Why do ranking systems need to model selection bias explicitly?

Industrial ranking systems face two distinct problems that interact. First, objectives conflict: engagement (clicks, watch time) and satisfaction (ratings, likes, shares) are not the same thing, and naive aggregation collapses them. YouTube's solution uses Multi-gate Mixture-of-Experts so each objective can choose which input experts it shares with others — soft parameter sharing rather than full-shared or fully-separate models.

Second, and more insidious: training data comes from logs of the current ranker. A user clicked a video because it was placed at position 1, not because they preferred it. Train on that data and you reinforce whatever the ranker did before — a positive feedback loop where the model keeps learning what it has already taught itself. The Wide & Deep extension here adds a shallow tower whose only job is to model position bias, factoring out the rank-induced effect from the engagement signal.

Two mechanisms because two failure modes: MMoE for objective conflict, shallow position tower for selection bias. Without explicit treatment of either, the model converges on a degenerate equilibrium.

RL-side echo — the same multi-objective problem, the same Pareto framing. DVAO confronts the recommender world's first problem (conflicting objectives) inside multi-reward GRPO for LLMs: accuracy, length, and format all push at once, and naive scalarization either explodes advantage magnitudes (Reward Combination) or ignores cross-objective correlation with static weights (Advantage Combination). Its fix — dynamically weighting each objective by its empirical reward variance within a rollout — is the RL analog of MMoE's soft per-objective parameter sharing: instead of a fixed mixing, let each objective's contribution adapt to where the live learning signal is. Both literatures converge on the same goal, a superior Pareto frontier across objectives rather than a single scalarized peak, and both reject fixed combination weights as the source of degeneration. The recommender's second problem — selection bias from logging the current policy — has no clean DVAO counterpart, but it rhymes with RLVR's shortcut-amplification: in both, training on data the current model generated reinforces whatever the model already did. DVAO does not address that feedback loop, which marks the boundary of the analogy.