The knowledge-leverage-based Takagi--Sugeno--Kang fuzzy system (KL-TSK-FS) modeling method has shown promising performance for fuzzy modeling tasks where transfer learning is required. However, the knowledge-leverage mechanism of the KL-TSK-FS can be further improved. This is because available training data in the target domain are not utilized for the learning of antecedents and the knowledge transfer mechanism from a source domain to the target domain is still too simple for the learning of consequents when a Takagi--Sugeno--Kang fuzzy system (TSK-FS) model is trained in the target domain. The proposed method, that is, the enhanced KL-TSK-FS (EKL-TSK-FS), has two knowledge-leverage strategies for enhancing the parameter learning of the TSK-FS model for the target domain using available information from the source domain. One strategy is used for the learning of antecedent parameters, while the other is for consequent parameters. It is demonstrated that the proposed EKL-TSK-FS has higher transfer learning abilities than the KL-TSK-FS. In addition, the EKL-TSK-FS has been further extended for the scene of the multisource domain.

This paper discusses parallel distributed genetics-based machine learning (GBML) of fuzzy rule-based classifiers from multiple data sets. We assume that each data set has a similar but different set of attributes. In other words, each data set has different missing attributes. Our task is the design of a fuzzy rule-based classifier from those data sets. In this paper, we first show that fuzzy rules can handle missing attributes easily. Next we explain how parallel distributed fuzzy GBML can handle multiple data sets with different missing attributes. Then we examine the accuracy of obtained fuzzy rule-based classifiers from various settings of available training data such as a single data set with no missing attribute and multiple data sets with many missing attributes. Experimental results show that the use of multiple data sets often increases the accuracy of obtained fuzzy rule-based classifiers even when they have missing attributes. We also discuss the learning from a data set under a severe privacy preserving policy where only the error rate of each candidate classifier is available. It is assumed that no information about each individual pattern is available. This means that we cannot use any information on the class label or the attribute values of each pattern. We explain how such a black-box data set can be utilized for classifier design.

When we generate association rules, continuous attributes have to be discretized into intervals while our knowledge representation is not always based on such discretization. For example, we usually use some linguistic terms (e.g., young, middle age, and old) for dividing our ages into some fuzzy categories. We describe the extraction of linguistic association rules and examine the performance of extracted rules. First we modify the definitions of the two basic measures (i.e., confidence and support) of association rules for extracting linguistic association rules. The main difference between standard and linguistic association rules is the discretization of continuous attributes. We divide the domain interval of each attribute into some fuzzy regions (i.e., linguistic terms) when we extract linguistic association rules. Next, we compare fuzzy discretization with standard non-fuzzy discretization through computer simulations on a pattern classification problem with many continuous attributes. The classification performance of extracted rules on unseen test patterns is examined under various conditions. Simulation results show that linguistic association rules with rule weights have high generalization ability even when the domain of each continuous attribute is homogeneously partitioned.