Hello Ehud, These are excellent, well-reasoned feature proposals for GaussianMixture! You've accurately identified potential gaps in flexibility, especially regarding parameter control and covariance structure.
Here's an assessment of your proposals in the context of scikit-learn's GaussianMixture implementation: 1. 🧊 Freezing Component Weights (weights_init) This is a very useful feature, particularly in transfer learning or when components are physically constrained (e.g., modeling known proportions in a mixture). Current State: The GaussianMixture implementation currently optimizes all parameters (means, covariances, and weights) during the Expectation-Maximization (EM) loop, even if initialized via weights_init. There is no built-in mechanism to freeze the weights. Implementation Feasibility: You are correct; this should be relatively easy to implement. The EM algorithm has a separate step for updating weights, called M-step for weights. To implement frozen_weights=True, you would simply add a conditional check in the M-step for weights: if the weights are frozen, skip the update calculation and keep the values from weights_init. Parameter Suggestion: A boolean parameter like frozen_weights (or perhaps fix_weights to align with common ML terminology) would be clear and appropriate, specifically acting as a flag when weights_init is provided <a href="https://textinvisible.org/";>https://textinvisible.org/</a> _______________________________________________ scikit-learn mailing list -- [email protected] To unsubscribe send an email to [email protected] https://mail.python.org/mailman3//lists/scikit-learn.python.org Member address: [email protected]
