CAN: Revisiting Feature Co-Action for Click-Through Rate Prediction

Author Zhou Guorui, a senior algorithm expert at Alibaba, presents the latest work of the Alibaba targeted advertising team on click‑through‑rate (CTR) modeling, titled CAN (Co‑Action Net), which has been deployed before the 2021 Double‑11 shopping festival with significant online gains.

The article narrates the development process in a chronological “charge” style, describing three major attempts: an early parameter‑bank idea for shared item embeddings, a “res‑embedding” approach using graph‑based similarity, and extensive experiments with feature‑cross methods that proved ineffective at large scale.

Observing that traditional feature‑cross models (FM, DeepFM, DCN, etc.) failed to capture item‑item co‑action information in massive e‑commerce data, the team explored Cartesian‑product embeddings, which model the joint probability P(Y|A,B) by creating a new ID for each item‑item pair, achieving better performance but suffering from huge parameter space, training/serving cost, and sparsity.

To overcome these drawbacks, the authors propose CAN, which treats one ID as input and the other as parameters of a multi‑layer perceptron (MLP). By using T slots per item and attention mechanisms, the model captures co‑action while keeping the number of queried IDs constant, thus reducing memory and CPU overhead.

Experiments on both internal Alibaba datasets and public benchmarks show that CAN outperforms the Cartesian‑product baseline with comparable or lower model size and acceptable online latency, leading to full production rollout and noticeable business growth.

The work highlights the importance of modeling input‑side interactions beyond simple feature crosses and suggests future directions for more efficient co‑action representations.