified

“…Parametro hauen guztien balio lehenetsiak orekatuak diren laginak (tamaina maximokoak eta estaldura maila %99koa) sortzeko ezarri dira, haiek izan baitira eskaini dizkiguten emaitzarik hoberenak azkeneko lanetan, [16], gehien bat klase desorekatuetako problemetan arreta jarrita.…”

Section: Igor Ibarguren Jesús M Pérez Javier Muguerzaunclassified

See 2 more Smart Citations

J48Consolidated WEKA paketea, adibide ezohikoen patroiak identifikatzeko tresna

Ibarguren

Pérez

Muguerza

2016

EKAIA

Self Cite

View full text Add to dashboard Cite

Laburpena: Artikulu honetan WEKA ikasketa automatikorako tresnarako CTC algoritmoaren inplementazioa aurkeztuko da: J48Consolidated paketea. CTC algoritmoak lagin multzo bat sortzen du eta lagin guztietan dagoen ezagutza kontuan hartuta sailkapen zuhaitz bakarra eraikitzeko gai da, kasu berrien sailkapenaren azalpena galdu gabe. Gainera, lan honetan J48Consolidated paketeak sortutako zuhaitzen emaitzak aztertuko dira, errealitateko 36 sailkapen problematarako, laginketa mota desberdinetan oinarrituta, eta jatorrizko laginaren estaldura-maila desberdinekin. Emaitzek erakusten dute estaldura-maila altuek orokorrean sailkatzeko gaitasuna handitzen dutela eta %75eko lagin estratifikatuak erabiltzea dela problema hauetan aukerarik lehiakorrena. Hitz-gakoak:Ikasketa automatikoa, Sailkapen gainbegiratua, Zuhaitz kontsolidatuak, CTC algoritmoa, Estalduran oinarritutako laginketa, J48Consolidated, WEKA.Abstract: This article presents the implementation of the CTC algorithm for the WEKA machine learning tool, the J48Consolidated package. The CTC algorithm creates a set of samples and taking the knowledge of all samples into account, is able to build a single classification tree, keeping the explanation of how new examples are classified. In addition, this work analyzes the results achieved by trees built by J48Con-solidated on 36 real world problems, using multiple sampling strategies and with different coverage values of the original sample. Results show that higher coverage values increase discriminating capacity and using stratified subsamples reduced to a 75% of the size give the most competitive results on these datasets.

show abstract