How Alibaba’s Deep Learning Transformed CTR Prediction: From MLR to Multi‑Interest Networks

Jing Shi, a researcher at Alibaba‑Mama, introduced his background in computer vision, recommendation systems, and computational advertising, and outlined the agenda: the evolution of deep learning under internet‑scale data, its application in advertising and search, and future challenges.

Internet‑Scale Data Characteristics

Internet data is massive, high‑dimensional, and rich in internal relationships, exemplified by the many‑to‑many connections between users and items in e‑commerce.

Importance of CTR Prediction

CTR estimation is a core technology for advertising, recommendation, and search, directly influencing platform revenue and serving as a fertile ground for deep‑learning research.

Traditional CTR Methods

Two classic approaches dominate: manually engineered strong features combined with GBDT, and high‑dimensional sparse inputs fed into large‑scale logistic regression (a generalized linear model).

From Linear to Piecewise Linear Models

To capture non‑linearity, Alibaba‑Mama introduced a piecewise linear model (MLR) that partitions the feature space into many regions, each with its own linear predictor, smoothly combined to approximate complex functions.

End‑to‑End Deep Learning

Embedding vectors generated by MLR were fed into a multilayer perceptron, but without end‑to‑end training the gains were limited. Introducing end‑to‑end learning, where embeddings and the MLP are trained jointly, yielded significant CTR and GMV improvements, forming Alibaba‑Mama’s first deep learning network.

Modeling Multi‑Peak User Interests

Recognizing that user interests are multi‑modal, a deep network was proposed to extract relevant subsequences from behavior logs, forming multiple interest representations that better predict CTR for candidate items.

Challenges with Rich Media Features

Incorporating image features dramatically increases data volume, requiring novel distributed training architectures where parameter servers also host sub‑models to process images, reducing communication overhead.

Full‑Library Retrieval via Tree‑Structured Search

A hierarchical tree search reduces exhaustive evaluation of billions of items to a few dozen node evaluations, dramatically improving recall while maintaining deep‑learning‑driven relevance.

Future Directions

Open problems include handling missing labels for user experience metrics, improving recommendation diversity, and developing better evaluation beyond recall, such as new‑category recall and addressing recommendation self‑reinforcement loops.