gemclus.add_mlcl_constraint¶
- gemclus.add_mlcl_constraint(gemini_model, must_link=None, cannot_link=None, factor=1.0)[source]¶
Adds must-link and/or cannot-link constraints to a discriminative clustering model. The contraints are ensure by respectively minimising or maximising the \(\ell_2\) norm between the prediction vectors. It is thus possible that not all constraints are fully satisfied.
- Parameters:
- gemini_model: MLP___, Linear___ or Categorical___
A GemClus model that involves gemini maximisation with gradient descent.
- must_link: ndarray of shape (n_constraints, 2) or None, default=None
The constraints of samples being together must be described by a list of pairs of indices [(i1,j1),…, (iN, jN)]. If set to None, no must-link constraint is applied on the model.
- cannot_link: ndarray of shape (n_constraints, 2) or None, default=None
The constraints of samples which must not be in the same cluster must be described by a list of pairs of indices [(i1,j1),…, (iN, jN)]. If set to None, no cannot-link constraint is applied on the model.
- factor: float, default=1.0
A weighting hyperparameter for the constraints in gradient descent.
- Returns:
- The model gemini model with decorated gradient functions to satisfy must-link / cannot-link constraints.
Examples using gemclus.add_mlcl_constraint¶
Consensus clustering with linking constraints on sample pairs