SFI.inference.sparse.greedy module

SFI.inference.sparse.greedy — Stepwise greedy selection

Classic forward / backward / bidirectional stepwise search.

• Forward: at each step, add the feature that maximises the information gain.

• Backward: start from the full model and drop the least useful feature.

• Bidirectional: alternate one forward step then one backward step, keeping whichever direction improves the score.

The algorithm naturally produces exactly one support per cardinality (forward path) or a monotonic path from full to empty (backward), yielding a clean Pareto front.

class SFI.inference.sparse.greedy.GreedyStepwiseStrategy(*, direction='forward', report_time=False)

Bases: SparsityStrategy

Forward / backward / bidirectional stepwise selection.

Parameters:

• direction ("forward" | "backward" | "both", default "forward") –

  Which direction(s) to search.

  • "forward": start empty, add one feature at a time.

  • "backward": start full, drop one feature at a time.

  • "both": run both directions and merge the Pareto fronts.

• report_time (bool, default False) – Log elapsed wall-clock time when done.

name: str = 'greedy'

Short identifier used in SparsityResult.method.

run(scorer, *, max_k, **_kwargs)

Execute the search and return a SparsityResult.

Parameters:

• scorer (SparseScorer) – Provides info_and_coeffs / vmap_info for evaluating candidate supports.

• max_k (int) – Maximum model size to explore.

Return type:

SparsityResult