Improving Sequential Recommendations with LLMs

The idea

Sequential recommendation — predicting what someone will interact with next from the order of what they have already done — had a decade of specialised models behind it when LLMs arrived. The question this work asks is not whether an LLM can replace them, but where in an existing recommender an LLM actually pays for itself. It sets out three orthogonal approaches, and hybrids of them, so that the answer is a comparison rather than a single system.

Contributions

• Three distinct ways of using an LLM in a sequential recommender, plus hybrids, framed so they can be compared rather than only demonstrated.
• An examination of the technical choices inside each approach — which language model, which baseline recommender, how the sequence is represented — with alternatives enumerated rather than fixed.
• Experiments over three datasets and a large number of configurations, with code and data released for reproducibility.

What it found

Initialising an established sequential model — BERT4Rec or SASRec — with embeddings taken from an LLM produced substantial accuracy gains. Fine-tuning an LLM for the recommendation task taught it not only the task but, to a degree, concepts of the domain; and fine-tuned OpenAI GPT did considerably better than fine-tuned Google PaLM 2. The overall reading is that the value of LLMs here is large and mostly not in replacing the recommender.