Recent Advances in Advertising Recommendation Algorithms and Their Applications

In recent years, advertising recommendation algorithms have made significant strides, evolving from foundational sequence learning, large‑scale graph learning, online and reinforcement learning, to multimodal recommendation, with both academic and industrial impact.

The JD.com advertising team emphasizes practical deployment, improving matching efficiency, user experience, and ecosystem health, and has presented related research at CIKM 2023.

Key developments in ranking algorithms include the rise of deep learning methods (CNN, RNN, Transformer), advanced sequence modeling (RNN, LSTM), self‑supervised learning for label generation, online learning and reinforcement learning for real‑time adaptation, multimodal recommendation integrating text, image, and video, and growing attention to explainability and fairness.

Breakthrough 1 – More Efficient Learning: An Incremental Update Framework with Data‑Driven Prior (DDP) introduces Feature Prior (FP) and Model Prior (MP) to stabilize online recommender updates, integrating feature‑level CTR estimates and Bayesian model priors, demonstrating superior performance on public and industrial datasets.

Breakthrough 2 – Finer Modeling: Dynamic Group Parameter Modeling (DGPM) automatically discovers user groups and generates group‑specific parameters, using a group information selection module, memory‑based group representation learning, and soft group assignment to enhance CTR prediction accuracy.

Breakthrough 3 – Better Interaction Capability: Bilateral Interactive Graph Convolutional Network (BI‑GCN) introduces early fusion between user and item trees via bilateral interaction, improving recommendation performance over traditional GCN‑based methods.

Breakthrough 4 – More Aesthetic Intelligence: A Relation‑Aware Diffusion Model for controllable poster layout generation combines visual‑text relationship awareness (VTRAM) and geometric relationship awareness (GRAM) within a diffusion framework, achieving state‑of‑the‑art results on the CGL‑Dataset V2.

Extensive experiments on public and industrial datasets validate each method, showing significant gains in CTR, RPM, and layout quality, and the paper outlines future challenges such as diversity vs. personalization, explainability, real‑time generation, cold‑start, privacy, and scalability.

All presented academic works have been patented and include intellectual property statements; inquiries should be directed to ads‑[email protected].