<p>Multi-label text classification (MLTC) faces challenges such as label dependencies, semantic misalignment, and long-tail distributions. This paper introduces the KHAT-MLTC framework, which addresses these issues with three key innovations. First, we integrate the Kolmogorov–Arnold network (KAN) into MLTC, enhancing feature fusion and enabling adaptive integration of diverse information. Second, we propose a dual-channel label relationship model combining PMI-based co-occurrence and semantic similarity, using an attention-enhanced graph convolutional network (GCN) to model label dependencies. Third, we introduce a hierarchical attention mechanism that combines label-guided self-attention, label–word interaction attention, and co-attention to improve semantic alignment between text and labels. A dual-scale temporal model captures both short-term and long-term label dependencies, improving the model’s robustness in recognizing low-frequency labels. The model’s complex architecture, which incorporates RoBERTa, TextCNN, BiGRU, GCN, and multiple attention mechanisms, requires high-performance computing (HPC) resources, including distributed processing and parallel computation, to handle the large-scale and computationally intensive tasks. The adoption of HPC is crucial for real-time performance and efficient training, particularly when processing massive datasets in applications such as health care, e-commerce, and financial risk assessment. The model demonstrates fast convergence in early training iterations, enabling quick adaptation to new data. Beyond academic contributions, KHAT-MLTC offers substantial practical applications. Its ability to handle long-tail distributions is particularly valuable in health care, where accurate disease classification is crucial; in e-commerce, for product tagging and recommendations; and in financial risk assessment, where dynamic label dependencies can enhance predictive models. With the support of HPC, KHAT-MLTC’s real-time predictive capabilities can be deployed in high-demand industries that require timely decision-making. Experimental results on the AAPD and Reuters-21578 datasets demonstrate that KHAT-MLTC outperforms existing methods, improving the Macro-F1 score by over 6% and setting new benchmarks in low-frequency label recognition. These results highlight the model’s ability to address long-tail distributions, offering promising applications in industries where label imbalance and evolving relationships are common.</p>

KAN-based hierarchical attention with temporal dynamics for multi-label text classification

期刊论文
发表日期：

2025年12月11日

查看全文

Zhen Yuan,
Xuesi Ma

摘要

Multi-label text classification (MLTC) faces challenges such as label dependencies, semantic misalignment, and long-tail distributions. This paper introduces the KHAT-MLTC framework, which addresses these issues with three key innovations. First, we integrate the Kolmogorov–Arnold network (KAN) into MLTC, enhancing feature fusion and enabling adaptive integration of diverse information. Second, we propose a dual-channel label relationship model combining PMI-based co-occurrence and semantic similarity, using an attention-enhanced graph convolutional network (GCN) to model label dependencies. Third, we introduce a hierarchical attention mechanism that combines label-guided self-attention, label–word interaction attention, and co-attention to improve semantic alignment between text and labels. A dual-scale temporal model captures both short-term and long-term label dependencies, improving the model’s robustness in recognizing low-frequency labels. The model’s complex architecture, which incorporates RoBERTa, TextCNN, BiGRU, GCN, and multiple attention mechanisms, requires high-performance computing (HPC) resources, including distributed processing and parallel computation, to handle the large-scale and computationally intensive tasks. The adoption of HPC is crucial for real-time performance and efficient training, particularly when processing massive datasets in applications such as health care, e-commerce, and financial risk assessment. The model demonstrates fast convergence in early training iterations, enabling quick adaptation to new data. Beyond academic contributions, KHAT-MLTC offers substantial practical applications. Its ability to handle long-tail distributions is particularly valuable in health care, where accurate disease classification is crucial; in e-commerce, for product tagging and recommendations; and in financial risk assessment, where dynamic label dependencies can enhance predictive models. With the support of HPC, KHAT-MLTC’s real-time predictive capabilities can be deployed in high-demand industries that require timely decision-making. Experimental results on the AAPD and Reuters-21578 datasets demonstrate that KHAT-MLTC outperforms existing methods, improving the Macro-F1 score by over 6% and setting new benchmarks in low-frequency label recognition. These results highlight the model’s ability to address long-tail distributions, offering promising applications in industries where label imbalance and evolving relationships are common.