Modeling Continuous Time Intervals in Sequential Recommendation

55 concepts53 papers1 proposal

Consider two users with the exact same purchase history: they buy a laptop, and then they buy a mouse. The first user buys the mouse five minutes after purchasing the laptop. The second user buys the mouse five months later.

To a traditional sequential recommendation model, these two histories are identical. They are both represented as the discrete sequence [Laptop, Mouse]. Yet the underlying user intent, urgency, and preference stability in these two scenarios are completely different.

This "time blindness"—formally known as temporal interval neglect—is a fundamental limitation of discrete-time sequence modeling. By treating user history as a simple ordered list, standard models assume that the intervals between actions are uniform. In reality, the continuous wall-clock time elapsed between interactions carries a vital signal about the stability, decay, and evolution of user preferences.

The Cost of Time Blindness

Why does ignoring the specific time interval matter? As Enneng Yang et al. point out in Uniform Sequence Better: Time Interval Aware Data Augmentation for Sequential Recommendation, temporal spacing is a direct indicator of preference stability. When a model ignores these intervals, it struggles to differentiate between short-term bursts of activity (like a focused shopping session) and long-term behavioral shifts over months (Chao Chen et al.).

Furthermore, human behavior is governed by distinct item-dependent rhythms. For instance, Cho-Jui Hsieh et al. demonstrate in Scalable Demand-Aware Recommendation that different categories of goods require different "refractory periods." Recommending a durable item like a television immediately after a purchase fails to account for the product's natural replacement cycle.

Similarly, in next-basket recommendation, Aleksey Romanov et al. note in Time-Aware Item Weighting for the Next Basket Recommendations that capturing item-specific repurchase cycles (such as buying milk every three days versus laundry detergent every month) requires explicit interval modeling. Without it, the predictive quality of a model rapidly degrades when there are significant gaps since a user's last interaction.

Injecting Time into Recurrent and Attention Architectures

To overcome this limitation, researchers have modified standard discrete architectures to accept continuous time inputs, primarily focusing on recurrent gates and attention mechanisms.

Time-Aware Recurrent Gates

In standard recurrent networks (like LSTMs or GRUs), hidden states update based purely on the step index ($t, t+1$). To incorporate irregular intervals, Yu Zhu et al. introduced the Time-LSTM architecture in What to Do Next: Modeling User Behaviors by Time-LSTM. Time-LSTM introduces specialized time gates that damp the cell memory based on the elapsed duration between events. If a long period of inactivity has passed, the network automatically discounts old, short-term interactions while preserving long-term preferences.

Continuous-Time Attention

In attention-based models like Transformers, standard positional encodings only capture ordinal rank ($1\text{st}, 2\text{nd}, 3\text{rd}$). To introduce temporal awareness, Chao Chen et al. proposed a self-modulating attention mechanism in Learning Self-Modulating Attention in Continuous Time Space with Applications to Sequential Recommendation. Instead of static weights, their model uses continuous-time intensity functions to dynamically amplify or attenuate the attention paid to past behaviors based on when they occurred.

A similar intuition applies to tracking human learning over time. In Rethinking and Improving Student Learning and Forgetting Processes for Attention based Knowledge Tracing Models, Youheng Bai replaced rigid decay functions with flexible Relative Positional Encodings (RPEs) to capture how a student's memory fades over irregular real-world intervals, demonstrating that wall-clock time is a much stronger predictor of cognitive forgetting than discrete interaction steps.

Modeling Preferences as Continuous Trajectories

While modifying discrete models helps, a more radical approach treats user preferences as a continuous trajectory moving through a latent space. Instead of jumping discretely from step to step, the user's state evolves continuously according to differential equations.

In ReCODE: Modeling Repeat Consumption with Neural ODE, Jun Xu leverages Neural Ordinary Differential Equations (ODEs) to learn return intervals directly from data without relying on rigid parametric assumptions. By modeling the user's latent state as a continuous-time function, the system can precisely predict when a user will repeat a consumption behavior.

Taking this a step further, Krishna Prasad Neupane et al. introduced evidential stochastic differential equations in Evidential Stochastic Differential Equations for Time-Aware Sequential Recommendation. Their framework mathematically links the continuous time elapsed between interactions to the uncertainty of a user's preferences:

Discrete Sequence:    [Item A] -------------------------> [Item B]
                       (Assumed uniform step; uncertainty is hidden)

Continuous Timeline:  [Item A] ====(Time Elapses)====> Preference Uncertainty Widens ====> [Item B]

As the time gap grows, the drift and diffusion terms in the SDE naturally expand the model's uncertainty, preventing it from making overconfident, outdated recommendations based on old data.

Why Time-Awareness Works: The Intuition

By explicitly modeling continuous time intervals, these frameworks unlock several critical capabilities that discrete sequence models miss:

  • Disentangling Multi-Interest Threads: Short time gaps typically imply a single, continuous search session with highly focused intent. Long gaps, however, signal a potential reset or evolution of user focus. As Qi Shen 0001 et al. demonstrate in Temporal aware Multi-Interest Graph Neural Network for Session-based Recommendation, injecting item-to-item time gaps allows graph networks to disentangle a single user history into multiple distinct, time-sensitive interest threads.
  • Capturing Non-Linear User Fatigue: Recommending the same category of content repeatedly causes psychological fatigue, but this effect is highly dependent on temporal proximity. In Modeling User Fatigue for Sequential Recommendation, Nian Li et al. show that rapid successions of similar items cause severe fatigue, whereas the same items distributed over longer intervals maintain their appeal. Continuous-time modeling allows systems to accurately weight historical interactions based on this temporal proximity.

By moving from discrete step sequences to continuous time intervals, modern recommendation systems transition from merely understanding what a user likes to predicting when their specific needs will resurface.

Key Foundational Work

A Maximum Entropy Approach to Collaborative Filtering in Dynamic, Sparse, High-Dimensional Domains (2002) — contributed 1 concept elements to this direction.
Does Time Matter? Modeling the Effect of Time with Bayesian Knowledge Tracing (2011) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_temporal_interval_neglect_e5abb33d,
  title = {temporal interval neglect},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/temporal-interval-neglect/rdr_e5abb33d1c1d413cbf91eb08349af680},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_temporal_interval_neglect_e5abb33d_v1,
  title = {temporal interval neglect},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/temporal-interval-neglect/rdr_e5abb33d1c1d413cbf91eb08349af680/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake