Large Language Models Enhanced Collaborative Filtering

Collaborative filtering learns from patterns across user-item interactions. An LLM can add information about product attributes, user preferences, and the compatibility between a history and a candidate item. A prompt about one user and one item covers only a small part of the interaction data. LLM-CF brings similar historical interactions into the prediction context, together with reasoning generated from them.

The dashed boundary in the figure separates offline generation from online serving. Below it, RecGen-LLaMA writes a reusable collection of reasoning examples. Above it, the recommendation model receives a learned feature derived from retrieved examples. Serving still requires feature encoding, retrieval, ICT, and the backbone model; it does not require a fresh LLM-generated response for each request.

Each stored example contains recommendation features, a generated chain of thought, and its historical interaction label. The teacher prompt on the right asks about the user's history, the candidate item's attributes, their compatibility, and possible interest in diversity. The historical label is supplied during this offline generation. The reported collection is sampled from one tenth of the training interactions.

For a current query, the retrieval module uses BGE to embed its textual user and item features and selects similar examples by cosine similarity. Examples from future time steps are excluded. The selected context contains equal numbers of positive and negative historical labels, preventing its label balance from becoming a shortcut. The current query's unknown label is not supplied to ICT.

At the upper right, each historical example becomes a feature embedding, a reasoning embedding, and a label embedding. ICT places these triples before the current query feature in a transformer-decoder sequence. With four examples, the sequence has thirteen positions. The query's final hidden representation becomes the LLM-enhanced collaborative-filtering feature. The reported decoder has two layers and two attention heads.

Training combines the recommendation objective with a cosine reconstruction loss for historical reasoning embeddings. This auxiliary objective encourages the learned feature representations to retain information from the reasoning examples. Ranking models concatenate the resulting feature with their original features; retrieval models add it to the user vector. For candidate retrieval, the current target item's information is masked to avoid exposing the answer.

The horizontal axis measures recommendation AUC; the vertical axis measures MMLU accuracy. Full-parameter tuning on recommendation data alone improves recommendation performance but sharply reduces MMLU accuracy. This trade-off matters because the teacher's next task is to generate informative reasoning examples.

The paper compares data mixing with LoRA, noise-based tuning, representation regularization, and weight interpolation. RecGen mixes recommendation examples with LIMA and alpaca-gpt4 instruction data and retains a useful position on both axes in these experiments. The figure establishes this measured balance; MMLU is the general-capability benchmark used here.

Each panel holds the backbone fixed and plots AUC on the three datasets. Several curves improve when moving from two to four examples, but their peaks occur at different context sizes. Individual settings favor two, six, or eight examples, while others decline with longer contexts. Four examples are the selected practical configuration, not a universal optimum across all curves.

The n4_noc setting removes the label-balance constraint. Its changes differ across datasets and backbones. Keeping the balanced setting makes the main comparison less dependent on the proportion of positive labels in the retrieved context. The ranking, retrieval, and ablation tables below show the main experimental evidence.