ABSTRACT Scores of machine learning algorithms depend on a good distance measure to achieve high performance. The inverted specific-class distance measure, simply ISCDM, is proposed to find reasonable distance measure between each pair of instances with nominal attributes only. ISCDM does not depend on the attribute value of the training instance, which makes it less sensitive to missing values in the training set and more robust to non-class attribute noise. However, in ISCDM, all attributes are assumed to be fully independent. It is obvious that the attribute independence assumption in ISCDM is rarely true in reality, which would harm its performance in the applications with complex attribute dependencies. In this paper, we single out an improved inverted specific-class distance measure by relaxing its unrealistic attribute independence assumption. We call it averaged one-dependence inverted specific-class distance measure, simply AODISCDM. We experimentally tested it on 29 classification problems from the University of California at Irvine (UCI) repository and found that it significantly outperforms ISCDM in terms of the negative conditional log likelihood (-CLL) and the root relative squared error (RRSE). Besides, the proposed AODISCDM maintains the computational simplicity (no search involved) and robustness that characterise ISCDM.

中文翻译：

---

名义属性的平均单依赖倒置特定类距离度量

摘要 机器学习算法的分数取决于良好的距离度量以实现高性能。提出了反向特定类距离度量，简称 ISCDM，用于在每对仅具有名义属性的实例之间找到合理的距离度量。ISCDM 不依赖于训练实例的属性值，这使得它对训练集中的缺失值不那么敏感，对非类属性噪声更加鲁棒。但是，在 ISCDM 中，假定所有属性都是完全独立的。很明显，ISCDM 中的属性独立假设在现实中很少是正确的，这会损害其在具有复杂属性依赖关系的应用程序中的性能。在本文中，我们通过放宽其不切实际的属性独立假设来挑选出改进的反向特定类距离度量。我们称其为平均单依赖倒置特定类距离度量，简称为 AODISCDM。我们在来自加州大学欧文分校 (UCI) 存储库的 29 个分类问题上对其进行了实验测试，发现它在负条件对数似然 (-CLL) 和相对平方误差 (RRSE) 方面明显优于 ISCDM。此外，所提出的 AODISCDM 保持了 ISCDM 特征的计算简单性（不涉及搜索）和鲁棒性。我们在来自加州大学欧文分校 (UCI) 存储库的 29 个分类问题上对其进行了实验测试，发现它在负条件对数似然 (-CLL) 和相对平方误差 (RRSE) 方面明显优于 ISCDM。此外，所提出的 AODISCDM 保持了 ISCDM 特征的计算简单性（不涉及搜索）和鲁棒性。我们在来自加州大学欧文分校 (UCI) 存储库的 29 个分类问题上对其进行了实验测试，发现它在负条件对数似然 (-CLL) 和相对平方误差 (RRSE) 方面明显优于 ISCDM。此外，所提出的 AODISCDM 保持了 ISCDM 特征的计算简单性（不涉及搜索）和鲁棒性。