從最佳化觀點推導多評準分類規則— 以生物及醫療資訊為例

Abstract

從資料中推導出關鍵的分類規則, 是科學研究的重要任務之一。一條有用的分類規則, 除其是最適外, 應同時滿足三項評準: 高正確度、高支持度、高精簡度。然而,目前的分類方法, 諸如約略集合理論、類神經網路、分類樹. . .等, 都只能推導得可行解規則, 而非最適規則。此外, 目前的方法推導得的規則只能同時滿足前述三項評準之一。本研究提出一個多評準的模式, 用以在較好的正確度、支持度及精簡度下, 推導得最適分類規則, 其是透過混合0-1線性多目標規化模型以推導分類規則。並以一些實際的生物及醫療資料進行測試, 其結果顯示所提方法能比目前方法推導得較佳的分類規則。

To induce critical classification rules from observed data is a major task in biological and medical research. A classification rule is considered to be useful if it is optimal and simultaneously satisfies three criteria: is highly accurate, has a high rate of support, and is highly compact. However, existed classification methods, such as rough set theory, neural networks, ID3, etc., may only induce feasible rules instead of optimal rules. In addition, the rules found by existed methods may only satisfy one of the three criteria. This study proposes a multi-criteria model to induce optimal classification rules with better rates of accuracy, support and compactness. A linear multiobjective programming model for inducing classification rules is formulated. Two practical data sets, one of HSV patients results and another of European barn swallows, are tested. The results illustrate that the proposed method can induce better rules than existed methods.