v0.12.9

BR-rerank

• Writes predictions and confidence scores to file
• Allows users to specify max GB training iteration

v0.12.8

Make pyramid compatible with Java 9+

v0.12.7

Fix tar long file issue

v0.12.6

New functions:

• Add the ability to normalize the feature matrix when creating datasets
• Add new calibrator type: identity (using uncalibrated confidence as final confidence) and zero (mapping everything to 0 confidence to avoid automation)
• Support F1 as calibration target
• Report Average Precision for each label classifier
• Unify LR and GB top feature file format

Internals:

• Use expected precision, recall, F1 as features for reranker calibrator
• Sample random support sets as negative calibration candidates
• Better bounds for isotonic regression output
• Improved CBM prediction stop condition based on the Threshold Algorithm

v0.12.5

release code and data for the ECML-PKDD 2019 paper "Learning to Calibrate and Rerank Multi-label Predictions"

v0.12.4

• smoothed isotonic regression label calibrator
• use un-interpolated isotonic regression for label calibrator and interpolated isotonic regression for ensemble set calibrator
• make support predictions consistent with top_sets.csv

v0.12.3

Fixed an issue with the top sets

v0.12.2

Fix a bug in CombSUM ensemble caused by empty set

v0.12.1

Use F1 based confidence in CombSUM ensemble

v0.12.0

• add precision, recall, F1, and ground truth to reports
• support F1 as the target metric in confidence threshold tuning
• average-confidence based ensemble
• isotonic regression outputs interpolated confidence