Applying Causal Inference to Limited‑Resource Decision‑Making

Introduction The talk introduces causal inference and its relevance to intelligent decision‑making under limited resources.

What is Causal Inference Causal inference seeks to estimate the effect of changing a variable (treatment) on an outcome, unlike correlation modeling which only predicts Y given X. An example shows how ice‑cream sales and drowning incidents are correlated due to a hidden confounder (temperature), illustrating the danger of mistaking correlation for causation.

Correlation vs. Causal Modeling Correlation models answer "what will Y be given X?" while causal models answer "how will Y change if we intervene on T?". A high‑AUC classifier does not guarantee accurate treatment‑effect estimation.

Limited‑Resource Decision Scenarios Examples include the classic knapsack problem, ride‑hailing subsidy allocation, and airline ticket pricing. Each scenario has a total resource constraint (capacity, budget, seats) and a decision variable (select/not select, subsidy amount, price) that influences the objective.

Technical Abstraction Two main components are needed: prediction services (target and cost prediction) and operations optimization (filtering, pruning, clustering candidate solutions and solving a constrained optimization problem).

Causal Inference Techniques for Intelligent Decisions Four major families of models are discussed: 1. Causal Forest – builds multiple causal trees to maximize heterogeneity of treatment effects. 2. Meta‑Learners – T‑learner, S‑learner, X‑learner, R‑learner, each combining separate models for control and treatment groups. 3. Representation Learning – uses deep networks to implement meta‑learner modules. 4. Double Machine Learning (DML) – relies on strong orthogonality assumptions between treatment and nuisance components.

Experimental Design Principles Three criteria for identifiable causal effects: (1) Exchangeability (homogeneous groups), (2) Positivity (non‑zero treatment probability), (3) SUTVA (no interference between units).

Evaluation Metrics Commonly used metrics are QiniScore (area between model‑sorted ITE curve and random), AUUC (normalized QiniScore using ATE), and AUCC (area under cost‑curve).

Feature Selection Guidelines - Prefer subtraction over addition of features. - Avoid instrumental variables in exploratory experiments. - Properly decompose confounders, adjustment, and instrumental variables. - Do not introduce collider or post‑treatment variables.

Future Work & Discussions Open challenges include causal inference when random experiments are prohibited by law, handling high‑dimensional treatment spaces, violations of SUTVA (network effects), and weak uplift signals where models struggle to learn reliable ITEs.

Conclusion Causal inference provides a principled framework for turning limited‑resource constraints into optimized, data‑driven decisions across various business domains.