Interaction-Aware Structured Pruning for Click-Through Rate Prediction Models with Data-Adaptive Noise Gating

Click-through rate (CTR) prediction models in real-time bidding (RTB) systems must remain accurate under strict latency budgets (about 20 ms for model inference). Explicit interaction architectures such as Deep & Cross Network V2 (DCN V2) combine large embedding tables with a cross-layer cost that grows quadratically with the input dimension. We propose Interaction-Aware Structured Pruning (IASP), which scores embedding dimensions by combining a Taylor-inspired importance score (magnitude and gradient) with each dimension’s participation in the cross network; a data-adaptive gate, whose threshold τ is a percentile of the interaction scores, restricts the synergy boost to dimensions with strong cross-layer participation. On the Avazu and Criteo datasets, IASP retains or improves the area under the ROC curve (AUC) at up to 99% embedding sparsity for DCN and DCN V2 and outperforms magnitude- and gradient-based criteria at 90–95% sparsity within a shared pruning pipeline. Physically removing the lowest-ranked dimensions (one dimension per field retained) reduces the memory footprint by about 16× and yields up to a 6.6× single-thread CPU speedup for DCN V2 when 200 candidates are ranked per request, whereas the speedup for single-candidate requests is marginal.