For tuning hyperparameters with Apache Spark ML / MLlib, when should I use Spark ML's built-in tuning algorithms vs. Hyperopt?

[Joseph_B]

When should I use Spark ML's CrossValidator or TrainValidationSplit, vs. a separate tuning tool such as Hyperopt?

[Joseph_B]

Both are valid choices. By default, I'd recommend using Hyperopt nowadays. Here's the rationale, as pros & cons of each.

Spark ML's built-in tools

• Pros: These fit the Spark ML Pipeline framework, so you can keep using the same type of APIs.
• Cons: These are designed for brute force grid search. That's fine for a small number (say up to ~3) hyperparameters, but it becomes inefficient when you have many hyperparameters or when you want to test many combinations.

Hyperopt

• Pros: This provides a more adaptive, iterative algorithm for tuning which can be more efficient in terms of the number of hyperparameter settings you need to try to reach a given accuracy. This is especially important when tuning many hyperparameters to testing many settings.
• Cons: (See pros of Spark ML.)

[Kaniz]

Hi @Joseph Bradley​ , Thanks for such an informative post!