Extend Lasso Class with Multitask and Uncertainty Options

[thomasATbayer]

Feature Request: Extend Lasso Class with Multitask and Uncertainty Options

Summary

Extend the existing Lasso model family in m_lasso.py with two new variants from scikit-learn:

1. MultiTaskLassoMother — wrapping sklearn.linear_model.MultiTaskLasso for simultaneous multi-output regression with shared sparsity.
2. ARDRegressionMother — wrapping sklearn.linear_model.ARDRegression for Bayesian linear regression with automatic relevance determination, providing native uncertainty estimates.

---

Motivation

1. MultiTaskLasso

The current LassoRegressorMother handles single-target regression only. In many life science and cheminformatics workflows, multiple related endpoints are predicted simultaneously (e.g. multiple ADMET properties, multiple assay readouts for the same compound series). MultiTaskLasso enforces a shared sparsity pattern across all outputs — features are selected or discarded jointly across all tasks — which is often more appropriate than fitting independent Lasso models when the targets share a common set of relevant features.

This fits naturally into Mother's existing multitask framework (e.g. MultitaskRandomForestMother) and would allow Lasso-style regularisation in multitask pipelines.

Reference: https://scikit-learn.org/stable/modules/generated/sklearn.linear_model.MultiTaskLasso.html

---

2. ARDRegression

ARDRegression (Automatic Relevance Determination) is a Bayesian linear model that places individual priors over each feature weight, effectively performing automatic feature selection while also producing calibrated predictive uncertainty in the form of a posterior over the weights. Compared to standard Lasso:

• It provides native uncertainty estimates (posterior mean and variance) without requiring conformal post-processing or ensembling.
• The regularisation strength is inferred from the data per-feature rather than set as a single global hyperparameter, which can be advantageous on high-dimensional datasets.
• It returns both a predicted mean and a predicted standard deviation, making it directly usable in Mother's uncertainty interface alongside existing uncertainty-aware models.

This would give users a lightweight, no-dependency uncertainty-aware linear baseline that complements the heavier ensemble and conformal approaches already in Mother.

Reference: https://scikit-learn.org/stable/modules/generated/sklearn.linear_model.ARDRegression.html

---

Proposed Changes

New classes in src/mother/ml/models/m_lasso.py

MultiTaskLassoMother

• Inherits from sklearn.linear_model.MultiTaskLasso and AbstractMotherPipeline.
• Implements get_hyperparameter_space() tuning the alpha regularisation parameter.
• Implements default_parameters().
• Consistent with the existing LassoRegressorMother pattern.

ARDRegressionMother

• Inherits from sklearn.linear_model.ARDRegression and AbstractMotherPipeline.
• Implements get_hyperparameter_space() covering the key Bayesian priors (alpha_1, alpha_2, lambda_1, lambda_2) and convergence settings.
• Implements default_parameters().
• Exposes predict_std() or integrates with Mother's existing uncertainty interface so that the posterior standard deviation is accessible downstream.

---

Affected Areas

Area	Change
src/mother/ml/models/m_lasso.py	Add MultiTaskLassoMother and ARDRegressionMother classes
src/mother/ml/__init__.py / model registry	Register both new model classes
test/unit/test_ml.py or new test file	Add tests for both new models
examples/	Optionally add or extend a notebook demonstrating multitask Lasso and ARD regression with uncertainty
mkdocs/docs/	Document the two new classes

---

Related

• LassoRegressorMother — existing single-task Lasso regression in m_lasso.py
• MultitaskRandomForestMother — existing multitask model for reference on how multitask models are integrated
• sklearn ARDRegression: https://scikit-learn.org/stable/modules/generated/sklearn.linear_model.ARDRegression.html
• sklearn MultiTaskLasso: https://scikit-learn.org/stable/modules/generated/sklearn.linear_model.MultiTaskLasso.html