An Empirical Comparison of High-Order Feature Interaction Operators for Conversion Rate Prediction in Sparse, High-Cardinality Message-Ads Traffic: Accuracy, Efficiency, and Offline--Online Consistency

Post-click conversion rate (CVR) prediction on message-ads traffic exposes feature interaction operators to an extreme regime of sparsity, label imbalance, and serving-latency constraints. While a decade of recommender research has produced an abundance of operators that differ in their treatment of explicit versus implicit, low-order versus high-order interactions, published comparisons typically optimize for click-through rate on dense public logs and seldom isolate the operator from confounding training pipelines. This study conducts a controlled empirical comparison of seven high-order interaction operators---plain MLP, FM, DeepFM, DCN, DCN-V2, xDeepFM, and AutoInt---across Criteo, Avazu, and Ali-CCP under a unified training protocol. We measure offline AUC and LogLoss, per-sample parameters, FLOPs, and inference latency, and further stratify AUC by user-activity quantile and by categorical-feature density. On Ali-CCP CVR, DCN-V2 attains the highest AUC (0.6289) while DCN matches it within 0.0011 AUC at 0.83× the latency; xDeepFM's compressed interaction component contributes the largest efficiency penalty without a proportionate accuracy gain. Rank correlation between offline AUC and an online CVR proxy drops from 0.93 on high-activity users to 0.41 on cold-start users, echoing documented offline--online inconsistencies. The findings provide operator-selection guidance grounded in measured efficiency and subgroup stability rather than on headline AUC deltas.