Feature request: Multi-label classification support

[hi-tech-jazz]

It would be great to have support for multi-label prediction. Methods of achieving this is briefly mentioned in the Sillas and Freitas survey of local hierarchical patterns. Implementation inspiration could possibly be drawn from this project, which shares its objective but unfortunately seems abandoned. Would this be possible?

[mirand863]

Hello,

Thanks for the interest in HiClass!

Can you please tell me what is your use-case?

Do you need multi-label prediction for all 3 hierarchical classifiers or only on the local classifier per node would be enough? I am currently rewriting the local classifier per level to pass sklearn's checks and support sparse matrices, but the next feature I will add is probably multi-label classification at least for the local classifier per node.

[hi-tech-jazz]

Nice to hear! It would be valuable for all, as I am trying to compare each method. The application is to annotate topical categories to blog posts.

[mirand863]

Nice to hear! It would be valuable for all, as I am trying to compare each method. The application is to annotate topical categories to blog posts.

@hi-tech-jazz do you happen to know of any public dataset similar to yours that I could use for testing?

[hi-tech-jazz]

Nice to hear! It would be valuable for all, as I am trying to compare each method. The application is to annotate topical categories to blog posts.

@hi-tech-jazz do you happen to know of any public dataset similar to yours that I could use for testing?

Sorry for replying late. This dataset is pretty similar to mine.

[mirand863]

Hi @hi-tech-jazz,

I am not sure this dataset is appropriate for multi-label classification. Furthermore, I have talked with my supervisor and we decided multi-label classification will not be implemented for now. Firstly, I have no proper dataset or metrics to test any implementation. Secondly, there are other features and documentation updates that we judged more important for a re-submission to JMLR. However, I will keep this issue open and revisit it in the future when I get more research done on the topic and circumstances change.

Kind regards,
Fabio

[mirand863]

Hello @hi-tech-jazz,

Just a quick update. I found a dataset to test multi-label classification. I will be working on this shortly and keep you posted.

Kind Regards,
Fabio

[mirand863]

Nice to hear! It would be valuable for all, as I am trying to compare each method. The application is to annotate topical categories to blog posts.

@hi-tech-jazz do you happen to know of any public dataset similar to yours that I could use for testing?

Sorry for replying late. This dataset is pretty similar to mine.

Hello @hi-tech-jazz,
Sorry for taking too long to start working on this. We are deciding on ways to represent the multi-label hierarchical data for training and predicting.
So far we are thinking about adding another dimension, for example:

[
    [
        [Woman, Mermaid], [Fish, Mermaid]
    ],
    [
        [Duck, Platypus], [Beaver, Platypus], [Otter, Platypus]
    ],
]

Do you have any suggestions that would make your life easier as a user?

[RobinRojowiec]

I have a similar problem. Wanted to have multiple labels on the leaves of the hierarchy tree. Your solution looks exactly like what I was looking for.
Hope that is added soon.

[mirand863]

Hi @RobinRojowiec

Thank you for the interest! There is an open pull request where we started a multi-label implementation for the local classifiers per node and per parent node #76. While it is still far from being marked as complete for merging, perhaps you could give it a try and let me know if you have any questions, issues or feedback.

Best regards,
Fabio