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

Next-basket recommendation (NBR) predicts the set of items a user will buy next based on past transaction sequences. A core challenge is balancing two often-conflicting user intents: habitual repurchase (repeating past buys) and exploratory interest (discovering new items). Existing models typically entangle these motives in a single representation, causing habits to dominate and ignoring the continuous time intervals between purchases, which carry critical signals about user behavior.

To solve this, researchers introduce TIDE (Time-Interval Disentangled Experts). The model incorporates a Hawkes-enhanced Fourier Time Encoding that learns item-specific periodicities and dynamic decay over time. It then uses a dual-expert architecture: a Habit Expert that captures recurring needs and a Pattern-Guided Exploration Expert that models discovery behavior. An item-aware gating mechanism adaptively balances these two experts for each prediction.

Evaluated on four diverse real-world datasets, TIDE consistently outperforms representative state-of-the-art NBR methods. The approach demonstrates that explicitly separating habitual and exploratory intents—while factoring in precise time intervals—significantly improves recommendation accuracy. This work has practical implications for e-commerce, subscription services, and any platform where users alternate between routine purchases and exploration.