Advances in Customer Lifetime Value Prediction for Online Advertising: Missing-aware Routing Fusion Network and Cross-domain Adaptive Learning

Feature Missing-aware Routing-and-Fusion Network for Customer Lifetime Value Prediction in Advertising (MarfNet) addresses the pervasive feature‑missing and label‑sparsity problems in LTV prediction for mobile‑game advertising. The method computes missing states of raw features and feature interactions, routes samples to specialized expert sub‑networks, and fuses their outputs into a missing‑aware representation. To further mitigate label sparsity, a batch‑in‑dynamic discriminative enhancement (Bidden) loss weighting scheme automatically assigns larger loss weights to hard samples. Offline experiments and online A/B tests on Tencent's Game Yuke Win platform demonstrate superior performance, influencing billions of daily impressions.

Cross-domain Adaptive Learning for Online Advertisement Customer Lifetime Value Prediction (CDAF) tackles data scarcity across domains by pre‑training an LTV model on a data‑rich related platform, then aligning domain‑invariant user representations via minimizing Wasserstein distance between source and target encodings. A dual‑predictor design preserves both domain‑invariant semantics and domain‑specific information, enhancing target prediction accuracy. Offline evaluation on real Tencent Game Yuke Win data shows notable AUC gains (e.g., +13.7% on DCNv2 for dataset G2).