Trustworthy Intelligent Decision-Making: Framework, Counterfactual Reasoning, Complex Payoffs, Predictive Fairness, and Regulated Decisions

The past years of AI research have focused on predictive tasks, yet many real‑world scenarios require decisions that go beyond prediction. This talk introduces a trustworthy intelligent decision framework and explores four key sub‑directions: counterfactual reasoning, complex payoffs, predictive fairness, and regulated decision making.

1. Decision Framework – Emphasizes that decisions are more critical than predictions, illustrating the prevalence of decision problems in commerce, sociology, and other domains. It contrasts two common approaches: reinforcement learning (simulator‑based) and prediction‑driven decision making, and discusses their limitations, especially the challenge of building accurate simulators.

2. Counterfactual Reasoning – Describes three scenarios (off‑policy evaluation, counterfactual prediction, and policy optimization) and reviews traditional methods (direct method, propensity‑score weighting) as well as newer estimators such as Focused Context Balancing (FCB) that directly re‑weight samples to mitigate distribution shift.

3. Individual‑Level Effect Prediction – Highlights the need to model heterogeneous users, introduces VSR (Variational Self‑Representation) that learns latent factors for high‑dimensional treatments, and shows its superiority over sample re‑weighting methods in offline experiments.

4. Policy Optimization – Differentiates policy evaluation from true optimization, proposing OOSR (Outcome‑Oriented Sample Re‑weighting) that emphasizes regions near the optimal intervention and demonstrates significant gains on benchmark datasets.

5. Complex Payoffs – Argues that short‑term gains (e.g., discounts) may not translate to long‑term value, and stresses the importance of jointly optimizing short‑ and long‑term rewards by understanding consumer choice models.

6. Predictive Fairness – Reviews demographic parity (DP) and equalized odds (EO), points out their shortcomings, and introduces conditional fairness, which balances fairness constraints with useful predictive variables. The DCFR algorithm is presented as a method that achieves a better trade‑off between accuracy and fairness.

7. Regulated Decision Making – Discusses personalized pricing and the risk of excessive profit extraction, proposing regulatory instruments (price caps, ratio limits) that embed constraints into the payoff function to protect consumer surplus while maintaining efficiency.

Overall, the framework demonstrates that trustworthy decision making requires integrating causal inference, fairness, long‑term optimization, and regulatory considerations, and it references several recent papers from KDD, NeurIPS, ICML, and WWW that substantiate each component.