AB Incremental Evaluation and Contamination Mitigation in Social Viral Experiments

AB testing serves as a critical compass for evaluating strategy effects in data‑driven businesses, especially when assessing the combined impact of multiple strategies on overall metrics.

The article first defines AB increment, shows how to compute DAU increment from per‑user visit rates, and explains that increment reflects both per‑user efficiency and total scale.

It then analyzes how social viral experiments cause contamination: in fully‑strategy experiments contamination is unidirectional (affecting only the control group), while in non‑fully‑strategy experiments it becomes bidirectional, potentially biasing increment estimates upward or downward.

To correct this bias, four probability‑estimation schemes are presented: exponential smoothing for stable traffic, same‑period expansion coefficient for bursty promotional periods, overall expansion coefficient when historic data are lacking, and homogeneous‑group sampling (including deterministic target sampling and PSM‑based sampling) that matches contaminated users with similar untreated peers.

The conclusion advises selecting the appropriate scheme according to the experimental design, viral mechanism, strategy scale and business characteristics to obtain a scientifically sound AB increment evaluation.