Platform Attention Allocation under Adolescent Welfare Constraints: A Multi-Objective Mixed-Integer Programming Approach

Video platforms allocate recommendation exposure primarily to maximize engagement, potentially creating behavioural welfare risks for adolescent users. Existing safeguards, such as screen-time reminders and forced interruptions, are largely ex post and do not alter the underlying allocation logic. This study formulates adolescent video recommendation as a constrained allocation problem and develops a multi-objective mixed-integer linear programming (MOMILP) framework that jointly incorporates predicted commercial utility, modeled behavioural risk, educational exposure, and content diversity. Using controlled synthetic experiments, this study compare the proposed framework with a myopic commercial baseline and a post-hoc hard-filtering baseline. Under the reported parameter setting, the proposed model reduces aggregate modeled risk by 58.9% relative to the commercial baseline while retaining 68.5% of its predicted commercial value at a nearly identical delivery cost. Compared with post-hoc filtering, joint optimization retains substantially more commercial value while coordinating multiple governance objectives within the allocation process. Sensitivity, solver-gap, and diversity analyses further demonstrate the trade-offs among commercial utility, modeled risk, computational performance, and exposure structure. These findings illustrate how adolescent-protection objectives can be embedded directly into recommendation allocation rather than added only as post-exposure controls. Results are based on synthetic data and should not be interpreted as causal estimates of real-world welfare improvement.