Scholarlink

<p>One of the key problems in Multi-Label Classification is label interdependence which is a critical factor in determining the performance of a given multi-label classifier. This problem has been attempted through the use of chaining classifiers or Multi-Layered stacking model architectures. However, these approaches neglect the order of labels while incorporating the label inter-dependencies which in itself is a <i>Non-Deterministic Polynomial Time Hard (NP-Hard)</i> problem, neither do they provide space for partial-dependency in a full-dependency architecture. Therefore, this work uses an Evolutionary approach to extract the optimal label order for arranging the labels. Additionally, novel crossover and mutation mechanisms namely <i>Ordered Slicing Crossover</i> and <i>Loop Shift Mutation</i> have also been introduced. Furthermore, a <i>Dependency Filtering Framework</i> is introduced to render partial dependency to avoid redundant labels and forced full dependency. The experiments were conducted on ten benchmark datasets with varying numbers of labels. Different performance metrics were used to evaluate the effectiveness of the proposed method, and it was compared to other state-of-the-art classifier models, showing an improvement of <i>14.73%</i> on the best-ranked model to date. The experiments also established reduced prediction time for the proposed approach. The code of the proposed method is available <a href="http://surl.li/spqjq">here</a>.</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

Permutation driven evolutionary ordering with dependency filtering for multi-label classification

  • 期刊论文

  • 发表日期：

    2025年1月16日

查看全文
  • Ankush Jain,
  • Daksh Gupta,
  • Sarthak Shukla,
  • Vishal Srivastava

摘要

One of the key problems in Multi-Label Classification is label interdependence which is a critical factor in determining the performance of a given multi-label classifier. This problem has been attempted through the use of chaining classifiers or Multi-Layered stacking model architectures. However, these approaches neglect the order of labels while incorporating the label inter-dependencies which in itself is a Non-Deterministic Polynomial Time Hard (NP-Hard) problem, neither do they provide space for partial-dependency in a full-dependency architecture. Therefore, this work uses an Evolutionary approach to extract the optimal label order for arranging the labels. Additionally, novel crossover and mutation mechanisms namely Ordered Slicing Crossover and Loop Shift Mutation have also been introduced. Furthermore, a Dependency Filtering Framework is introduced to render partial dependency to avoid redundant labels and forced full dependency. The experiments were conducted on ten benchmark datasets with varying numbers of labels. Different performance metrics were used to evaluate the effectiveness of the proposed method, and it was compared to other state-of-the-art classifier models, showing an improvement of 14.73% on the best-ranked model to date. The experiments also established reduced prediction time for the proposed approach. The code of the proposed method is available here.