TAST: Task-Aware Sparse Topology Learning for Multi-Task Recommendation

Multi-task recommendation improves predictive performance by sharing knowledge across related tasks, but existing dense architectures and predefined expert-routing mechanisms do not explicitly adapt fine-grained parameter connectivity to individual tasks, potentially leading to parameter competition, negative transfer, and deployment overhead. We propose TAST, a Task-Aware Sparse Topology Learning framework that constructs task-specific subnetworks within a shared network. A task-conditioned topology generator jointly learns neuron- and connection-level masks, while explicit density budgets discourage trivial dense or excessively sparse solutions and enable controllable capacity allocation. TAST first learns shared parameters and task-specific topologies, hardens the learned masks into binary structures, and then optionally applies a lightweight sample-conditioned refinement module for local adaptation. Validation-guided branch selection retains the refined branch only for tasks with sufficient validation gains, allowing unused refinement branches to be removed during deployment. Experiments on four AliExpress subsets and the UserBehavior dataset show that TAST achieves the highest observed mean AUC across all five datasets, with statistically significant improvements over the fixed MoME comparator on AliExpress-US and UserBehavior after multiple-comparison correction. On AliExpress-NL, TAST achieves a slightly higher mean AUC than MoME while reducing deployment parameters and measured batch-inference latency by 87.7% and 77.0%, respectively. The current implementation still evaluates the hardened masked weights using dense GPU operators; therefore, the learned sparsity represents structural adaptation rather than physical sparse execution. These results demonstrate a favorable balance among predictive accuracy, structural adaptation, and deployment efficiency.