Scholarlink

<p>Monotonic classification is a specialized endeavor commonly encountered in the real world, where improvements in feature values should not result in worse decision outcomes. The Fuzzy Monotonic <i>K</i>-Nearest Neighbors (FMKNN) is an important machine learning method for addressing such tasks. However, similar to the classical KNN classifier, FMKNN also faces challenges like low computational efficiency and sensitivity to noise. Although FMKNN can mitigate the impact of noisy labels to some extent, its inherent high computational cost and strict monotonicity constraints limit the algorithm’s classification performance and adaptability in practical scenarios. In this paper, we propose an instance selection method named Adaptive-Condensed Fuzzy Monotonic <i>K</i>-Nearest Neighbors (Ada-CFMKNN). This algorithm skillfully integrates the fuzzy ordinal decision entropy function with a performance prediction function to construct a novel comprehensive evaluation function. The core aim of this function is to dynamically optimize the balance between maintaining monotonicity and improving test accuracy. Through this adaptive balancing mechanism, Ada-CFMKNN can select the most representative instance set, effectively reducing the dataset size and significantly simplifying the subsequent computation process. Extensive experimental results demonstrate that Ada-CFMKNN can achieve up to 58.74% data reduction while preserving or even enhancing prediction accuracy of FMKNN algorithm on the designed data set. Additionally, compared with the instance selection methods MONIPS, MENN, and MCNN, Ada-CFMKNN shows different degrees of enhancement in ACCU, MAE, and NMI.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Adaptive condensed fuzzy monotonic K-nearest neighbors for monotonic classification

  • 期刊论文

  • 发表日期：

    2024年12月31日

查看全文
  • Jiankai Chen,
  • Zhongyan Li

摘要

Monotonic classification is a specialized endeavor commonly encountered in the real world, where improvements in feature values should not result in worse decision outcomes. The Fuzzy Monotonic K-Nearest Neighbors (FMKNN) is an important machine learning method for addressing such tasks. However, similar to the classical KNN classifier, FMKNN also faces challenges like low computational efficiency and sensitivity to noise. Although FMKNN can mitigate the impact of noisy labels to some extent, its inherent high computational cost and strict monotonicity constraints limit the algorithm’s classification performance and adaptability in practical scenarios. In this paper, we propose an instance selection method named Adaptive-Condensed Fuzzy Monotonic K-Nearest Neighbors (Ada-CFMKNN). This algorithm skillfully integrates the fuzzy ordinal decision entropy function with a performance prediction function to construct a novel comprehensive evaluation function. The core aim of this function is to dynamically optimize the balance between maintaining monotonicity and improving test accuracy. Through this adaptive balancing mechanism, Ada-CFMKNN can select the most representative instance set, effectively reducing the dataset size and significantly simplifying the subsequent computation process. Extensive experimental results demonstrate that Ada-CFMKNN can achieve up to 58.74% data reduction while preserving or even enhancing prediction accuracy of FMKNN algorithm on the designed data set. Additionally, compared with the instance selection methods MONIPS, MENN, and MCNN, Ada-CFMKNN shows different degrees of enhancement in ACCU, MAE, and NMI.