Causal Inference Knowledge Map: Framework, Application Evaluation, Typical Algorithms, Implementation Challenges, and JD Tech Credit Decision Model

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

Main Contents 1. Overall framework 2. Application‑scenario evaluation (decision problems) 3. Typical causal algorithms 4. Practical difficulties in deploying causal inference 5. Case study – JD Technology’s credit‑limit decision model 6. Future development

Causal Inference Tasks Causal inference mainly involves three tasks: (a) discovering causal structure (identifying causal relations among variables), (b) estimating causal effects (quantifying the impact of one variable on another under intervention), and (c) correcting bias when training and application data distributions differ.

Application Scenarios The technique is especially useful in decision‑making contexts, industrial processes with complex, long‑running workflows, and situations demanding robust and interpretable models. It also supports post‑decision effect evaluation.

Evaluating Scenario Suitability A decision problem must be clearly defined: the action, constraints, and the objective to be maximized. For example, in coupon marketing, one must decide whether to issue coupons under a budget constraint to maximize sales. If the action’s effect on the objective is clear and data satisfy the assumptions of causal models (e.g., stable unit treatment value, ignorability, overlap), causal inference can be applied.

Typical Causal Algorithms • Causal structure discovery: constraint‑based methods, score‑based methods (e.g., maximizing likelihood of a DAG), and additive‑noise‑model approaches. • Effect estimation: instrumental variable (IV) methods, Difference‑in‑Differences (DID), synthetic control, propensity‑score based methods (matching, stratification, weighting), doubly‑robust and double‑machine‑learning approaches, meta‑learning methods such as S‑learning and X‑learning, tree‑based uplift models, and causal representation learning.

Practical Challenges 1. Weak causal signals often comparable to random noise, requiring high‑capacity models while avoiding over‑fitting. 2. Insufficient data conditions: violations of overlap, lack of randomization, and scarcity of test data. 3. Multi‑treatment decision problems demand decomposition or deep‑learning‑based continuous treatment modeling. 4. Fixed allocation mechanisms and high cost of random testing make model validation difficult. 5. Complex target prediction may need multi‑task learning or simplified key‑metric prediction.

Case Study – JD Technology Credit‑Limit Decision Model Using causal inference, JD Tech predicts user behavior under different credit limits, aligns predictions with business goals (profit, scale, user activity), and determines the optimal credit limit for each user through a two‑step process: causal prediction followed by goal‑oriented optimization.

Future Development Future research should address the limitations of current causal models by building large‑scale, non‑linear models, advancing causal representation learning for better modularity and transferability, and reducing overly strong assumptions to create more universally applicable algorithms.

Conclusion The presentation summarized the above topics and thanked the audience.