M-scan: A Multi-Scenario Causal-driven Adaptive Network for Recommendation

We primarily focus on the field of multi-scenario recommendation, which posesa significant challenge in effectively leveraging data from different scenariosto enhance predictions in scenarios with limited data. Current mainstreamefforts mainly center around innovative model network architectures, with theaim of enabling the network to implicitly acquire knowledge from diversescenarios. However, the uncertainty of implicit learning in networks arisesfrom the absence of explicit modeling, leading to not only difficulty intraining but also incomplete user representation and suboptimal performance.Furthermore, through causal graph analysis, we have discovered that thescenario itself directly influences click behavior, yet existing approachesdirectly incorporate data from other scenarios during the training of thecurrent scenario, leading to prediction biases when they directly utilize clickbehaviors from other scenarios to train models. To address these problems, wepropose the Multi-Scenario Causal-driven Adaptive Network M-scan). This modelincorporates a Scenario-Aware Co-Attention mechanism that explicitly extractsuser interests from other scenarios that align with the current scenario.Additionally, it employs a Scenario Bias Eliminator module utilizing causalcounterfactual inference to mitigate biases introduced by data from otherscenarios. Extensive experiments on two public datasets demonstrate theefficacy of our M-scan compared to the existing baseline models.