Repository for "Synthesizing Optimal Action Sequences for Modifying Model Decisions" (Published in AAAI-2020)
$ pip install -r requirements.txt$ python run.py --target-model german \
--ckpt model.h5 \
--target-data data.npy \
--mode vanilla \
--l 4 \
--actions WaitYears Naturalize ChangeCreditAmount ChangeLoanPeriod AdjustLoanPeriod GetGuarantor GetUnskilledJob \. <package root>
+-- actions # contains definitions for the Action and Feature classes, and precondition helpers
+-- common # contains logger / path definitions
+-- heuristics # contains the code to implement the score functions
+-- models # contains model specific data, including model/data classes, custom action definitions and data
+-- <model>
+-- actions.py # list of model-specific action classes. Each action takes the entire features.json metadata
each action requires definition of following methods
- apply(): action transformation method. Can be used for both tensors and numpy arrays
- get_cost(): cost calculations. Can be used for both tensors and numpy arrays
- precondition():
- during optimization, conditions are converted to continuous functions and added to the cost
- during condition check, the conditions output boolean outputs
+-- dataset.py # wrapper class for npy dataset, as well as feature meta-info extraction routine
+-- features.json # json definition for each feature in dataset
+-- model.h5 # model checkpoint file in h5 format
+-- model.py # wrapper class to load (and train/save) each model
+-- data.npy # numpy array containing 100 random negatively classified points from the test set
+-- adult
+-- fanniemae
+-- recourse # main search module
+-- config.py # contains configuration
+-- result.py # contains classes to store instance and sequence level results
+-- search.py # main search algorithm - contains two classes for search:
- SequenceSearch for outer-loop that selects the next sequence to optimize
- ParamSearch for inner-loop that performs the minimization
+-- utils.py # miscellaneous utilities
Sample results from running the following command can be found in the results folder:
$ python .\run.py --target-model german --ckpt model.h5 --mode vanilla --exp-name sample_length_1 --l 1