Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes

Auto Encoding Neural Process for Multi-interest Recommendation

AAAI 2025 · Probabilistic and scientific modeling › Uncertainty and Gaussian processes · approach: Priors and functional variational uncertainty · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes NP-Rec, which combines a Mamba sequence model with a neural process (MLP encoder-decoder with latent variable and categorical classifier) to model multi-interest recommendation as learning a distribution over functions, providing predictions with uncertainty.

Problem

Multi-interest sequential recommendation lacks adaptability to diverse user preference patterns and calibrated uncertainty over users' shifting current intents, which NN-based single- and multi-point interest representations do not provide.

Approach

Treat next-item recommendation as a neural process: learn a distribution over preference functions via a latent-variable encoder-decoder over context/target splits of the history, sample K latent codes, and average the resulting categorical distributions, with entropy of the mean giving uncertainty.

Key results

• NP-Rec achieves the best HR and NDCG on all four datasets, with average improvements up to 13.94% over the strongest baselines per the abstract (e.g., H@5 on ML-100K improves 39.52% over the strongest baseline).
• Ablation: removing the neural process module (w.o. NP) lowers HR@10 on ML-100K from 22.75 to 15.49, and removing the Mamba sequence model (w.o. SM) drops it to 9.66.
• Plugging the NP model onto SIR backbones (TriMLP, GRU4Rec, NextItNet, SASRec) largely boosts their recommendation performance.

Measured effects reported in the paper

Surprising findings

• Removing the sequence model causes a much larger drop than removing the neural process module, indicating sequential modeling is essential for meaningful interest representations in the NP framework.

Limitations

• Model sensitivity to the number of latent samples K: best K differs per dataset (1, 5, 15), requiring dataset-specific tuning.
• Pooling for the order-invariant aggregator is simple averaging, and autoregressive inference is not explored (stated as future work).

Experimental setup

Datasets: MovieLens ML-100K MovieLens ML-1M Foursquare NYC Foursquare TKY

Benchmarks: ML-100K ML-1M NYC TKY

Metrics: Hit Ratio (HR@5) HR@10 NDCG@5 NDCG@10

Baselines: GRU4Rec NextItNet SASRec TriMLP MIND ComiRec PIMI Re4 REMI

Builds on: Neural Processes Mamba ANP TriMLP

Closest papers by meaning

• NP-MiSR: Neural Process-based Multi-Interest Learning for Session-Based Recommendation (AAAI 2026 · similarity 0.91)
• Dynamic Multi-Interest Graph Neural Network for Session-Based Recommendation (AAAI 2025 · similarity 0.87)
• Efficient Heterogeneous Collaborative Filtering without Negative Sampling for Recommendation (AAAI 2020 · similarity 0.86)
• Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations (ICML 2024 · similarity 0.86)
• Preference Diffusion for Recommendation (ICLR 2025 · similarity 0.85)
• Density-based User Representation using Gaussian Process Regression for Multi-interest Personalized Retrieval (NeurIPS 2024 · similarity 0.84)