Problem

Watch-time feedback in video recommendation is confounded by video duration. Longer videos mechanically accumulate more watch time, so prediction models drift toward duration effects instead of true preference. Existing backdoor-adjustment approaches discretize duration into groups and combine group-wise models, which the authors argue is prone to misspecification under feature distribution shift across groups.

Approach

Duration-Invariant Feature Learning (DIFL) reinterprets the debiasing task as invariant learning. A kernel-based regularizer forces learned representations to stay invariant across duration groups, reducing sensitivity to how the groups are drawn and supporting counterfactual watch-time prediction.

Public evidence

Experiments on public benchmarks and a large-scale production dataset report state-of-the-art performance against duration-bias baselines. The public abstract does not give numeric metrics, so treat the claim as benchmark-level evidence rather than measured online lift.

Takeaway

Invariant learning gives the counterfactual watch-time line a way to separate preference signal from duration conditions without committing to one discretization of time.

Bibliographic details

Authors

Chenghou Jin, Yixin Ren, Hongxu Ma, Yewei Xia, Yi Guan, Hao Zhang, Jiandong Ding, Jihong Guan, Shuigeng Zhou

Venue

AAAI 2026

Status

Published