事例間の類似性に基づく負例の生成によるデータベースからの知識獲得

概要

論文の詳細を見る
This paper presents a method of generating negative instances based on similarity between positive instances for inductive learning systems, which construct the concept descriptions from positive/negative instances of a target concept. Given only positive instances, such systems may face a problem of overgeneralization of the resulting descriptions. This problem prevents existing useful inductive learning systems from being applied to KDD(Knowledge Discovery in Database), since most databases consist of only positive instances. To solve this problem, We generate negative instances from positives instances. We concentrate on similarity between positive instances, and construct the "approximated space" of the target concept, which contains all positive instances and instances close to them on the basis of the similarity. Then the instances which are not in the approximated space but are most similar to the positive instances are generated as the negative instances of the target concept. Such negative instances behave as near-miss instances and protect the over-generalization.
社団法人人工知能学会の論文
2000-09-01