Enterprise Knowledge Recommendation System at Alibaba: Architecture, Challenges, and Large Model Applications

The presentation begins with an overview of enterprise digital transformation, describing three stages—informationization, digitization, and intelligence—and emphasizing the crucial support provided by technical communities such as Alibaba's internal ATA platform.

It then outlines the value and challenges of enterprise knowledge recommendation, highlighting short content lifecycles, long document length, high knowledge barriers, and cold‑start problems caused by sparse user behavior.

The core architecture of Alibaba's knowledge recommendation system is described in four layers:

Feature layer: content, attribute, context, and collaborative‑filtering features.

Engine layer: offline data‑works processing, OSS storage, machine‑learning training, and online services (RTP, TPP, graph storage).

Ranking layer: multi‑stage recall, coarse ranking, fine ranking, and re‑ranking, with six recall methods (collaborative, tag‑based, semantic, rule‑based, etc.).

Function layer: homepage recommendation, technical‑person recommendation, search‑term recommendation, and related‑article recommendation.

Recall methods include collaborative (Swing and relationship graphs), tag‑based (category and skill tags), semantic (BERT embeddings), and popularity‑based approaches, with tag‑based recall showing the best click‑through performance.

The fine‑ranking stage evolves from linear FTRL models to increasingly sophisticated deep models such as Wide&Deep, DeepFM, DCN, AFM, NFM, DIN, AutoInt, FiBiNet, and reinforcement‑learning‑based DRN, ultimately adopting a Wide&DCN model that improves CTR by about 10% over FTRL.

To address long‑document and user‑interest challenges, a large‑model recommendation framework is introduced, comprising an Article Encoder (feature embedding, BERT title encoder, LLM content encoder using ChatGLM‑6B), an Instant User Interest (IUI) module with cross‑attention over recent browsing sequences, and a Constant User Interest (CUI) module that incorporates user profiles via LLM encoders.

Extensive offline and online evaluations show that the proposed large‑model approach outperforms six SOTA baselines on AUC, MRR, and nDCG metrics, and ablation studies confirm the importance of both constant and instant user interest components.

The article concludes by summarizing the system’s architecture, multi‑recall implementation, progressive ranking models, and the ongoing large‑model research aimed at better representing long technical documents and dynamic user interests.