ADSNet: Cross-Domain LTV Prediction with an Adaptive Siamese Network in Advertising

At KDD 2024, the Tencent Advertising Technology team presented a paper titled "ADSNet: Cross-Domain LTV Prediction with an Adaptive Siamese Network in Advertising" addressing the challenge of sparse paid conversion data in ad systems and the need for personalized user value estimation to align bidding with advertiser ROI goals.

The authors define user Lifetime Value (LTV) as the revenue contributed by a user over a period and formulate the pLTV prediction task as estimating LTV for a given pair. They highlight the difficulty caused by sparse conversion data and the negative transfer that can arise when simply aggregating multi‑domain data.

To overcome these issues, the paper proposes (1) expanding the paid sample pool by incorporating external‑channel data and (2) a novel Adaptive Difference Siamese Network (ADSNet) that evaluates information gain to reject noisy samples and improve cross‑domain transfer.

ADSNet consists of three parts: a pseudo‑siamese network that measures gain between source and target domains, a gain‑measurement strategy that quantifies the benefit of external samples, and a domain‑adaptation module that bridges distribution gaps at both embedding and tower levels using knowledge distillation.

The backbone model is an ordinal‑classification‑based multi‑expert network with an encoder, MoE expert layer, and a tower that jointly predicts payment probability and multi‑granular payment amount, handling the long‑tail and multi‑modal nature of LTV distribution.

During inference, the final pLTV is computed by combining the predicted payment probability with the weighted average LTV of each ordinal segment, using binary cross‑entropy losses for both probability and amount predictions.

Training objectives combine losses from the gain network, the original network, and the domain‑adaptation network, with an iterative alignment strategy that first warms up on internal data and then jointly trains with external data while periodically syncing parameters.

Extensive offline experiments and online A/B tests demonstrate that ADSNet outperforms several SOTA LTV models, especially when external data is used, mitigating negative transfer and significantly improving performance on long‑tail samples (up to 15.2% GINI gain).

Ablation studies confirm the effectiveness of each ADSNet component, and visualizations show the model progressively learning to reject negative‑gain samples during training.

The paper concludes that leveraging external‑channel data together with the adaptive siamese architecture effectively addresses data sparsity and cross‑domain challenges in advertising LTV prediction.

References include recent works on large‑scale LTV prediction, multi‑task learning, cross‑domain recommendation, and related machine‑learning techniques.