RecPFN Introduces In-Context Learning for Sequential Recommendations

This research introduces RecPFN, a novel neural network architecture designed for sequential recommendation tasks. Unlike traditional models, RecPFN leverages in-context learning, allowing it to perform Bayesian-style inference efficiently from a small set of examples without requiring weight updates during inference. The model is pre-trained exclusively on synthetic clickstream data generated from a broad causal prior. During deployment, RecPFN utilizes a lightweight decoder-only transformer. This component conditions on a few domain-specific sequences to generate next-item predictions in a single forward pass. This approach makes it highly deployment-efficient and resilient to changes across different domains. Empirical evaluations across eight public benchmarks demonstrate RecPFN's superior zero-shot performance. It also remains highly competitive with supervised methods, particularly in scenarios with limited computational resources or scarce data, offering a practical pathway towards more generalizable and data-efficient recommendation systems.