Post Jan-Feb 2025 Workshop Suggestions

[BalzaniEdoardo]

User Suggested Enhancements

This list is meant to be extended as more feedback comes in! So far, extension proposal included:

1. Simplify the get/set_params and the sklearn parameter grid definition, using labels to retrieve individual basis components from tree structured composite basis.
2. Allow a partial specification of model coefficients: some coefficients may be fixed and provided by the users, while the remaining coefficient will be learned.
3. Confidence measure on coefficients (Bayesian GLM).
4. HMM-GLM implementation.

Fixes

1. Adding a transformer basis with a regular basis results in an error message that is difficult to understand.
2. Setting a negative regularization strength doesn't result in an error.

[billbrod]

• set_input_shape is confusing. we can come up with better default behavior, so that the (hopefully) user only really needs to worry about it with a composite basis that combines atomic basis objects that act on more than 1d inputs

Suggested behavior:
- all atomic basis objects should assume 1d input on init (composite basis behavior unchanged)
- compute_feature should reset input shape everytime (atomic and composite)
- more helpful error message if they try to call split_by_feature without it being set
- have transformer behave similarly: atomic basis should reset input_shape everytime on transformer
- composite basis can't do this so have a helpful error message

[billbrod]

• add ability to predict behavioral variable, which would require a different observation model
• as part of that, should have an example of how to implement your own observation model? in addition to the developer docs