Automunge

- rethinking of defaults for binary classification labels transform
- recast from 'bnry' (single column boolean integer) to 'text' (one-hot encoding)
- found further validation that label interpretability benefits from seperate predictions for each class, ie plausability of each does not have to sum to unity as would be case for bnry encoding
- partly inspired by discussions in paper by Freund, Y., Schapire, R. E., et al. "Experiments with a new boosting algorithm"

- some soul searching for shft family transforms
- realized the infill options are innappropriate given row shifts
- so special case for shft family transforms to just default to adjacent cell infill
- without NArw aggregation
- this applies to shft family transforms,
- the dxdt family sequential transforms, as retain a basis for target row, is left in place

- new shft family entries for shf7 and shf8
- where shf7 applies upstream retn, shf4, shf5 and a downstream retn on shf4 and shf5
- and shf8 applies upstream retn, shf4, shf5, shf6 and a downstream retn on shf4, shf5, and shf6
- these are intended to simplify application for cases where user desires to apply mulitple shift operations at different time step periods on same data stream
- new infill type 'naninfill' to return data with inputation of NaN for infill points
- as some external libraries may prefer data without inputation
- added support for dimensionality reduction via Binary and PCA when excl_suffix passed as False
- corrected info_retention on mnm5 to True

- new 'shft' family of transforms for sequential data
- inspired by ICML Autonomous Driving workshop discussions by Sanjiban Choudhury and Arun Venkatraman
- shft is similar to dxdt family of transforms, but instead of taking relative difference between time steps it simply shifts a prior time step forward
- accepts parameter 'periods' for number of time steps
- and parameter 'suffix' for returned column suffix appender (as may be useful if applying the transform multiple times such as to distinguish)
- offered in library in six varients: shft, shf2, shf3, shf4, shf5, shf6
- where shft/shf2/shf3 are comparable with default periods of 1/2/3 respectively
- and shf4/shf5/shf6 return two seperate column: one derived from source column with a retn transform applied, and the second derived from the source column with either a shft/shf2/shf3 followed by a downstream retn normalziation
- where retn normalizaiton is similar to a min/max scaling but retains the sign of the source data (ie +/-)
- again this is sort of like a simpler version of the various dxdt family of sequential transformations already in library
- also inversion now supported for shft, shf2, shf3, shf4, shf5, shf6, lbnm, nmbd, copy, mnm3, mnm4, mnm5, mnm6, year

- revision of the oversampling method
- to accomodate edge case when ordinal labels have supplemental columns such as when NArw included

- revision of the oversampling method
- for cases where numeric target labels are aggregated into bins
- now this method is generalized to support binning by custom transformation functions