CAJ | 학술논문

有序回归是特殊的机器学习范式，其目标是利用数据间内在的序标号以划分模式。尽管众多算法相继提出，但经典的最小平方回归( LSR)尚未应用于有序回归场景。为此，文中采用累积标号编码和间隔扩大策略，在LSR基础上提出判别最小平方有序回归( DLSOR)。 DLSOR在对回归函数无需施加约束的前提下，仅通过改造标号实现有序信息的嵌入和类间间隔的扩大，从而确保DLSOR在与LSR具有相当模型复杂度的同时，既保证较高的分类精度，又获得较低的平均绝对误差。实验验证DLSOR在提升有序回归性能上的优越性。
유서회귀시특수적궤기학습범식，기목표시이용수거간내재적서표호이화분모식。진관음다산법상계제출，단경전적최소평방회귀( LSR)상미응용우유서회귀장경。위차，문중채용루적표호편마화간격확대책략，재LSR기출상제출판별최소평방유서회귀( DLSOR)。 DLSOR재대회귀함수무수시가약속적전제하，부통과개조표호실현유서신식적감입화류간간격적확대，종이학보DLSOR재여LSR구유상당모형복잡도적동시，기보증교고적분류정도，우획득교저적평균절대오차。실험험증DLSOR재제승유서회귀성능상적우월성。
Ordinal regression is a special machine learning paradigm and its objective is to classify patterns by using a between-class natural order property between the labels. Although many algorithms are proposed, the classical least squares regression ( LSR) is not applied to the ordinal regression scenario. In this paper, a discriminative least squares ordinal regression ( DLSOR) is proposed by using the cumulative labels and the margin-enlarging technique. Without constraints imposed on the regression function, DLSOR can embed ordinal information and expand between-class margin only through the label transformation. Thus, a high classification accuracy and low mean absolute errors can be guaranteed with the premise that the model complexity of DLSOR is consistent with that of LSR. The experimental results demonstrate the superiority of the proposed method in improving the ordinal regression performance.