Towards the Personalization of Algorithms Evaluation in Data Mining
Towards the Personalization of Algorithms Evaluation in Data Mining
复制标题
DOI：
--
复制
发表时间：
1998-08
期刊：
影响因子：
--
通讯作者：
G. Nakhaeizadeh;A. Schnabl
中科院分区：
文献类型：
--
作者：
G. Nakhaeizadeh;A. Schnabl
与统计学中的模型选择一样，选择合适的数据挖掘算法（DM-Algorithms）是知识发现过程中的一项重要任务。由于这一事实，有必要有复杂的指标，可以用作比较，以评估替代dm算法。文献表明，数据包络分析（DEA）是开发多标准评估指标的合适平台，可以考虑-与单标准指标相反- dm算法的所有积极和消极性质。我们讨论了DEA的不同扩展，这些扩展可以考虑dm算法的定性性质和在开发评估指标时考虑用户偏好。这些结果在统计学和机器学习中关于模型选择的一般性辩论中开启了新的讨论。
Like model selection in statistics, the choice of appropriate Data Mining Algorithms (DM-Algorithms) is a very important task in the process of Knowledge Discovery. Due to this fact it is necessary to have sophisticated metrics that can be used as comparators to evaluate alternative DM-algorithms. It has been shown in literature, that Data Envelopment Analysis (DEA) is an appropriate platform to develop multi-criteria evaluation metrics that can consider -in contrary to mono-criteria metrics — all positive and negative properties of DM-algorithms. We discuss different extensions of DEA that enable consideration of qualitative properties of DM-algorithms and consideration of users preferences in development of evaluation metrics. The results open new discussions in the general debate on model selection in statistics and machine learning.