Multi-Task Feature
Hi/Hello from Matraman Indonesia.
I’m trying to understand whether Dataiku Visual ML currently supports a true multi-target prediction task beyond multi-target regression.
From the documentation, I can see that Multi-target Regression allows multiple numerical targets. However, I’m wondering about a case where I have multiple targets that are categorical, for example:
- Target A → binary classification
- Target B → binary classification
- Target C → multiclass classification
Or even a mixed case such as:
- Target A → classification
- Target B → regression
Does Dataiku currently support training these targets as a single multi-target / multi-task model, rather than creating a separate prediction model for each target?
I’ve mostly worked with single-target prediction in Dataiku, and I’ve seen the Multi-target Regression option, but I haven’t found an equivalent option for multi-target classification or mixed target types.
If this isn’t natively supported, is there a recommended approach in Dataiku for this kind of use case?
Thanks!
Dataiku version used: 14.7
Comments
-
Hi!
Your understanding is correct. In Dataiku 14.7, Visual ML doesn’t natively support multi-target classification or a mixture of classification and regression targets within a single model. The native multi-target capability is currently limited to regression with multiple numerical targets.
The recommended approach is to create a separate Visual ML model for each target, while reusing the same prepared dataset and features. You can also automate the training and scoring of these models through a Dataiku workflow.
If jointly learning the targets is essential to your use case, you would need to implement a custom model in code—for example, a multi-head neural network—with suitable losses and evaluation metrics for each target type.