Temporary repository for our paper "DACE: Distribution-Aware Counterfactual Explanation by Mixed-Integer Linear Optimization," in IJCAI-20. This repository includes the code and scripts for reproducing the experimental results in our full version of our IJCAI paper.
DACE is a framework of Counterfactual Explanation (CE) that optimizes the following objective function: C(a | x) = d_M(x, x+a) + alpha * q_1(x+a), where
- a: action for the given input instance x
- d_M(x, x+a): l1-Mahalanobis distance (l1-MD)
- q_1(x+a): 1-Local Outlier Factor (1-LOF)
- alpha: Trade-off parameter between l1-MD and 1-LOF
If you want to use DACE, please set the following parameters:
- cost_type = 'DACE' (utilize l1-MD as the basic cost)
- alpha > 0 (Trade-off parameter between l1-MD and 1-LOF)
Basic running examples are provided in example.py:
$ python example.py
==================== TLPS ====================
* Action (Default: Yes -> No):
* duration_in_month: 12 -> 4 (-8)
* age: 23 -> 45 (+22)
* job: skilled_employee/official -> unemployed/unskilled-non-resident
* Scores:
* Time: 0.06543192
* alpha: 0.00000000
* Objective: 81.17015841
* Cost (TLPS): 81.17015841
* Mahalanobis: 8.34391339
* 10-LOF: 1.19636343
==================== MAD ====================
* Action (Default: Yes -> No):
* purpose: (vacation-does_not_exist?) -> retraining
* Scores:
* Time: 0.03359125
* alpha: 0.00000000
* Objective: 0.46314260
* Cost (MAD): 0.46314260
* Mahalanobis: 18.88119622
* 10-LOF: 1.13466846
==================== PCC ====================
* Action (Default: Yes -> No):
* other_debtors: none -> guarantor
* job: skilled_employee/official -> unemployed/unskilled-non-resident
* Scores:
* Time: 0.06525796
* alpha: 0.00000000
* Objective: 0.07530073
* Cost (PCC): 0.07530073
* Mahalanobis: 9.98328166
* 10-LOF: 1.13634779
==================== DACE ====================
* Action (Default: Yes -> No):
* duration_in_month: 12 -> 5 (-7)
* installment_as_income_perc: 4 -> 2 (-2)
* age: 23 -> 29 (+6)
* Scores:
* Time: 2.85304871
* alpha: 0.01000000
* Objective: 8.97257815
* Cost (l1-Mahal.): 8.96435471
* Cost (LOF): 0.82234381
* Mahalanobis: 2.27827220
* 10-LOF: 1.09590389
Quick examples for each classifier can be ran as follows:
$ python linear_ce.py 3
# Demonstration: linear_ce.py
- Dataset: FICO HELOC Dataset
- Classifier: l2-Regularized Logistic Regression Classifier
# 1st Individual with high risk of default
* Action (RiskPerformance: Bad -> Good):
* ExternalRiskEstimate: 70 -> 80 (+10)
* NetFractionRevolvingBurden: 70 -> 4 (-66)
* NumBank2NatlTradesWHighUtilization: 2 -> 0 (-2)
* Scores:
* Time: 9.92181595
* alpha: 0.01000000
* Objective: 8.11828936
* Cost (l1-Mahal.): 8.10626566
* Cost (LOF): 1.20237024
* Mahalanobis: 2.76675010
* 10-LOF: 1.19098847
# 2nd Individual with high risk of default
* Action (RiskPerformance: Bad -> Good):
* ExternalRiskEstimate: 69 -> 78 (+9)
* MSinceOldestTradeOpen: 155 -> 292 (+137)
* AverageMInFile: 81 -> 134 (+53)
* Scores:
* Time: 14.56293435
* alpha: 0.01000000
* Objective: 7.25777212
* Cost (l1-Mahal.): 7.21427372
* Cost (LOF): 4.34983913
* Mahalanobis: 2.26445112
* 10-LOF: 1.35521965
# 3rd Individual with high risk of default
* Action (RiskPerformance: Bad -> Good):
* ExternalRiskEstimate: 59 -> 84 (+25)
* MaxDelq2PublicRecLast12M: 4 -> 8 (+4)
* MaxDelqEver: 6 -> 8 (+2)
* Scores:
* Time: 18.00154517
* alpha: 0.01000000
* Objective: 11.94607759
* Cost (l1-Mahal.): 11.85384015
* Cost (LOF): 9.22374429
* Mahalanobis: 4.39399600
* 10-LOF: 1.39338479
$ python mlp_ce.py 3
# Demonstration: mlp_ce.py
- Dataset: Diabetes Dataset
- Classifier: Multi-Layer Perceptron Classifier
# 1st Individual with high risk of diabetes
* Action (DiseaseProgression: Bad -> Good):
* BMI: 35.0000 -> 29.8580 (-5.1420)
* Lamotrigine: 4.0431 -> 4.0273 (-0.0158)
* BloodSugarLevel: 91 -> 92 (+1)
* Scores:
* Time: 18.92564264
* alpha: 0.01000000
* Objective: 1.79501897
* Cost (l1-Mahal.): 1.78394779
* Cost (LOF): 1.10711775
* Mahalanobis: 1.42359425
* 10-LOF: 1.06446166
# 2nd Individual with high risk of diabetes
* Action (DiseaseProgression: Bad -> Good):
* TCells: 194 -> 202 (+8)
* LowDensityLipoproteins: 126.6000 -> 131.9600 (+5.3600)
* HighDensityLipoproteins: 43.0000 -> 44.8000 (+1.8000)
* Scores:
* Time: 19.74549178
* alpha: 0.01000000
* Objective: 0.86038674
* Cost (l1-Mahal.): 0.84653713
* Cost (LOF): 1.38496115
* Mahalanobis: 0.35309675
* 10-LOF: 1.06654302
# 3rd Individual with high risk of diabetes
* Action (DiseaseProgression: Bad -> Good):
* BMI: 28.2000 -> 25.0180 (-3.1820)
* BloodPressure: 112.0000 -> 112.7000 (+0.7000)
* Lamotrigine: 4.9836 -> 4.9674 (-0.0162)
* Scores:
* Time: 21.07739703
* alpha: 0.01000000
* Objective: 1.07244689
* Cost (l1-Mahal.): 1.06348294
* Cost (LOF): 0.89639446
* Mahalanobis: 0.88916807
* 10-LOF: 1.04295307
$ python forest_ce.py 3
# Demonstration: linear_ce.py
- Dataset: German Dataset
- Classifier: Random Forest Classifier
# 1st Individual with high risk of default
* Action (Default: Yes -> No):
* duration_in_month: 36 -> 14 (-22)
* credit_amount: 15857 -> 4558 (-11299)
* installment_as_income_perc: 2 -> 3 (+1)
* Scores:
* Time: 10.37286215
* alpha: 0.01000000
* Objective: 22.33710194
* Cost (l1-Mahal.): 21.51082580
* Cost (LOF): 82.62761388
* Mahalanobis: 4.82378028
* 10-LOF: 0.98032141
# 2nd Individual with high risk of default
* Action (Default: Yes -> No):
* duration_in_month: 48 -> 33 (-15)
* installment_as_income_perc: 3 -> 2 (-1)
* age: 24 -> 30 (+6)
* Scores:
* Time: 99.48086417
* alpha: 0.01000000
* Objective: 8.96938964
* Cost (l1-Mahal.): 8.04373529
* Cost (LOF): 92.56543557
* Mahalanobis: 1.96893252
* 10-LOF: 0.81689791
# 3rd Individual with high risk of default
* Action (Default: Yes -> No):
* credit_amount: 1842 -> 1373 (-469)
* age: 34 -> 37 (+3)
* Scores:
* Time: 44.72032961
* alpha: 0.01000000
* Objective: 2.22400200
* Cost (l1-Mahal.): 2.20602798
* Cost (LOF): 1.79740260
* Mahalanobis: 0.46215243
* 10-LOF: 1.54871136
@inproceedings{ijcai2020-395,
title = {DACE: Distribution-Aware Counterfactual Explanation by Mixed-Integer Linear Optimization},
author = {Kanamori, Kentaro and Takagi, Takuya and Kobayashi, Ken and Arimura, Hiroki},
booktitle = {Proceedings of the Twenty-Ninth International Joint Conference on
Artificial Intelligence, {IJCAI-20}},
publisher = {International Joint Conferences on Artificial Intelligence Organization},
editor = {Christian Bessiere},
pages = {2855--2862},
year = {2020},
month = {7},
note = {Main track}
doi = {10.24963/ijcai.2020/395},
url = {https://doi.org/10.24963/ijcai.2020/395},
}