A ranking method separates engagement from its baseline

When a Feed score predicts likes, how much reflects a particular post and how much reflects a viewer's usual liking? If a score mixes both, should ranking remove the viewer baseline, or keep it as a separate signal whose influence can be adjusted?

The paper adds a separately weighted estimate of a candidate's contribution above a baseline while retaining its original engagement prediction.

The paper asks how a Feed ranking score should handle an engagement prediction that reflects both a candidate and a baseline, such as a viewer’s usual behavior. Its method keeps the raw prediction and adds a separately weighted estimate of how much the candidate exceeds a viewer, item, or request baseline.

In four-arm live tests of six deployed Feed terms, the combined score had higher point estimates than either term alone in 11 of 13 judged term-and-metric pairs; 17 of 26 direct comparisons significantly favored both, while one significantly went against it.

Because these comparisons tested each term at production weights, they support the conclusion that neither term was generally redundant at that tested operating point.

The six tested terms are a case series, not a representative sample, and the study does not establish a general link between the offline diagnostic and online results. The ablation tests deployed weights; it does not show whether a retuned single-term model could match the combined score.