Skip to content

Exposing the tree for multiclass classification #6370

Closed
Closed
Exposing the tree for multiclass classification#6370
Labels
Priority:1Work that is critical for the release, but we could probably ship withoutarea-TreesTree-based trainers, including FastTree and LightGBMenhancementNew feature or requestquestionFurther information is requestedyellowmaintenance only
Milestone
ML.NET 5.0

Description

@oberma-schueco
oberma-schueco
opened on Oct 12, 2022

Hey.

We are currently porting our Code from the Accord framework to ML.net and struggle to do so. In the Accord.net framework we are using the C4.5 decision tree algorithm to process our data and so we were looking into multiclass trainers from ML.net.

With Accord.net we have the ability to inspect the trained tree and expose this via text form to the user. Are we able to get the same information with ML.net?

After reading into #4275 we are actively looking into the LightGBM trainer, but I am only able to find the trained tree ensemble for all the other options except the multiclass classification version of LightGBM.

Activity

  1. ghost added
    untriagedNew issue has not been triaged
    on Oct 12, 2022
  2. michaelgsharp
    added
    questionFurther information is requested
    on Nov 28, 2022

  3. michaelgsharp commented on Nov 28, 2022

    @michaelgsharp
    michaelgsharp
    on Nov 28, 2022
    Contributor

    So you should be able to use the same steps for LightGBM as you do the other tree models. When we use LightGBM after training we actually convert it to the same representation as the other tree models.

    @luisquintanilla any other suggestions?

  4. oberma-schueco commented on Nov 30, 2022

    @oberma-schueco
    oberma-schueco
    on Nov 30, 2022
    Author

    @michaelgsharp What method are you using exactly? For all the LightGBM extensions there are ModelParameters that expose the tree, except for the multiclass version of it. And that is what we are actually looking into.

    Can you explain how you do the conversion? Maybe there is a conversion from multiclass to another method?

  5. luisquintanilla
    added
    lightgbm
    on Mar 29, 2023
  6. luisquintanilla
    added this to the ML.NET 3.0 milestone on Mar 29, 2023
  7. ghost removed
    untriagedNew issue has not been triaged
    on Mar 29, 2023
  8. luisquintanilla
    assigned
    michaelgsharp
    on Mar 29, 2023
  9. luisquintanilla
    added
    P1
    untriagedNew issue has not been triaged
    on Mar 29, 2023

  10. luisquintanilla commented on Mar 29, 2023

    @luisquintanilla
    luisquintanilla
    on Mar 29, 2023
    Contributor

    Hi @sp44rk,

    Thanks for this issue. We've added this to our backlog.

  11. luisquintanilla
    added
    enhancementNew feature or request
    on Mar 29, 2023
  12. michaelgsharp
    modified the milestones: ML.NET 3.0, ML.NET 4.0 on Jan 23, 2024
  13. ghost removed
    untriagedNew issue has not been triaged
    on Jan 23, 2024
  14. ericstj
    modified the milestones: ML.NET 4.0, ML.NET 5.0 on Jan 16, 2025
  15. rosebyte
    added
    Priority:1Work that is critical for the release, but we could probably ship without
    area-TreesTree-based trainers, including FastTree and LightGBM
    and removed
    P1
    lightgbm
    on Sep 8, 2026
  16. rosebyte
    unassigned
    michaelgsharp
    on Sep 9, 2026
  17. rosebyte
    added
    yellowmaintenance only
    on Sep 21, 2026

  18. matouskozak commented on Oct 23, 2026

    @matouskozak
    matouskozak
    on Oct 23, 2026
    Member

    Triage:
    I see two possible workarounds are:

    • Export to ONNX and write code to read the tree structure and display it as text. This requires a compatible model and custom parsing.
    • Train separate binary models, one per class, and inspect each model’s TrainedTreeEnsemble. You would need to combine their predictions yourself.

    Hope this helps a bit, feel free to reach out if you still have issues.

  19. matouskozak
    closed this as completedon Oct 23, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    Priority:1Work that is critical for the release, but we could probably ship withoutarea-TreesTree-based trainers, including FastTree and LightGBMenhancementNew feature or requestquestionFurther information is requestedyellowmaintenance only

    Type

    No type

    Projects

    No projects

      Milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions