Relation Guided Message Passing for Multi-label Classification
Abstract: A well known-challenge in multi-label classification is modelling the dependencies between the labels. Most of the attempts in the literature focus on label dependencies that exhibit themselves through co-occurrences. Co-occurrences represent a pulling type of relationship between labels, meaning that labels that are observed together in training samples are more likely to co-occur. But other label relationships are common, such as a group of labels that never occur together. We call this a pushing relation. Successfully modeling such relations and the dependencies they induce can also lead to improved prediction performance.
In this work, we develop a graph-based dependency module that models multiple types of relations between labels and thus captures richer dependencies. The module is designed to be flexible so that it can be integrated into most embedding-based multi-label classification approaches. We propose a generic method to extract pulling and pushing relations between labels for any multi-label data. We then present Relation Guided Message Passing (RGMP), a Transformer based classifier for multi-label classification that uses the proposed label dependency module. Experiments on benchmark datasets show that RGMP yields similar or superior performance compared to state-of-the-art methods and the approach imposes only minor additional computational and memory overheads.
Submission Length: Regular submission (no more than 12 pages of main content)
Previous TMLR Submission Url: https://openreview.net/forum?id=k6E4Cxw2NP&referrer=%5BAuthor%20Console%5D(%2Fgroup%3Fid%3DTMLR%2FAuthors%23your-submissions)
Changes Since Last Submission: We have rephrased a sentence about our previous work to anonymize the submission thanks to desk's careful and quick investigation! We also removed the 'Acknowledgments' section for now due to anonymity concerns.
Assigned Action Editor: ~Vlad_Niculae2
Submission Number: 1552
Loading