🤖 AI

Decoupled Temporal Encoding for Generative Recommendation

This paper proposes Decoupled Temporal Encoding (DTE), a lightweight framework for generative recommendation that improves upon existing methods by separating temporal dynamics from item order through two complementary modules: a personalized macro-temporal module for broad temporal patterns and a time-gated micro-sequential module for dense local interactions.

Original authors: Pengfei Jia, Jingjian Wang, Jingmao Li, Ge Zhang, Feng Shi

Published 2026-08-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Pengfei Jia, Jingjian Wang, Jingmao Li, Ge Zhang, Feng Shi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the digital world, recommendation systems act as the quiet curators of our daily lives, suggesting what to watch, buy, or eat next. For years, these systems have relied on a powerful type of artificial intelligence known as the Transformer, which excels at understanding sequences. Think of a user's history as a story: the system reads the chapters of items a person has interacted with to guess the next chapter. To make sense of a story, the AI needs to know not just what happened, but when and in what order. Standard methods for teaching AI this sense of order were borrowed from language models, treating every item in a list as a simple step in a sequence, like words in a sentence. However, human behavior in the real world is far more complex than a simple list. It is shaped by the clock, the calendar, and the rhythm of daily life. A person's hunger at lunchtime, their shopping habits on a weekend, or their reaction to a sudden sale are all temporal patterns that standard methods often struggle to capture alongside the simple order of events.

Researchers at Alibaba's Rajax Network Technology, working within the high-pressure environment of the Taobao app's food delivery and instant retail services, identified a gap in how these systems understand time. They observed that user behavior is driven by two distinct forces that often get tangled together. One force is the broad, sweeping context of time: the fact that it is Tuesday, that it is near lunch, or that a specific shop is experiencing a sudden surge in traffic. The other force is the tight, immediate sequence of actions: the fact that a user clicked on a noodle shop, then a drink, then a dessert within a few minutes. Existing systems tried to feed both of these signals into the AI through a single, unified channel, much like trying to describe a storm by only measuring the wind speed and ignoring the rain. This approach made it difficult for the model to distinguish between the general mood of the day and the specific, rapid-fire choices a user was making in the moment.

To solve this, the team developed a new framework called Decoupled Temporal Encoding. Instead of forcing the AI to learn time and order from the same source, they built two separate, lightweight modules that work in harmony. The first module, which they call the macro-temporal encoder, focuses on the big picture. It injects compact signals into the system to tell the AI about the time of day, the day of the week, and how much time has passed since the last interaction. This allows the model to understand that a user ordering dinner at 7 PM on a Friday is likely in a different mindset than the same user ordering lunch at 12 PM on a Monday. The second module, the micro-sequential encoder, handles the fine details. It acts as a gatekeeper, only stepping in to emphasize the strict order of events when interactions happen so quickly that the timestamps alone become blurry. If a user clicks through several items in rapid succession, this module ensures the AI pays attention to the precise sequence of those clicks, rather than just the fact that they happened close together.

The researchers tested this approach on a massive dataset containing billions of interactions from the Taobao app, as well as on a public benchmark for video recommendations. The results showed that separating these two types of time information significantly improved the system's ability to predict what a user would want next. In offline tests, the new method outperformed all previous state-of-the-art models, achieving higher accuracy in predicting clicks and reducing errors. Crucially, the team found that the system worked best when users had clear, repeating patterns in their behavior, such as regular lunch orders, or when their interactions were very dense. The analysis of the AI's internal attention patterns revealed that the system learned to reinforce these periodic habits without losing track of the immediate sequence of events.

The true test of this research came when the team deployed the system into the live traffic of the Taobao app. In a rigorous three-week experiment involving millions of real users, the new method delivered measurable business improvements. It increased the click-through rate by 1.8 percent and the revenue per thousand impressions by 3.0 percent, all while adding only a negligible amount of delay to the system's response time. Because the gains were consistent and the system remained stable, the company rolled out the update to 100 percent of the traffic for the Shangou advertisement recommendation system in January 2026. This deployment demonstrates that by untangling the broad context of time from the specific order of actions, recommendation systems can become more attuned to the complex, rhythmic nature of human life, offering suggestions that feel not just timely, but truly personal.

Technical Summary: Decoupled Temporal Encoding for Generative Recommendation

1. Problem Definition and Motivation

Generative recommendation models, which treat user interaction histories as autoregressive item sequences, rely heavily on Transformer architectures. However, standard self-attention mechanisms are permutation-invariant and lack inherent sequence order, necessitating positional encoding. While existing methods (e.g., sinusoidal embeddings, RoPE) effectively model discrete token order, they often fail to capture the rich, continuous-time semantics inherent in recommendation data.

In real-world scenarios like food delivery and instant retail, user behavior is shaped by multi-level temporal effects distinct from simple sequential order:

Recency: Recent interactions often carry stronger predictive value.
Periodicity: Demand concentrates around specific times (e.g., meal times) and shifts between weekdays and weekends.
Burstiness: Promotions or holidays cause short-term traffic surges.

Current time-aware approaches often inject these heterogeneous temporal signals into a unified representation or a single modeling pathway. This conflation makes it difficult to distinguish broad temporal dynamics from local sequential order cues. Furthermore, complex temporal modeling (e.g., large interval embedding tables) often incurs high memory costs and latency, hindering deployment in strict production environments.

2. Methodology: Decoupled Temporal Encoding (DTE)

To address these limitations, the authors propose Decoupled Temporal Encoding (DTE), a lightweight framework that explicitly separates "macro" temporal dynamics from "micro" sequential order effects using two complementary modules.

3.1 Macro-Temporal Encoding

This module operates at the input layer to inject compact temporal context into item embeddings. Instead of large lookup tables, it uses a small set of temporal primitives combined via an adaptive mixture:

Recency Decay: An exponential decay function based on the time gap between the interaction and the current prediction time.
Intra-day Periodicity: A sinusoidal function modeling daily cycles (e.g., lunch/dinner peaks).
Weekly Periodicity: A sinusoidal function modeling weekly shifts (weekday vs. weekend).
Burst-Sensitive Variation: A component capturing short-term deviations in target-side traffic (e.g., shop-level traffic surges) relative to historical baselines.

These components are weighted by a gating network that adapts the importance of each primitive based on user-specific features (derived from average-pooled historical embeddings). The final enhanced embedding is computed as:
$\mathbf{e}'_i = \mathbf{e}_i + \mathbf{r} \cdot m(t_i)$
where $m(t_i)$ is the adaptive mixture of the four temporal signals, and $\mathbf{r}$ is a learnable projection vector.

3.2 Micro-Sequential Encoding

This module operates at the attention level to provide relative-order bias, specifically when timestamps are too dense to distinguish interactions reliably.

Time-Gated Mechanism: A soft gate determines whether to apply a relative-order bias based on the absolute time gap ( $\Delta t_{ij}$ ) between two interactions.
Conditional Activation: The bias is only activated when interactions are temporally dense (i.e., when the time gap is below a learnable threshold $\tau$ ). This ensures that sequential order acts as a complementary signal rather than a replacement for timestamp information.
Attention Modification: The standard attention score is augmented with this gated bias:
$A'_{ij} = A_{ij} + B_{ij}$
where $B_{ij}$ is the time-gated relative-order bias.

3. Key Contributions

Identification of a Limitation: The paper highlights that existing positional encodings often conflate heterogeneous temporal effects (recency, periodicity) with local sequential order, limiting their effectiveness in modeling evolving user intent.
Architectural Decoupling: DTE introduces a novel framework that decouples macro-temporal context (input-level) from micro-sequential order (attention-level), allowing the model to handle broad temporal dynamics and dense local interactions separately.
Efficiency and Deployability: The design is parameter-efficient, requiring minimal additional parameters and memory, and is compatible with strict latency constraints, enabling "plug-and-play" integration into existing Transformer-based ranking systems.
Comprehensive Validation: The method is validated through offline experiments on public and industrial datasets, as well as extensive online A/B testing and full-traffic deployment.

4. Experimental Results

Offline Evaluation

Experiments were conducted on a public benchmark (KuaiRand 1K) and a large-scale industrial dataset (Shangou from Taobao).

Performance: DTE achieved state-of-the-art results, outperforming order-only encodings (Sinusoidal, RoPE, ALiBi) and time-aware baselines (BST, TiSASRec, RAB).
- On the Industrial dataset, DTE achieved a GAUC of 0.7098 (vs. 0.6982 for the best baseline BST+ALiBi).
- On KuaiRand 1K, DTE achieved a GAUC of 0.9243 (vs. 0.9110 for the best baseline).
Ablation Studies: Removing either the Macro or Micro module resulted in significant performance degradation, confirming the complementary nature of the two components. The "Soft Gate" mechanism was also shown to be critical, as uniform application of order bias was less effective.
Efficiency: DTE introduced only 238 additional parameters (0.47 KB FP16 memory) and an offline inference latency of 3.02 ms, comparable to lightweight baselines like RAB and significantly lower than embedding-heavy methods like TiSASRec.

Online Evaluation

DTE was deployed in the ranking stage of the Shangou advertisement recommendation system in the Taobao App.

Business Metrics: Online A/B tests (20% treatment vs. 20% control) over three weeks showed a +1.8% increase in Click-Through Rate (CTR) and a +3.0% increase in Revenue Per Mille (RPM).
Latency: The serving latency increased by only +0.3% (from 15.4 ms to 15.5 ms).
Deployment: Following successful A/B tests, DTE was rolled out to 100% of the Shangou advertisement traffic in January 2026, with no observed systematic regression in business or serving metrics.

5. Significance and Claims

The paper claims that DTE offers a practical solution to the challenge of modeling temporal effects in generative recommendation. By decoupling broad temporal context from local sequential order, the method improves ranking quality without the heavy computational or memory overhead associated with traditional time-aware models.

The authors emphasize that the success of DTE in a production environment demonstrates the value of separating macro-temporal dynamics (like recency and periodicity) from micro-sequential cues. The framework is presented as broadly applicable beyond the specific food delivery scenario, as evidenced by its success on the public KuaiRand benchmark. The paper concludes that while DTE yields smaller gains for users with sparse histories or weak periodicity, its efficiency and performance make it a viable, scalable solution for industrial generative recommendation systems.