Changelog¶

This is a list of changes that have been made between releases of the seismometer library. See GitHub for full details.

Breaking changes may occur between minor versions prior to the v1 release; afterwhich API changes will be restricted to major version updates.

0.3.0¶

Addresses #77 by removing Aequitas and replacing with a great_tables based fairness audit. (#77)
Added ExploreBinaryModelMetrics to see plots of individual metrics, including number needed to treat. (#86)
Reorganize methods making initial import and public api more standard (#102)
Includes confusion matrix rates into Binary Fairness metrics (#108)
Renames Flagged to Flag Rate for clarity (#108)
Add function to load example datasets (#113)

Remove remaining references to -1 ‘invalidation’; validate directly on time comparison when needed (#100)
Fixes scaling issue for binary classfier scores that use the range 0-100 rather than 0-1. (#101)
Fixes a few minor ux issues. (#109)
- Disable dropdowns with only one valid option.
- Fix the Sensitivity/Specificity/PPV plot to move the label to the lower right.
- Fix the Legend in the new Fairness Audit table to improve readability.
- Add right border to the count column.
Remove NotebookHost class that was no longer in use. (#114)

Added additional “merge_strategy” event configuration option. (#76)
Viable merge strategies are “first”, “last”, “nearest”, “forward”, and “count”. (#76)
Added ExploreSubgroups as a drop in replacement for sm.cohort_list (#82)
Added ExploreModelScoreComparison to compare two scores againts a shared target (#82)
Added ExploreModelTargetComparison to compare a single score across two targets (#82)
Added MultiselectDropdownWidget as a new widget for selecting cohort_dicts, uses a drop down and dismissalbe tags to keep the UX neater. (#82)
Updated handling around primary_output and outputs, so that if primary_output is in outputs, it does not get added in again during startup. (#82)
Cast features to dtypes in the dictionary yml when specified in the dictionary configuration. (#92)
Cast event _Values to dtypes in the dictionary yml when specified in the dictionary configuration, done after imputation. (#92)
Add a configuration helper that can generate a dictionary file for events or predictions. (#93)

Limits control max-widths to 1200px in most cases, allowing row wrap when needed. (#81)
Exclude time columns from ydata-profiling. (#88)
Temporarily remove CLI - variation in templates is not stable enough to ensure robustness. (#90)

Add number needed to treat (NNT) with fixed rho=1/3 and net benefit, from med_metrics package. (#78)

Warning message on censored cohorts is coerced to string before logging. (#65)
Log a warning if cohort source column is not found in the dataframe. (#67)
Improved support for datasets with a large number of cohort columns. Allowing row wrapping of cohort selection. (#70)
Improved the Fairness audit iframe support by increasing the height a bit to account for cohort error messages. (#70)
Allow long filenames by hashing the ends, this fix allow large cohort selection lists when creating reports and fairness audits which are cached to disk. (#70)
Handle merging without context id; assumes events dataframe columns. (#71)
Simplify merge logic for the single ‘first’ strategy. This removes coalecsing logic by assuming an event type always has times (or never does). (#72)
Hardens restriction of events (with times) to occur after prediction time + window offset, not having the unintuitive partial information for early, late, and unknown timings. (#72)

Updated integration_guide/index.rst to add a table documenting examples from the Seismometer Community of open source tools that are compatible/integrated with Seismometer. (#74)

Changed box and whisker plots to violin plots (#36)
Added Exploration controls to allow setting dynamic values for thresholds, targets, outcomes, etc. (#41)
Add an optional aggregation_method to event objects of usage_config. (#48)
Modify the accessors of ConfigProvider.events to return a dictionaries of events instead of a list. (#48)
Update model and cohort performance plots to respect the aggregation_method: supports max (default), min, first and last. (#48)
Configuration now supports multiple event types as the source for a single event. Values for the source events are assumed to be compatible, like both being Boolean. (#54)
Updates the Seismogram class constructor to take precisely a DataLoader and a ConfigProvider, improving separation between the class and configuration loading. (#56)
Adds a post-load hook for modifying the otherwise ready dataframe. This hook is not accessible via normal run_startup, and requires direct initialization of Seismogram. (#56)

Fixes issue with IFrames not being displayed in vscode. (#53)
Fixes issue with ipywidgets being hidden in sphinx docs. (#55)
Fixes issue with iframe sources not being found in sphinx docs. (#55)
Update “Generate Report” button width for Cohort Comparison widget. (#58)
Update various plot titles to <h4> headings for consistent size/theming. (#58)
New CensoredResultException, for when a plot cannot render due to too few results. (#58)

Added a developer guide section for custom visualizations. (#64)
Updated docstrings for exploration widgets and improved sphinx documentation layout. (#64)

Added FilterRule.all() and FilterRule.none() class methods for matching all or no rows of a dataframe. (#27)
Updated plots to use HTML and SVG over pngs from matplotlib (#28)
Added DiskCachedFunction to allow disk caching of HTML content (#28)
Restructures Seismogram load to allow prioritizing in memory dataframe over loading predictions and/or events. (#20)
seismometer.run_startup() can now accept preloaded prediction and event dataframes that take precendence over loading from configuration (#34)

Fixes the header of sm.show_info() to start the table zebra stripe on the right row. (#24)
Updated the defaulting for censor_threshold` in _plot_leadtime_enc (#35)
Take len of column list for count (#42)

Initial release!