Key Insights on Causal Inference: Motivation, Applications, Challenges, and Links to A/B Testing, ML, and Deep Learning

01 Why Do Causal Inference

One driver is academic research: Professor Cui Peng at Tsinghua noticed that deep learning, despite its success, cannot solve problems such as out‑of‑distribution (OOD) issues, fairness, interpretability, and actionability, all of which can be framed as causal analysis. Another driver is industry practice: prediction models often suffer from instability and poor generalisation, and businesses need incremental impact rather than just predicting Y from high‑dimensional X. Causal inference helps identify the true effect of interventions, enabling cost‑effective targeting.

02 Typical Application Scenarios

Causal inference is used mainly for two categories of problems: prediction (more accurate, stable, and interpretable) and decision‑making (pricing, logistics, recommendation, marketing). These are essentially counterfactual questions—what would happen if we intervene versus if we do not—often under business constraints such as risk or cost.

1. Intelligent Decision‑Making – In marketing, uplift models estimate the incremental conversion caused by a specific outreach strategy. Companies abstract uplift into a general framework that combines causal effect estimation with optimization algorithms for growth scenarios.

2. Prediction Scenarios – In recommendation systems, when training and inference are not i.i.d., causal techniques are used to debias models, integrating causal reasoning into machine learning pipelines.

03 Challenges of Deploying Causal Inference

Validation is the biggest obstacle. Causal inference relies on untestable assumptions and observational data, making it hard to prove correctness. It requires close collaboration with business to perform case‑by‑case analysis, unlike standard ML models that can be applied generically.

Data quality is another issue: randomized controlled trial (RCT) data are scarce, and combining RCT with observational data, or adapting causal methods to big‑data environments, remains challenging.

The pragmatic solution is to focus on concrete business problems and evaluate causal effects using A/B tests where possible.

04 Causal Inference and A/B Testing

1. Relationship – A/B testing is a subset of causal inference; its core value is to determine causal impact. While A/B testing is mature, there is still room for improvement in speed, accuracy, and cost.

2. Priority – When an A/B test can be run, it is preferred because it provides a gold‑standard causal estimate. However, causal inference often lacks universal metrics like AUC, so A/B tests are used to validate causal predictions.

3. Differences – A/B tests are expensive and explore a limited candidate space. The trend is to relax constraints, reduce cost, and expand policy optimization, moving from evaluation to automation and learning, where causal inference can deliver user‑level decision insights.

05 Causal Inference, Machine Learning, and Deep Learning

1. Mutual Influence – Since 2016‑2017, ML and causal inference have begun to intersect. Introducing causal reasoning into ML can enhance stability and generalisation, especially under OOD conditions, while ML provides tools to handle high‑dimensional data.

2. Priority – Current efforts focus on embedding causal concepts into existing ML frameworks rather than rebuilding systems from scratch, because replacing established recommendation or search pipelines is costly.

3. Causal and Deep Learning – Representation learning is a promising direction, e.g., separating confounders in latent space, which helps apply causal ideas within deep learning models.

--- IID: Independent Identically Distributed OOD: Out Of Distribution ML: Machine Learning Causal Inference: 因果推断 Actionability: 可行性 Validation: 验证 RCT: Randomized Controlled Trial