How Offline Causal Inference Unlocks 3D Product Value on Taobao

Background and Motivation

Traditional statistics often reveal correlations—such as chocolate consumption and Nobel laureates—but cannot establish causation because experiments on many real‑world factors are infeasible. Causal inference aims to answer "what would happen if we intervene" by controlling confounding variables and estimating treatment effects from observational data.

Fundamental Concepts

Causal inference has two main frameworks: the causal graph model (Judea Pearl) for identifying causal structure, and the Potential Outcomes (Rubin) model, which underlies tools like Propensity Score Matching (PSM) and instrumental variables.

Treatment : whether a user receives an intervention (e.g., a push notification).

Potential outcomes : the outcomes that would occur under treatment and under control for the same individual.

Observed outcome : the realized outcome for the actual treatment received.

Confounders : variables that affect both treatment and outcome and must be balanced.

Key Causal Methods

Propensity Score Matching (PSM) : summarizes multiple confounders into a single score and matches treated units with control units of similar scores to create comparable groups.

Bayesian Causal Graphs : computes entropy‑based relationships between variables to infer directed causal edges, helping to pinpoint root causes in complex behavior chains.

Difference‑in‑Differences (DID) : compares changes over time between a treatment group and a control group, assuming parallel trends, to estimate average treatment effects.

Illustrative Example: Push Notification Impact

For users ID 1‑3 a push is sent (T=Y) and for users ID 4‑6 it is not. Potential outcomes Y1 and Y0 illustrate the counterfactual gap. Observed outcome is the actual reading count (9) for the treated user; the untreated outcome (7) remains unobserved. Using the “god‑view” we compute ATE = 1.2 and ATT = 1.7, confirming the push increases reading counts.

When and How to Apply Causal Inference

AB testing is ideal when randomization is possible, but many scenarios (e.g., smoking effects, 3D model adoption) cannot be randomized or are too costly. Offline causal inference provides a scientific way to predict experiment outcomes and prioritize high‑impact interventions.

Case Study: Offline Causal Inference for Taobao 3D Sample Rooms

Business Context

Taobao’s 3D sample rooms aim to boost conversion by offering immersive product experiences. However, low model coverage and limited merchant participation hindered growth.

Analysis Framework

The team built a value‑analysis framework and applied three causal methods:

PSM : matched 20+ user and preference features to create a control group. Results showed the treatment group achieved +24.85% add‑to‑cart rate, +27.68% dwell time, +29.53% average order value, +5.98% traffic, and a –5.75% decision cycle.

Bayesian Causal Graph : constructed a graph from 10+ in‑room actions and 20+ user attributes, identifying “home ownership” as a key driver and highlighting gaps in the new‑user guide funnel.

DID : combined with PSM, the DID estimate indicated a +6.73% add‑to‑cart rate, +1.26 additional items, and +17.26 minutes of session time over three weeks.

Operational Recommendations

Target users with the “homeowner” label for precise campaigns and improve the guide flow beyond the “room switch” step, which later raised add‑to‑cart rate by 28.93%.

Conclusion

Observational causal inference methods—PSM, Bayesian graphs, and DID—provide complementary tools for quantifying product impact when randomized experiments are infeasible. Combining multiple techniques and domain expertise yields robust, actionable insights, as demonstrated by the 3D sample‑room analysis.

References

Judea Pearl & Dana Mackenzie, The Book of Why .

Hernán & Robins, Causal Inference: What If .

王乐, “概率图模型之贝叶斯网络”.

Various academic papers on PSM, DID, and causal mechanisms.