Generalized Inverse Classificationmichaeltlash.com/slides/lash_sdm2017_slides.pdfGeneralized Inverse...
Transcript of Generalized Inverse Classificationmichaeltlash.com/slides/lash_sdm2017_slides.pdfGeneralized Inverse...
GENERALIZED INVERSECLASSIFICATIONSDM ’17
Michael T. Lash1, Qihang Lin2, W. Nick Street2, JenniferG. Robinson3, and Jeffrey Ohlmann2
1Department of Computer Science, 2Deparment of ManagementSciences, 3Department of Epidemiology
www.michaeltlash.com
What is inverse classification?
# The process of making meaningful perturbations to a testinstance such that the probability of a desirable outcome ismaximized.
1
What is inverse classification?
# The process of making meaningful perturbations to a testinstance such that the probability of a desirable outcome ismaximized.
1
What is inverse classification?
# The process of making meaningful perturbations to a testinstance such that the probability of a desirable outcome ismaximized.
1
What is inverse classification?
# The process of making meaningful perturbations to a testinstance such that the probability of a desirable outcome ismaximized.
1
What is inverse classification?
# The process of making meaningful perturbations to a testinstance such that the probability of a desirable outcome ismaximized.
1
What is inverse classification?
# The process of making meaningful perturbations to a testinstance such that the probability of a desirable outcome ismaximized.
1
What is inverse classification?
# The process of making meaningful perturbations to a testinstance such that the probability of a desirable outcome ismaximized.
1
What is inverse classification?
# The process of making meaningful perturbations to a testinstance such that the probability of a desirable outcome ismaximized.
What about the meaningful part of the definition?1
Meaningful Perturbations
Well...lets visit some past work!# Michael T. Lash, Qihang Lin, W. Nick Street, and
Jennifer G. Robinson, “A budget-constrained inverseclassification framework for smooth classifiers”, arXivpreprint arXiv:1605.09068, submitted.
2
Meaningful Perturbations
Begin with a basic formulation.
2
Meaningful Perturbations
2
Meaningful Perturbations
2
Meaningful Perturbations
Some regressor
Segment features.
2
Meaningful Perturbations
Some regressor
Estimate indirectly changeable.
2
Meaningful Perturbations
Some regressor
Update objective function. Add constraints.
2
Meaningful Perturbations
Some regressor
Cost-change function
2
Meaningful Perturbations
Some regressor
Budget
2
Meaningful Perturbations
Some regressorBounds
2
Main Contributions
1. Relax assumptions about f (·).
Some regressorBounds
3
Main Contributions
1. Relax assumptions about f (·).
Some regressorBounds
3
Main Contributions
1. Relax assumptions about f (·).
Some regressorBounds
Generalized inverse classification 3
Main Contributions
1. Relax assumptions about f (·).2. Quadratic cost-change function.
Some regressor
4
Main Contributions
1. Relax assumptions about f (·).2. Quadratic cost-change function.
Some regressor
4
Main Contributions
1. Relax assumptions about f (·).2. Quadratic cost-change function.
Some regressor
4
Main Contributions
1. Relax assumptions about f (·).2. Quadratic cost-change function.3. Three real-valued heuristic optimization
methods and two sensitivity analysis-basedoptimization methods.
* Projection operator to maintain feasibility.
5
Optimization Methodology
Heuristic
# Hill Climbing + Local Search (HC+LS)# Genetic Algorithm (GA)# Genetic Algorithm + Local Search (GA+LS)
Sensitivity Analysis
# Local Variable Perturbation – First Improvement (LVP-FI)# Local Variable Perturbation – Best Improvement (LVP-BI)
6
Experiment Decisions and Data
# f (·): Random forest# H(·): Kernel regression# Dataset 1: Student Performance (UCI Machine Learning
Repository).# Dataset 2: ARIC# One f for optimization, separate f for heldout evaluation.
7
Results: Student Performance
8
Results: Student Performance
8
Results: ARIC
9
Results: ARIC
Need sparsity constraints 9
Conclusions
# Generalized Inverse Classification: can use virtually anylearned f (as shown by experiments w/ Random Forestclassifier).
# Our proposed methods were successful, although thisvaried by dataset.
10
GENERALIZED INVERSECLASSIFICATIONSDM ’17
Michael T. Lash1, Qihang Lin2, W. Nick Street2, JenniferG. Robinson3, and Jeffrey Ohlmann2
1Department of Computer Science, 2Deparment of ManagementSciences, 3Department of Epidemiology
www.michaeltlash.com
10
Causality and Inverse Classification
# Yes! ....# What we’re doing:
1. Imposing our own causal structure (DAG).2. We’re not taking the usual counterfactual approach.
# Future work will focus on incorporating causalmethodology...
11
Causality and Inverse Classification
# Yes! ....
# What we’re doing:1. Imposing our own causal structure (DAG).2. We’re not taking the usual counterfactual approach.
# Future work will focus on incorporating causalmethodology...
11
Causality and Inverse Classification
# Yes! ....# What we’re doing:
1. Imposing our own causal structure (DAG).2. We’re not taking the usual counterfactual approach.
# Future work will focus on incorporating causalmethodology...
11
Causality and Inverse Classification
# Yes! ....# What we’re doing:
1. Imposing our own causal structure (DAG).2. We’re not taking the usual counterfactual approach.
# Future work will focus on incorporating causalmethodology...
11