Causal Inference Knowledge Map: Framework, Application Evaluation, Typical Algorithms, Practical Challenges, and JD Technology Case Study

Introduction – The talk, titled “Causal Inference Knowledge Map,” was delivered by Wang Dongdong, an algorithm engineer at JD Technology.

1. Overall Framework

The main content is organized into six parts: overall framework, application‑scenario evaluation (decision problems), typical causal algorithms, practical challenges of causal inference deployment, a JD Technology credit‑limit decision model case, and future development.

2. Application‑Scenario Evaluation (Decision Problems)

Evaluating whether a scenario suits causal inference starts with clarifying the decision problem: the action, constraints, and the objective to be maximized. The need for a causal model depends on whether the action influences the objective and whether data conditions satisfy the model’s assumptions (e.g., overlap, stable unit treatment value).

Key assumptions for potential‑outcome models include stability of individual causal effects, independence of treatment and potential outcome given features, and overlap (each user must have a non‑zero probability of receiving each decision).

Structural causal models assume known causal relationships, which are often hard to verify.

Understanding prior knowledge, the data‑generation mechanism, and business experience is crucial for feature engineering and handling bias.

3. Typical Causal Algorithms

Algorithms are grouped into three categories:

Causal structure discovery (graph‑based conditional independence, score‑based methods, additive‑noise models).

Causal effect estimation – including instrumental variable methods, DID, synthetic control, propensity‑score based methods (matching, stratification, weighting), doubly robust and double machine learning, meta‑learning (S‑learning, X‑learning), tree‑based uplift models, and causal representation learning.

Meta‑learning approaches that treat treatment as a feature and learn separate models for treatment/control groups.

4. Practical Deployment Challenges

Weak causal signals mixed with noise make modeling difficult and risk over‑fitting.

Insufficient data conditions – violations of overlap, lack of randomization, and scarcity of test data hinder reliable evaluation.

Multi‑treatment decisions increase complexity; continuous treatment modeling may require deep‑learning‑based functional assumptions.

Fixed allocation mechanisms and high cost of random testing limit model validation.

Target prediction often involves many factors; multi‑task learning can simplify but still faces difficulty.

5. Case Study – JD Technology Credit‑Limit Decision Model

The case demonstrates how JD Technology uses causal inference to predict user behavior under different credit limits and then selects the optimal limit based on business goals such as profit and scale.

6. Future Development

Future research directions include building large‑scale models for complex non‑linear relationships, advancing causal representation learning for better modularity and transferability, and designing more flexible algorithms that relax strong assumptions to lower deployment barriers.

Overall, the presentation provides a detailed roadmap for applying causal inference in decision‑making systems, from theory to practice.