Time-Aware Sequential Recommendation

This direction is split into other directions: temporal interval neglect. Its last version stays readable and citable here.

Version 1 of 1179 papers209 conceptsfrozen 2026-10-04

Most sequential recommendation systems treat a user's history as a simple ordered list of items: $i_1, i_2, \dots, i_n$. In this discrete-step view, the transition from a movie click to a purchase is treated the same whether it happened five minutes later or five months later.

However, researchers have increasingly identified "time-blindness" as a fundamental flaw in these models. The actual time elapsed between interactions—the temporal interval—is often a more powerful signal of user intent than the order itself. Recent work focuses on transforming recommendation from a sequence of discrete steps into a continuous-time modeling problem.

The Signal in the Gap

Standard models like Factorized Personalized Markov Chains (FPMC) or basic RNNs assume that the relationship between items is static. But as Enneng Yang and colleagues argue in Uniform Sequence Better, the stability of a user's preference is directly reflected in the variance of their interaction intervals.

When a user interacts with items at consistent, "uniform" intervals, their intent is predictable and stable. Conversely, irregular gaps often signal preference drift—a shift in interest that traditional models fail to capture because they only see the next item in the chain, not the silence between them. Yang addresses this with TiCoSeRec, which uses time-aware data augmentation to help models handle these irregular temporal gaps.

From Discrete Steps to Continuous Trajectories

If user interest is a living, breathing thing, treating it as a sequence of "jumps" is insufficient. Several researchers propose modeling interest as a continuous trajectory:

  • Ordinary Differential Equations (ODEs): Shu-Ting Shi and Jianke Zhu introduced the Deep Time-Stream (DTS) framework, which uses ODEs to track how interest evolves over continuous time. This allows the model to "fill in the blanks" during periods of inactivity Deep Time-Stream Framework.
  • Self-Modulating Attention: Chao Chen developed a mechanism that uses continuous-time intensity functions to dynamically decay the importance of past actions based on exactly how much time has passed, rather than how many items have been clicked since Learning Self-Modulating Attention.
  • Hawkes Processes: To capture both short-term bursts and long-term cycles, Andreas Peintner and Eva Zangerle proposed HyperHawkes. This model uses Hawkes Processes—a mathematical tool for modeling events that "trigger" future events—to predict when a specific intent (like a yearly gift-shopping trip) is likely to resurface Hypergraph-based Temporal Modelling.

Demand and the Refractory Period

In e-commerce, time-awareness is critical for distinguishing between durable and nondurable goods. Cho-Jui Hsieh and Jinfeng Yi point out that recommending a television immediately after a user buys one is a failure of temporal logic. Their DAROSS model learns "inter-purchase durations" to respect the necessary refractory period for durable goods while identifying the rapid replenishment cycles of nondurable goods like snacks Scalable Demand-Aware Recommendation.

Similarly, in "Next Basket" recommendations, Enneng Yang uses hypergraph convolutions to model how different product categories have unique temporal rhythms, ensuring that milk is recommended weekly while lightbulbs are recommended every few months Basket Representation Learning.

The Global Timeline Paradox

A surprising finding by Aixin Sun and Chenliang Li challenges the "more data is better" mantra. They discovered that "loyal" users with long histories often receive worse recommendations because models over-rely on the order of their interactions while ignoring their position on the global timeline.

If a user was highly active three years ago but just returned today, their old interactions are often treated with the same weight as a new user's interactions from yesterday. Sun argues that recency on the global timeline is a more significant predictor of accuracy than the total volume of historical data Do Loyal Users Enjoy Better Recommendations?.

Temporal Awareness in Modern Architectures

As the field moves toward Large Language Models (LLMs) and Transformers, the challenge of "temporal oversight" remains.

  • CETRec, developed by Yutian Liu and Xiang Wang, uses counterfactual tuning to force LLMs to recognize that the chronological order of events is a causal factor, not just a formatting choice Counterfactual Tuning.
  • MOJITO uses Gaussian mixtures of attention to integrate external temporal context—like the hour of the day or the season—directly into the Transformer's self-attention mechanism, allowing the model to learn that a user's music preference at 8:00 AM (commute) is fundamentally different from their preference at 11:00 PM (sleep) Attention Mixtures.

Papers in this version

Showing 60 of 179 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_time_aware_sequential_recommendation_5b30e264,
  title = {Time-Aware Sequential Recommendation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/time-aware-sequential-recommendation/rdr_5b30e264ff45488da434b3a949aef764},
  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_time_aware_sequential_recommendation_5b30e264_v1,
  title = {Time-Aware Sequential Recommendation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/time-aware-sequential-recommendation/rdr_5b30e264ff45488da434b3a949aef764/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}