Applying Large Language Models to Recommendation Systems at Ant Group

Introduction: This article shares Ant Group's research and deployment of large models in recommendation scenarios.

Background: Traditional recommendation pipelines suffer from exposure and popularity bias; integrating large language models can inject world knowledge to mitigate bias.

Approach 1 – Knowledge Extraction: A two‑stage pipeline where the LLM generates structured or textual knowledge graphs from online corpora, using relation type selection, entity generation, and prompt‑driven filtering.

Approach 2 – LLM as Teacher: Distilling GPT‑3.5/ChatGPT into smaller models (LLAMA2‑7B) to provide reasoning‑rich recommendation rationales, followed by generative loss fine‑tuning and embedding of the rationales.

Approach 3 – Efficient Deployment: Producing seed‑user knowledge embeddings and serving them online, reducing per‑user inference cost while preserving performance.

Experiments: Evaluated on backbone models GRU4Rec, SASRec, SRGNN; the DLLM2Rec model improves long‑tail recommendation, reduces popularity bias, and outperforms baseline methods.

Challenges & Future Work: Issues with knowledge reliability, cross‑modal distillation, and ranking‑based distillation are addressed with ranking loss and embedding alignment.

Q&A: Discussed model size (LLAMA2‑7B), offline vs online inference, and handling data sparsity in low‑frequency scenarios.

Conclusion: Large models can be integrated into production recommendation systems through knowledge extraction, teacher‑student distillation, and efficient serving, delivering better accuracy and robustness.