Skip to article frontmatterSkip to article content
Site not loading correctly?

This may be due to an incorrect BASE_URL configuration. See the MyST Documentation for reference.

Tasks, constraints, and the search space

A discovery task is a precise statement of which descriptions are eligible and how to rank them. The same task can be passed to different compatible algorithms.

Loading...

Three different limits

result_set_size limits the number of returned descriptions, not the number examined. A request for ten results can still require exploring many thousands of candidates.

depth limits the number of conditions. It controls both interpretability and combinatorial complexity. Start shallow and increase it for a reason.

min_quality supplies a quality floor. Returned qualities must be strictly above it. The effective threshold can rise during top-k search once enough better candidates have been found. An eligible result list may contain fewer than the requested number of subgroups.

Absolute and relative size constraints

MinSizeFilter(10) requires ten rows. MinSizeFilter(0.1) requires at least 10% of the task dataset. Integer and floating-point arguments deliberately have different meanings; 10.0 is not an alternative spelling of ten rows.

For relative minimum size, a fractional threshold is rounded up to a row count; relative maximum size is rounded down. A relative threshold on validation or discovery subsets refers to their own sizes when their tasks are constructed.

Pruning constraints versus result filters

A refinement adds conditions, so it cannot gain rows. Therefore a group that already has fewer than ten rows cannot produce a refinement with ten rows. Minimum size is anti-monotone and can safely prune the entire branch.

Target share behaves differently. A subgroup with a low positive share can contain a strongly positive refinement. Rejecting its whole branch would lose valid results.

FilterPrunes a failing branch?Reason
MinSizeFilterYesRefinements cannot gain rows
MinPositivesFilterYesRefinements cannot gain positives
MinMajorityCountFilterYesRefinements cannot gain rows of any class
MaxSizeFilterNoA refinement can become small enough
MinTargetShareFilterNoA refinement can become enriched
CustomFilterOnly if explicitly declaredThe analyst must justify anti-monotonicity
Loading...

Filters supplied to a task participate in search and result eligibility. Post-hoc filters instead process an already ranked list; they cannot recover a candidate that never entered that list.

The language is part of the question

For the standard searches:

With fixed numeric bins, the search sees those bins. With thresholds, it sees the supplied one-sided cuts. An exhaustive search over one language need not match a search over the other. Build interval candidates explicitly when a bounded interval should be available as one condition.

DSSD differs: its refinement operator permits certain repeated conditions on a feature and, by default, recomputes numeric thresholds inside each subgroup. Use n_bins=None to keep its numeric candidates fixed. This makes search-space choices important when comparing DSSD with other methods.

Change one factor at a time

Loading...

The best attainable quality cannot decrease when the allowed depth grows and all other eligibility rules remain fixed. But the computation can grow quickly, and more complex descriptions are not automatically more useful.

Experiment. Choose a constraint that is not anti-monotone, such as “at least 80% positives.” Construct a small parent subgroup that fails and a refinement that passes. Explain why treating it as a pruning filter would be incorrect.