Fuzzy rule-based similarity model enables learning from small case bases
Graphical abstract
Highlights
► A fuzzy similarity model is proposed for case-based reasoning. ► Similarity degrees between cases are evaluated via fuzzy rule based reasoning. ► Fuzzy similarity rules can be learned from rather small case bases. ► The proposed method presents a new paradigm for relation-oriented learning.
Introduction
Case-based reasoning (CBR) presents an important cognitive methodology in Artificial Intelligence, which advocates the use of previous experiences to solve new problems [1]. A fundamental principle that underlies CBR is the hypothesis that similar problems have similar solutions. Hence a CBR system first retrieves cases in the case base that are similar to a query problem and then refines the solutions of the retrieved cases to tackle the new situation at hand.
Similarity assessment plays a key role in CBR in that it decides the quality of retrieved cases. A competent similarity model has to reflect the real utility/relevance of cases for solving new problems [2]. So far a wealth of similarity measures has been established for successful applications of CBR in various real-world scenarios. Cunningham [3] proposed a coherent taxonomy which organized the broad range of similarity mechanisms into the four categories (direct, information-based, transformation-based and emergent measures). The work of this paper belongs to the first category and aims to develop direct similarity models for cases with feature-value representation.
Our objective is to build the similarity model as a knowledge container to guide the CBR process [4]. Fuzzy if–then rules are adopted in this paper as the form of knowledge representation due to the following two reasons. First, fuzzy rules provide a flexible means to express the knowledge and criteria for similarity assessment. Second, fuzzy rule based systems are proved universal approximators [5], able to produce accurate reasoning results for similarity evaluation. The learning of fuzzy similarity rules is implemented by exploiting the case base. We consider the case base a valuable resource with hidden knowledge for similarity learning. A sample of similarity is created from a pair of known cases in which the vicinity of case solutions reflects the similarity of case problems. We do pair-wise comparisons of cases in the case base to derive adequate training examples for learning fuzzy similarity rules. The empirical studies have demonstrated that the proposed approach is capable of discovering fuzzy similarity knowledge from a very limited number of cases, giving rise to the competence of CBR systems to work on a small case library.
The paper is organized as follows. Section 2 discusses related works. Section 3 outlines a general CBR paradigm used in the paper. The fuzzy similarity model for case matching is addressed in Section 4. Then, in Section 5, we discuss the issue of how to learn these fuzzy similarity rules from the case base. In Section 6, we present experimental results for evaluation of the proposed method. Finally, concluding remarks are given in Section 7.
Section snippets
Related works
The issue of similarity has received much research attention from the CBR community. Plaza et al. [6] discussed the ways to exploit similarity information for explaining CBR results in classification tasks. They indicated that suitable explanation can be derived from building symbolic descriptions of similar aspects among cases. They also illustrated that symbolic descriptions of similarity can be utilized to support various steps (including retrieve, reuse, revise and retain) within a CBR
Case-based reasoning: a general paradigm
The general idea of the case-based approach is exploitation of information in the previous cases to solve a new problem. A general CBR paradigm used in this paper is shown in Fig. 1. It starts with similarity matching between a query problem and known cases in the case library. A properly defined similarity function has to be employed in this stage. As the evaluated similarity values reflect the utility or appropriateness of solutions of the known cases, they offer important information to be
Fuzzy rules based similarity model
This section explains how similarity matching between cases can be implemented by fuzzy rules based reasoning. We will start with discussing the benefits of fuzzy rules as similarity model in Section 4.1. Then we explain the general rule structure and fuzzy reasoning procedure employed for similarity assessment in Section 4.2.
Learning fuzzy rules from case bases
Supervised learning is performed in this paper to generate fuzzy rules for the similarity model. We need a “teacher” to specify samples of desired similarity values for various pairs of cases as training examples. The task of learning is to revise fuzzy similarity rules to reduce the discrepancy between the desired similarity scores as given by the “teacher” and the estimated similarity values produced by the (similarity) model. The general learning paradigm is shown in Fig. 3 in which a
Experimental evaluations
To evaluate the capability of the proposed method, we show in this section the experimental results on six well-known data sets from the UCI Machine Learning Repository [27]. All these data sets contain cases characterized by numerical features and discrete classes, with the numbers of features ranging from 4 to 13 and the numbers of classes between 2 and 6, as illustrated in Table 1. We used the classification accuracy based on small case bases as the criterion to evaluate the learning ability
Conclusion
This paper puts forward a new method of employing fuzzy rules as the representation of similarity models in CBR research. Fuzzy rules are considered a more powerful vehicle to accommodate rich domain knowledge than conventional similarity metrics. Fuzzy rule-based reasoning is conducted to estimate the degrees of similarity between cases in the case library and a new problem. Further we explain that fuzzy similarity rules can be generated by exploiting the information from a rather small case
Acknowledgement
The author would like to sincerely thank the anonymous referees for their valuable comments and suggestions for improvement of the paper.
References (32)
- et al.
Reduction of fuzzy control rules by means of precise learning – method and case study
Fuzzy Sets and Systems
(2002) Finding useful fuzzy concepts for pattern classification using genetic algorithm
Information Sciences
(2005)- et al.
Analysis of interpretability-accuracy tradeoff of fuzzy systems by multiobjective fuzzy genetics-based machine learning
International Journal of Approximate Reasoning
(2007) Retrieval, reuse, revision and retention in case-based reasoning
The Knowledge Engineering Review
(2005)- et al.
Utility-oriented matching: a new research direction for case-based reasoning
A taxonomy of similarity mechanisms for case-based reasoning
IEEE Transactions on Knowledge and Data Engineering
(2009)- et al.
Case-based reasoning foundations
The Knowledge Engineering Review
(2006) Mamdani fuzzy system: universal approximator to a class of random processes
IEEE Transactions on Fuzzy Systems
(2002)- et al.
The explanatory power of symbolic similarity in case-based reasoning
Artificial Intelligence Review
(2005) - et al.
Weighting features
Acquiring customer preferences from return-set selections
Improving recommendation ranking by learning personal feature weights
Using evolution programs to learn local similarity measures
Genetic algorithms to optimize CBR retrieval
Global optimization of feature weights and the number of neighbors that combine in a case-based reasoning system
Expert Systems
Supporting case-based retrieval by similarity skylines: basic concepts and extensions
Cited by (19)
kNN-P: A kNN classifier optimized by P systems
2020, Theoretical Computer ScienceCitation Excerpt :It is based on the representation of multiple choices of two key parameters of fuzzy NN: one is applied in the definition of the membership function, and the other is used in the computation of the voting rule. Xiong et al. [17,18] discuss the similarity modelling in case-based reasoning. Membrane computing is a class of distributed parallel computing models inspired from the structure and functioning of living cells as well as interaction of living cells in tissues and organs, known as P systems [19,20].
Integrating multi-granularity model and similarity measurement for transforming process data into different granularity knowledge
2018, Advanced Engineering InformaticsCitation Excerpt :In the last decades, most of the research developments are more focus on the stage of retrieval and similarity of problem pairs. As a consequence, many different methods have been developed such as feature-based similarity measure [10,17,18], improved cosine similarity measure [19], combining case indexing and similarity matrix based on the constructed ontology [20], part similarity measurement using ontology and multi-criteria decision making technique [21], similarity evaluation based on the fuzzy rules [22,23], the nearest neighbor retrieval [24] and artificial bee colony [25]. In addition, some cases can only be described by hierarchical tree structures [14].
Case-based reasoning classifier based on learning pseudo metric retrieval
2017, Expert Systems with ApplicationsCitation Excerpt :So far, various types of similarity measures have been proposed and used for CBR systems (Liao, Zhang, & Mount, 1998). Among that, some fuzzy rule-based similarity measurement methods were developed for problems with uncertainty attribute (Wang, & Qu, 2013; Xiong, 2013). Mitra, Kundu, and Pedrycz (2012) employed a structural similarity approach for feature selection and the merits of the proposed similarity measures have been verified in both classification and clustering tasks.
Application of Case-based Reasoning Methodology to Bingo Lottery Benchmark Model
2016, Procedia Computer ScienceTCBR-HMM: An HMM-based text classifier with a CBR system
2015, Applied Soft Computing JournalCitation Excerpt :The purpose of this hybrid model is to provide a solution to another kind of text classification problem where a previously created model needs to be adapted to each new document in an iterative learning frame. The CBR system [19–23] solves new problems by recalling and adapting previous solutions. Consequently, CBR is useful for tasks using predicates that are poorly defined, lead to inconsistent outcomes, and have incomplete rules to apply [24,23].
Neural network models for situation similarity assessment in hybrid-CBR
2023, Journal of Intelligent and Fuzzy Systems