Generate Actionable Counterfactuals w/ Multi-objective Particle Swarm Optimization
summary
A simple counterfactual generator uses multi-objective particle swarm optimization with a mixed discrete PSO variant, box constraints and a diversity preservation term to control proximity and sparsity, plus k-means selection to pick diverse solutions. It handles categorical features (one hot or label encoding), enforces immutability on non-actionable features, and produces multiple CFs (experiments generate five per query). Results are comparable to state of the art—better on some datasets but not uniformly superior—showing practical control rather than dramatic accuracy gains.
AI-written summary · about
chapters (9)
- 0:00Actionable counterfactuals intro
- 2:13Counterfactual explainability challenges
- 12:10Particle Swarm Optimization basics
- 14:21PSO for counterfactuals
- 19:15Objective functions in PSO
- 22:06Proximity and sparsity control
- 25:18Algorithm workflow & selection
- 26:49Experimental results
- 28:53Conclusion & future work
related talks
07mSwarm Intelligence Algorithms - Session Interview With Frances Buontempo
33mA Hands-On Introduction To Multi-Objective Optimization
10mMulti-objective Optimization for COVID Vaccine Management
50mOptimal control, no! Than what? (...)
14mSequential Monte Carlo Program Synthesis with Refinement Proposals
01h 15mAn Introduction to Swarm Intelligence Algorithms
34mImproved Decision Making with Pareto Fronts
30mEvolutionary Algorithms Perfecting the Art of "Good Enough"