Adversarial Adaptive Framework for Cold-Start Cross-Domain Recommendation

The paper, based on the CIKM 2022 publication "Cross-domain Recommendation via Adversarial Adaptation," addresses the critical role of recommendation systems in online advertising, focusing on click‑through rate (CTR) and conversion rate (CVR) prediction using deep neural networks.

When user‑item interaction data are sparse, recommendation performance degrades, prompting the use of cross‑domain recommendation (CDR) to leverage abundant source‑domain data for a target domain. Existing CDR methods struggle with cold‑start scenarios and domain shift, where source and target data distributions differ.

The authors identify two main challenges in cold‑start CDR: (1) domain shift between source and target domains, and (2) severe data imbalance caused by a scarcity of positive samples in the target domain.

To tackle these issues, they propose an adversarial adaptive learning framework that aligns source and target feature distributions via a domain discriminator, producing domain‑invariant features and transferring knowledge from a pre‑trained source expert network to the target model.

The framework consists of three components—source expert network, target model, and discriminator—and follows four training stages: (1) pre‑train the source model on abundant source data, (2) adversarially align source and target feature distributions, (3) calibrate the target model using supervised target data to retain task‑specific information, and (4) inference on the target domain.

Loss functions for CTR and CVR tasks are defined using standard cross‑entropy and multi‑task formulations, with additional adversarial and gradient‑penalty terms for the discriminator.

Domain shift is mitigated by adversarial learning with a domain discriminator.

Target‑domain data imbalance is alleviated by leveraging source‑domain distribution via the expert network.

Task‑specific calibration preserves target‑domain information despite alignment.

Extensive experiments on real business datasets combine the proposed method with nine popular single‑domain CTR/CVR models (e.g., DeepFM, NFM, DCN). The adversarial adaptive approach consistently outperforms baselines such as target‑only training, cross‑test, mixed training, and fine‑tuning.

t‑SNE visualizations show that after applying the framework, target‑domain feature encodings align closely with source‑domain encodings, whereas mixed‑data training leads to chaotic feature distributions.

In conclusion, the adversarial adaptive framework effectively addresses cold‑start cross‑domain recommendation challenges—domain shift, positive‑sample scarcity, and data imbalance—by aligning domains and transferring knowledge, achieving significant performance improvements across multiple CTR and CVR tasks.