Data-Driven Uncertainty Revenue Modeling for Computation Resource Allocation in Recommendation Systems

In recent years, as the resource consumption of computation-intensive recommendation systems (RS) significantly increased, and the supply of large-scale resources encountered bottleneck, computation resource allocation for improving computing efficiency grabbed the attention of the industry. The simulation of revenue is a focal point in the allocation problem. However, due to the complex engineering architecture of RS, no existing research proposes a simulation model that addresses the relationship between resource allocation strategies and benefits. This paper, based on real data from Alipay's advertising RS and integrating queuing theory, models the relationship from resource allocation decisions to revenue considering exposure randomness of traffic. We further merge allocation tasks with capacity planning (CP) to establish a two-stage joint optimization model and use the revenue model above as the objective. The proposed model outperforms the baseline with 1.9% in revenue, and represents flexible adaptation of exposure rate, providing insights for the simulation of industrial RS.