Repository logo
 

Publication:
Agrupamento e agregação de medidas objetivas para ranqueamento de regras em classificadores associativos

Loading...
Thumbnail Image

Advisor

Carvalho, Veronica Oliveira de

Coadvisor

Graduate program

Ciência da Computação - FC/FCT/IBILCE/IGCE 33004153073P2

Undergraduate course

Journal Title

Journal ISSN

Volume Title

Publisher

Universidade Estadual Paulista (Unesp)

Type

Master's thesis

Access right

Acesso abertoAcesso Aberto

Abstract

Abstract (portuguese)

Em diversos problemas é interessante o uso de algoritmos inerentemente interpretáveis, uma vez que facilita o entendimento do conhecimento obtido e das predições realizadas. Dentre estes algoritmos encontram-se os classificadores associativos (CAs). Os mesmos mesclam as tarefas de associação e classificação e são, portanto, induzidos em etapas. No que se refere a etapa de ordenação, a mesma é, em geral, realizada via o uso de medidas objetivas (MOs), utilizadas, entre outros propósitos, para ranquear as regras pelo seu grau de importância. Em geral, o método CSC (Confiança, Suporte, Cardinalidade) é o utilizado para se ordenar as regras, tendo como base as MOs padrão de regras de associação. Contudo, ao longo dos anos inúmeras MOs foram propostas, visando superar limitações de algumas das medidas até então existentes. Este grande número de MOs gerou um problema secundário, uma vez que não existe uma medida que seja adequada a todas as explorações. Neste contexto, novas propostas surgiram visando modificar a etapa de ranqueamento dos CAs, seja por meio: (i) de novas medidas e/ou medidas já existentes em outros contextos, mas aplicadas nesta etapa de maneira singular; (ii) da união (agregação) de medidas existentes de maneira a usá-las simultaneamente. A estratégia (ii), adotada neste trabalho, traz a vantagem de diminuir a necessidade da escolha de uma só medida, considerando também diferentes aspectos (semântica) para a ordenação das regras. Trabalhos recentes nesta linha vêm sendo propostos, os quais têm apresentado resultados promissores. Contudo, os mesmos apresentam problemas em relação ao desempenho e/ou a interpretabilidade dos modelos gerados. Em todos eles é possível verificar uma relação inversa entre desempenho e interpretabilidade, i.e., quando o desempenho do modelo é alto, a interpretabilidade é baixa (e vice-versa). Diante do exposto, este trabalho apresenta um método de ranqueamento via agregação de MOs, denominado AC.Rankₐ a ser incorporado a fluxos de indução de CAs, visando a obtenção de modelos que apresentem um melhor equilíbrio entre desempenho e interpretabilidade. Para que o AC.Rankₐ funcione, o mesmo deve ser instanciado com um conjunto de MOs e um método de agregação. Os conjuntos de MOs explorados foram tanto extraídos da literatura quanto propostos neste trabalho. Já os métodos de agregação foram extraídos tanto de abordagens de agregação de rankings como de análise de decisão multicritérios. O método proposto foi avaliado em diversos fluxos de indução de CAs em relação a desempenho, medido via F1-Macro, e interpretabilidade, medida via tamanho do modelo. A análise foi realizada comparando-se diversos fluxos de indução quando o ranqueamento acontece via CSC (baseline) e via AC.Rankₐ. Os resultados obtidos demonstram que o AC.Rankₐ consegue manter o desempenho dos modelos, porém com uma melhor interpretabilidade, diferentemente dos trabalhos da literatura.

Abstract (english)

In several problems the use of inherently interpretable algorithms is interesting, as it facilitates the understanding of the obtained knowledge and the predictions made. Among these algorithms are the associative classifiers (ACs). They merge the tasks of association and classification and are, therefore, induced in stages. Regarding the ranking step, it is, in general, carried out through the use of objective measures (OMs), used, among other purposes, to rank the rules according to their level of importance. In general, the CSC (Confidence, Support, Cardinality) method is used to sort the rules, based on the standard OMs of association rules. However, over the years, many OMs have been proposed, aiming to overcome limitations of some of the previously existing measures. This large number of OMs led to a secondary problem, since there is no measure that is suitable for all explorations. In this context, new proposals have emerged aiming to modify the ACs ranking step, either through: (i) new measures and/or measures that already exist in other contexts, but applied at this step in a unique way; (ii) the union (aggregation) of existing measures in order to use them simultaneously. Strategy (ii), adopted in this work, has the advantage of reducing the need to choose a single measure, also considering different aspects (semantics) for ranking the rules. Recent works in this line have been proposed, which have shown promising results. However, they present problems in relation to the performance and/or interpretability of the generated models. In all of them it is possible to verify an inverse relationship between performance and interpretability, i.e., when model performance is high, interpretability is low (and vice versa). Considering the above, this work presents a ranking method via aggregation of OMs, named AC.Rankₐ, to be incorporated into ACs induction flows, aiming to obtain models that present a better balance between performance and interpretability. For AC.Rankₐ to work, it must be instantiated with a set of OMs and an aggregation method. The sets of OMs explored were both extracted from the literature and proposed in this work. The aggregation methods were extracted from both ranking aggregation and multi-criteria decision analysis approaches. The proposed method was evaluated in several ACs induction flows in terms of performance, measured via F1-Macro, and interpretability, measured via model size. The analysis was carried out by comparing different induction flows when ranking takes place via CSC (baseline) and via AC.Rankₐ. The results obtained demonstrate that AC.Rankₐ can maintain the performance of the models, but with better interpretability, unlike works in the literature.

Description

Keywords

Classificadores associativos, Medidas objetivas, Agregação de medidas objetivas, Ranqueamento de regras, Agrupamento de medidas objetivas, Associative classifiers, Objective measures, Aggregation of objective measures, Rules ranking, Clustering of objective measures

Language

Portuguese

Citation

Related itens

Units

Departments

Undergraduate courses

Graduate programs