Evolution of Alibaba's Advertising CTR Prediction Models: From Linear Methods to Deep Interest Evolution Networks

The talk begins by describing the unique traits of e‑commerce data personalization, where ads are displayed based on inferred user intent rather than explicit queries, and outlines three key dimensions—explicit content, account‑level IDs, and final feedback—to characterize the data.

It then traces Alibaba Mama's model evolution since 2012, starting with large‑scale feature‑rich linear regression (MLR) and its non‑linear extension (MLR with piecewise linear learning), followed by the first generation of deep CTR models that combine embedding layers with simple MLPs, achieving significant performance gains.

To address the diversity of user interests, the Deep Interest Network (DIN) introduces an attention‑based activation unit that dynamically selects relevant historical behaviors, improving CTR by 10%, CVR by 3.3%, and GPM by 12.6%.

Subsequent work incorporates visual features through the CrossMedia network, handling massive image data by storing compressed embeddings on remote servers.

Recognizing the limitations of sequential RNNs for heterogeneous interests, the Deep Interest Evolution (DIE) framework adds an interest extraction layer (attention‑augmented GRU) and an interest evolution layer, enabling the model to capture evolving user preferences over time.

To balance model complexity and online latency, a teacher‑student distillation approach is employed, featuring collaborative training, parameter sharing, and gradient blocking to compress the model while preserving performance.

The presentation concludes with two future research directions: developing disentangled representations tailored to e‑commerce items, and building more interpretable white‑box models that expose the concepts influencing user decisions.

Author bio: Zhou Guorui, Alibaba algorithm expert, Ph.D. candidate at Beijing University of Posts and Telecommunications, focuses on large‑scale machine learning, NLP, computational advertising, and recommendation systems; core developer of Alibaba's XDL deep learning framework.