A new arXiv paper, announced as arXiv:2610.27423, addresses bias in Instagram's feed ranking. The authors describe a feed ranker that scores candidates using a value model, which is a weighted sum of engagement predictions. The problem, they argue, is that each prediction conflates two distinct quantities: a baseline — what the viewer, item, or request would produce anyway — and the incremental amount contributed by the specific ranking decision.
The paper proposes a decomposition of these predictions to separate the baseline from the incremental effect. This separation is intended to improve debiasing across three dimensions: user, item, and request. By isolating the incremental impact, the ranker can better estimate the true causal effect of showing a particular item in a particular context, rather than being skewed by pre-existing tendencies.
The source is an abstract only, so details on methodology and results are not yet available. The significance lies in the explicit framing of ranking predictions as a mix of baseline and incremental signals, which is a common challenge in recommendation systems. The work suggests a path toward more accurate and fair feed ranking by untangling these components.