Using Class-Based Reasoning to Improve the Accuracy of Symbolic Rules in a Hybrid Possibilistic Approach

A common strategy used in rule inductive algorithms is to assign an unseen example, not covered by any rule, to a static default class fixed at the inductive time and not updated thereafter. This paper presents a rule-based system using a Hybrid Possibilistic Inference Mechanism, which combines a Possibilistic Rule-based with a Class-based Reasoning. The inference process gives pre-eminence to Possibilistic Rule-based Reasoning, which selects the most suitable rule used to reach a conclusion in response to input facts. The proposed approach encodes relationship dependencies existing between facts and rules through Possibilistic Networks and quantifies these relationships by means of two measures: possibility and necessity. If the Possibilistic Rule-based Reasoning is blocked due the lack of satisfied rules, the Hybrid Possibilistic Inference Mechanism favours the Possibilistic Class-based Reasoning, which is the main contribution of this paper as it dynamically assigns a default class to each specific fact base not covered by any rule. To do so, we use a possibilistic network which searches for the most plausible class by quantifying relationship between facts and classes through a distance measure. Experimentation results demonstrate that the hybrid approach leads to accuracy improvement of the system.