Powered by OpenAIRE graph
Found an issue? Give us feedback
image/svg+xml Jakob Voss, based on art designer at PLoS, modified by Wikipedia users Nina and Beao Closed Access logo, derived from PLoS Open Access logo. This version with transparent background. http://commons.wikimedia.org/wiki/File:Closed_Access_logo_transparent.svg Jakob Voss, based on art designer at PLoS, modified by Wikipedia users Nina and Beao Universidade de Lisb...arrow_drop_down
image/svg+xml Jakob Voss, based on art designer at PLoS, modified by Wikipedia users Nina and Beao Closed Access logo, derived from PLoS Open Access logo. This version with transparent background. http://commons.wikimedia.org/wiki/File:Closed_Access_logo_transparent.svg Jakob Voss, based on art designer at PLoS, modified by Wikipedia users Nina and Beao
image/svg+xml Jakob Voss, based on art designer at PLoS, modified by Wikipedia users Nina and Beao Closed Access logo, derived from PLoS Open Access logo. This version with transparent background. http://commons.wikimedia.org/wiki/File:Closed_Access_logo_transparent.svg Jakob Voss, based on art designer at PLoS, modified by Wikipedia users Nina and Beao
UTL Repository
Doctoral thesis . 2012
Data sources: UTL Repository
versions View all 2 versions
addClaim

Tratamento de dados omissos e métodos de imputação em classificação

Authors: Lorga Da Silva, Ana;

Tratamento de dados omissos e métodos de imputação em classificação

Abstract

Le but de ce travail est d'étudier l’effet des données manquantes en classification de variables, principalement en classification hiérarchique ascendante, et aussi en classification non hiérarchique (ou partitionnement). L'étude est effectuée en considérant les facteurs suivants: pourcentage de données manquantes, méthodes d'imputation, coefficients de ressemblance et critères de classification. On suppose que les données manquantes sont du type MAR («missing at random») données manquantes au hasard, mais pas. complètement au hasard.. Les données manquantes satisfont un schéma majoritairement monotone. Nous avons utilisé comme techniques sans imputation les méthodes lisîwise et pairwise et comme méthodes d'imputation simple: l'algorithme EM, le modèle de régression OLS, l’algorithme NIPALS et une méthode de régression PLS., Comme méthodes d'imputation multiple nous avons adopté une méthode basée sur le modèle de régression OLS associé à des techniques bayesiennes; on a aussi proposé un nouveau modèle d'imputation multiple basé sur les méthodes de régression PLS. Pour combiner les structures de classification résultant des méthodes d'imputation multiple nous avons proposé une combinaison par la moyenne des matrices de similarité et deux méthodes de consensus. Nous avons utilisé comme méthodes de classification hiérarchique des méthodes classiques et probabilistes, ces dernières basées sur la famille de méthodes VL (Vraisemblance du Lien), comme méthodes de classification hiérarchique, le saut minimal, le saut maximal, la moyenne parmi les groupes et aussi les AVL et AVB; pour les matrices de ressemblance, le coefficient d'affinité basique (pour les données continues) - qui correspond à l'indice d'Ochiai; pour les données binaires, le coefficient de corrélation de Bravais-Pearson et l'approximation probabiliste du coefficient d'affinité centré et réduit par la méthode-W. L'étude est basée principalement sur des données simulées et complétée par des applications à des données réelles. Nous avons travaillé sur des données continues et binaires. Le coefficient de Spearman est utilisé pour comparer les structures hiérarchiques obtenues sur des matrices complètes avec les structures obtenues à partir des matrices ; où les données sont «effacées» puis imputées. L'indice de Rand est utilisé pour comparer les structures non hiérarchiques. Enfin, nous avons aussi proposé une méthode non hiérarchique qui «s'adapte» aux données manquantes. Sur un cas réel la méthode de Ward est utilisée dans les mêmes conditions que pour les simulations; mais aussi sans satisfaire un schéma monotone; une méthode de Monte Carlo par chaînes de Markov sert pour l'imputation multiple.

Neste trabalho, pretende-se estudar o efeito dos dados omissos em classificação de variáveis, principalmente em classificação hierárquica ascendente, de acordo com.òs seguintes factores: percentagens de dados omissos, métodos de imputação, coeficientes de semelhança-e métodos de classificação. Supõe-se que os dados omissos são do tipo MAR ("missing at random"), isto é, a presença de dados omissos não depende dos valores omissos, nem das variáveis com dados omissos, mas depende de valores observados sobre outras variáveis da matriz de dados. Os dados omissos satisfazem um padrão maioritariamente monótono. Utilizaram-se as técnicas, em presença de dados omissos "listwise" e "pairwise"; como métodos de imputação simples: o algoritmo EM, o modelo de regressão OLS, o algoritmo MPALS e um método de regressão PLS. Como métodos de imputação múltipla, adoptou-se um método baseado sobre o modelo de regressão OLS associado a técnicas bayesianas; propôs-se também um novo método de imputação múltipla baseado sobre os métodos de regressão PLS. Como métodos de classificação hierárquica utilizaram-se classificações clássicas e probabilísticas, estas últimas baseadas na família de métodos VL (validade da ligação). Os métodos de classificação hierárquica utilizados foram, "single", "complete" e "average" "linkage", AVL e AYB. Para as matrizes de semelhança utilizou-se o coeficiente de afinidade básico (para dados contínuos) - que corresponde ao índice d'Ochiai para dados binários; o coeficiente de correlação de Pearson e a aproximação probabilística do coeficiente de afinidade centrado e reduzido pelo método-W. O estudo foi baseado em dados simulados e reais. Utilizou-se o coeficiente de Spearman, para comparar as estruturas de classificação hierárquicas e para as classificações não hierárquicas o índice de Rand.

In this work we aimed to study the effect of missing data in classification of variables; mainly in ascending hierarchical classification, according to the following factors: amount of missing data, imputation techniques, similarity coefficient and classification-criterion. We used as techniques in presence of missing data, listwise and pairwise; as simple imputation methods, an EM algorithm, the OLS regression method, the NIPALS algorithm and a PLS regression method. As multiple imputation, we used a method based on the OLS regression and a new one based on PLS, combined by the mean value of the similarity matrices and an ordinal consensus. As hierarchical methods we used classical and. probabilistic approaches, the latter based on the VL-family. The hierarchical methods used were single, complete and average linkage, AVL and AVB. For the similarity matrices we used the basic affinity coefficient (for continuous data) - that corresponds to the Ochiai index for binary data; the Pearson's correlation coefficient and the probabilistic approach of the affinity coefficient, centered and reduced by the W-method.. The study was based mainly on simulated data, complemented by real ones. We used the Spearman.coefficient between the associated ultrametrics to compare the structures of the hierarchical classifications and, for the non hierarchical classifications, the Rand's index.

Doutoramento em Matemática Aplicada à Economia e à Gestão

Country
Portugal
Keywords

Missing Data, Imputation Methods, Méthodes d'imputation, Données manquantes, Classification

  • BIP!
    Impact byBIP!
    selected citations
    These citations are derived from selected sources.
    This is an alternative to the "Influence" indicator, which also reflects the overall/total impact of an article in the research community at large, based on the underlying citation network (diachronically).
    0
    popularity
    This indicator reflects the "current" impact/attention (the "hype") of an article in the research community at large, based on the underlying citation network.
    Average
    influence
    This indicator reflects the overall/total impact of an article in the research community at large, based on the underlying citation network (diachronically).
    Average
    impulse
    This indicator reflects the initial momentum of an article directly after its publication, based on the underlying citation network.
    Average
    OpenAIRE UsageCounts
    Usage byUsageCounts
    visibility views 86
    download downloads 11
  • 86
    views
    11
    downloads
    Powered byOpenAIRE UsageCounts
Powered by OpenAIRE graph
Found an issue? Give us feedback
visibility
download
selected citations
These citations are derived from selected sources.
This is an alternative to the "Influence" indicator, which also reflects the overall/total impact of an article in the research community at large, based on the underlying citation network (diachronically).
BIP!Citations provided by BIP!
popularity
This indicator reflects the "current" impact/attention (the "hype") of an article in the research community at large, based on the underlying citation network.
BIP!Popularity provided by BIP!
influence
This indicator reflects the overall/total impact of an article in the research community at large, based on the underlying citation network (diachronically).
BIP!Influence provided by BIP!
impulse
This indicator reflects the initial momentum of an article directly after its publication, based on the underlying citation network.
BIP!Impulse provided by BIP!
views
OpenAIRE UsageCountsViews provided by UsageCounts
downloads
OpenAIRE UsageCountsDownloads provided by UsageCounts
0
Average
Average
Average
86
11
Upload OA version
Are you the author of this publication? Upload your Open Access version to Zenodo!
It’s fast and easy, just two clicks!