DiffNBR: A Spatiotemporal Diffusion and Information‑Bottleneck Approach for Next‑Basket Recommendation

INTRODUCTION

Next‑basket recommendation aims to predict the unordered set of items a user will buy in a future basket. Existing methods often ignore “threshold‑meeting” behavior, where users add seemingly unrelated low‑price items to satisfy free‑shipping or discount thresholds, treating them as noise. Moreover, current models lack a theoretically analyzable mechanism to decouple repetitive purchases from exploratory ones. DiffNBR is the first next‑basket recommendation (NBR) framework built on diffusion models to address these challenges.

Core Idea of DiffNBR

DiffNBR captures two complementary dimensions of user purchase behavior: a spatial dimension that identifies latent combination strategies caused by threshold‑meeting, and a temporal dimension that models the consistency and evolution of these behaviors. The framework employs two parallel denoising diffusion probabilistic models (DDPMs) to learn joint spatiotemporal representations. Inspired by information‑bottleneck (IB) theory, a directed decoupling mechanism adjusts information flow to compress redundant information in the generative representation, forcing the model to focus on exploratory patterns.

Model Framework

The model consists of three key stages:

Spatiotemporal Encoding (Feature Reconstruction) – Normalized embeddings remove user‑side and item‑side bias. Spatial aggregation compresses co‑occurring item embeddings to extract temporal features, while temporal aggregation applies exponential decay to compress each item’s historical basket sequence, extracting spatial features. A target‑basket‑generated guidance signal serves as a prior for the diffusion process.

Spatiotemporal Diffusion (Generative Modeling) – Two independent sub‑modules, TDiff (handling temporal evolution) and SDiff (handling spatial co‑occurrence), reconstruct the target basket representation. Unlike traditional noise‑only corruption, the model injects a Gaussian‑noised target representation into the encoded historical sequence or item set, using guidance embeddings to balance randomness and accelerate learning. During reverse diffusion, stacked Transformer networks act as approximators, iteratively denoising under guidance to recover a high‑quality target basket vector.

Information‑Bottleneck (Decoupling & Fusion) – The IB principle minimizes mutual information between habitual and generated representations, filtering redundant repetitive information and achieving directed decoupling of repeat vs. exploratory purchases. An adaptive gating mechanism dynamically balances the two feature streams before mapping to candidate item scores.

Experimental Results

DiffNBR outperforms eight representative baselines on four real‑world datasets. On the Dunnhumby dataset, Recall and NDCG improve by 8.45%–14.6% over the strongest baseline.

**Ablation studies** confirm the necessity of temporal diffusion (T), spatial diffusion (S), and the IB constraint (H); removing any component degrades performance, and the H module exhibits strong synergy when paired with the generative representation.

**Repeat vs. Exploration** – While repeat‑purchase prediction slightly declines, exploratory recommendation (‑exp) shows a dramatic boost, e.g., a 245.32% increase on Instacart.

**Threshold‑Meeting Modeling** – A case study shows DiffNBR accurately identifies strategic “paper‑towel” and “dental‑floss” purchases driven by threshold‑meeting, assigning them high scores compared to the M2 model that relies solely on item association.

**Representation Quality** – Visualization demonstrates that DiffNBR’s embeddings are more uniformly distributed, aligning core points of the true distribution and isolated outliers representing strategic purchases.

Conclusion

DiffNBR successfully brings diffusion models into the NBR domain through a novel spatiotemporal dual‑path design and an IB‑based decoupling mechanism, overcoming the difficulty of capturing threshold‑meeting behavior and balancing repeat vs. exploratory patterns. Experiments validate its high‑precision prediction capability, and the theoretical framework provides solid support for understanding strategic purchase decisions, laying groundwork for future generative recommendation research.