关于学习空间及分类风险评价的非单调一致性

About Learning Space and Non-Monotonicity in Assessment of Classification Risk

允许经验风险不为0是现代模式分类器构造方法区别于传统模式分类器构造方法的标志.为了进一步研究分类器构造观点的变化对模式分类系统所产生的更深入的影响,拓展模式分类系统的学习空间,作者讨论了限制经验风险必须为0的传统模式分类系统在分类性能问题上所受的限制,分析了影响模式分类系统分类性能的关键因素,给出了学习空间可拓展的必要条件,并构造了一种投机学习方法,证明了学习空间可拓展的充分条件.同时,在实验中观察到,分类器评价与测试集上的分类风险是非一致单调的.这一结论对于模式识别及其应用研究是严峻的.

The characteristic of modern pattern classification methods is to admit empirical risk non-zero, whereas inseparable feature set never provides a chance for learning algorithms to make a classifier with zero of empirical risk. In order to investigate the potential connection between in- separable feature set, which is usually thought as trustless on intuition, and the modern idea on learning problem, this paper argues the necessary condition of the availability of inseparable feature set, by which elaborates an opportunistic learning method to validate the sufficient condition experimentally. Experimental evidences show that inseparable feature subset can make important contributions for improving the performance of pattern classifier. Further more, the relation between the assessment of classification and the predictive performance on test set is proved to be non-monotone in experiments. Both the analytical results and experimental studies reflect that this conclusion may be a challenge to pattern classification and its applications in the future.