Performance evaluation of seven multi-label classification methods on real-world patent and publication datasets

Purpose Many science, technology and innovation (STI) resources are attached with several different labels. To assign automatically the resulting labels to an interested instance, many approaches with good performance on the benchmark datasets have been proposed for multilabel classification task in the literature. Furthermore, several open-source tools implementing these approaches have also been developed. However, the characteristics of real-world multilabel patent and publication datasets are not completely in line with those of benchmark ones. Therefore, the main purpose of this paper is to evaluate comprehensively seven multi-label classification methods on real-world datasets. Design/methodology/approach Three real-world datasets (Biological-Sciences, Health-Sciences, and USPTO) from SciGraph and USPTO database are constructed. Seven multilabel classification methods with tuned parameters (dependency-LDA, MLkNN, LabelPowerset, RAkEL, TextCNN, TexRNN, and TextRCNN) are comprehensively compared on these three real-world datasets. To evaluate the performance, the study adopts three classification-based metrics: Macro-F1, Micro-F1, and Hamming Loss. Findings The TextCNN and TextRCNN models show obvious superiority on small-scale datasets with more complex hierarchical structure of labels and more balanced documentlabel distribution in terms of macro-F1, micro-F1 and Hamming Loss. The MLkNN method works better on the larger-scale dataset with more unbalanced document-label distribution. Research limitations Three real-world datasets differ in the following aspects: statement, data quality, and purposes. Additionally, open-source tools designed for multi-label classification also have intrinsic differences in their approaches for data processing and feature selection, which in turn impacts the performance of a multi-label classification approach. In the near future, we will enhance experimental precision and reinforce the validity of conclusions by employing more rigorous control over variables through introducing expanded parameter settings. Practical implications The observed Macro F1 and Micro F1 scores on real-world datasets typically fall short of those achieved on benchmark datasets, underscoring the complexity of real-world multi-label classification tasks. Approaches leveraging deep learning techniques offer promising solutions by accommodating the hierarchical relationships and interdependencies among labels. With ongoing enhancements in deep learning algorithms and large-scale models, it is expected that the efficacy of multi-label classification tasks will be significantly improved, reaching a level of practical utility in the foreseeable future. Originality/value (1) Seven multi-label classification methods are comprehensively compared on three real-world datasets. (2) The TextCNN and TextRCNN models perform better on small-scale datasets with more complex hierarchical structure of labels and more balanced document-label distribution. (3) The MLkNN method works better on the larger-scale dataset with more unbalanced document-label distribution.