Time-Interval-Aware Disentangled Expert Modeling for Next-Basket Recommendation

Abstract: Next-basket recommendation (NBR) is a type of recommendation that aims to predict a set of items a user will purchase based on their historical transaction basket sequences. It is governed by a dynamic interplay between two distinct user intents: habitual repurchase, which involves repeating past behaviors, and exploratory interest, which involves discovering new items. However, existing NBR methods generally suffer from two limitations: (1) they often entangle these conflicting motives within a single representation, causing habits to overshadow discovery, and (2) they rely on discrete sequential modeling that ignores continuous-time intervals and item-specific periodicities. In this paper, we propose a novel solution named Time-Interval Disentangled Experts (TIDE) to address these challenges. TIDE incorporates a Hawkes-enhanced Fourier Time Encoding to capture item-specific temporal periodicities and dynamic decay. To decouple user intentions, TIDE utilizes a dual-expert architecture that integrates a Habit Expert for recurring needs and a Pattern-Guided Exploration Expert for discovery. Combined with an item-aware gating mechanism, TIDE adaptively balances repurchase and exploration. Extensive experiments on four diverse real-world datasets demonstrate that TIDE consistently outperforms representative state-of-the-art NBR methods.

• The paper introduces TIDE, a framework that disentangles habitual repurchase and exploratory interests through time-interval-aware modeling.
• It employs Hawkes-enhanced Fourier encoding and an adaptive gating mechanism to capture non-monotonic repurchase cycles and latent item associations.
• Empirical results on diverse Amazon datasets demonstrate superior performance over baselines, highlighting enhanced precision and robustness.

Time-Interval-Aware Disentangled Expert Modeling for Next-Basket Recommendation

Problem Statement and Motivation

Next-basket recommendation (NBR) represents a complex sequential prediction challenge, wherein the objective is to forecast a set of items a user will purchase in her next transaction, conditional on past basket-level purchase sequences. Existing approaches predominantly entangle two fundamentally distinct, often competing, user intentions within a unified latent representation: habitual repurchase (repetition) and exploratory interests (novelty-seeking). Classical linear-order or deep sequential NBR models typically overlay these factors, leading to habitual bias—where frequent purchases overshadow the detection of exploratory patterns—ultimately harming the system's ability to recommend genuinely novel, user-aligned items. Furthermore, state-of-the-art NBR frameworks frequently assume monotonic temporal decay in user interest (i.e., utility diminishes exponentially with time since last interaction), disregarding empirically observed, item-specific, non-monotonic periodicities (e.g., daily necessities vs. long-tail durable goods). These limitations degrade the capacity to model precise, real-world replenishment cycles, as supported by the detailed empirical analysis across several Amazon product categories.

Technical Contributions

The proposed TIDE (Time-Interval Disentangled Experts) framework directly addresses these gaps by:

1. Temporal Disentanglement via Hawkes-Enhanced Fourier Encoding: TIDE incorporates a Hawkes-influenced process with a Fourier-based time encoding module. For each item in a user's history, the approach captures not only recency and frequency but also personalizes periodicity through learnable spectral components and item-specific dynamic intensity decay. This enables the non-monotonic, spike-like repurchase probability modeling that classic exponential or position-based temporal encoders miss.
2. Dual-Expert Disentangled Architecture: User preference is separated into two explicit subspaces: a Habit Expert models precise replenishment cycles via enriched, temporally-aware item embeddings and basket-level additive attention; a Pattern-Guided Exploration Expert discovers collaborative, novel interests by propagating signals through a global item-pattern bipartite graph. This global structure captures latent item associations beyond the user's personal basket history.
3. Per-Item Adaptive Gating and Contrastive Alignment: To integrate both intent-specific signals into the final recommendation, an item-aware gating mechanism dynamically weighs the relevance of each expert for each candidate item. Integration is further regularized via an InfoNCE-based contrastive loss, ensuring semantic coherence while maintaining functional orthogonality between the experts.
4. Extensive Empirical Validation: TIDE is evaluated on diverse, large-scale Amazon datasets (Beauty, Sports, Grocery, Home), spanning domains with varying repeat/explore dynamics and sparsity levels. Ablations rigorously demonstrate the necessity of each modular component, while supplementary analyses confirm TIDE’s robustness to graph-construction metrics and alternative sequence encodings.

Experimental Results

TIDE consistently achieves top performance across all datasets and standard metrics (Recall@10, NDCG@10, Recall@20, NDCG@20). For instance, in the Beauty domain, TIDE's Recall@10 of 0.1276 outperforms the strongest baseline (SemTHy: 0.1078), yielding an 18.4% relative improvement. In sparsity-dominated Home, and others with complex consumption rhythms, TIDE significantly outpaces entangled or static-decay baselines, illustrating the efficacy of both the Hawkes kernel for habit modeling and the collaborative bipartite pattern graph for exploratory retrieval.

Ablation studies reveal:

• Removing the Exploration Expert (w/o Graph) devastates exploratory recommendation, especially for low-repeat domains (89.4% drop in Home Recall@10).
• Excluding temporal modeling (w/o Time) damages performance where periodicity dominates (e.g., Grocery, Beauty).
• Disabling contrastive alignment increases semantic drift between experts, degrading fusion quality.

Visualization via t-SNE shows clear specialization and partial alignment of the habitual and exploratory subspaces after training, in contrast to the entangled chaos at model initialization.

Alternate pattern mining (NPMI, Jaccard, Lift) in the Exploration Expert shows TIDE’s robustness, with domain-specific optimality but no collapse, supporting the generality of the architecture.

Time interval-aware encoding (Δt via Hawkes/Fourier) outperforms both static position encodings and recurrent models (GRU), particularly as repurchase cycles grow less regular and basket sparsity increases.

Theoretical and Practical Implications

On the theoretical side, TIDE establishes that fine-grained temporal modeling (explicit periodicity and excitation via Hawkes processes) is essential for accurate next-basket recommendation; simple position or monotonic decay fails in the presence of real-world item life cycles. Furthermore, explicit representation disentanglement—habitual vs. exploratory—enables balanced prediction unaffected by popularity skew or habitual dominance, especially in sparse domains.

Practically, TIDE’s architecture is flexible: the integration of a stationary item-pattern bipartite graph allows for extension toward dynamic/topological graph models (e.g., transformer-based GNNs), and the InfoNCE-based contrastive objective is amenable to future information-theoretic generalization. The per-item adaptive gating enables fine-grained recommendation adaptation, crucial as user intent drifts between exploration and exploitation in varied retail or service environments.

Limitations and Future Directions

TIDE currently employs a static global item-pattern bipartite graph, potentially limiting responsiveness in rapidly evolving commercial environments. Dynamic graph neural architectures or meta-learned structural adaptation could further improve system reactivity and scalability. While contrastive alignment suffices for semantic regularization between intent subspaces, more principled information-theoretic disentanglement could offer stronger guarantees against latent space collapse or drift. Integration with advanced generative or retrieval architectures may enable real-time pattern adaptation and further improvements for cold-start, evolving, or adversarial settings.

Conclusion

TIDE presents a rigorously validated, modular NBR framework with explicit, temporally-aware intent disentanglement. Its strong empirical performance, along with ablation analyses, underline the value of domain-aligned dual-expert separation and the necessity of accurate temporal dynamics for basket-level recommendation. Theoretical advances in representational geometry and practical engineering in item-gating and collaborative graph mining position TIDE as a robust foundation for further research in next-basket and general sequence-based recommendation systems.

Reference: "Time-Interval-Aware Disentangled Expert Modeling for Next-Basket Recommendation" (2605.00499)