T2ST: Data-Efficient Spatio-Temporal Foundation Models via Temporal Data and Knowledge Transfer

Yue Xu, Kaiyun Wei, Wenying Duan, Zichen Xu, Shuai Ma, Xiaoxi HearXiv 2026 · arXiv

TL;DR

T2ST (Temporal-to-Spatio-Temporal) is a novel framework that drastically reduces the data requirements for training Spatio-Temporal Foundation Models (STFMs). By augmenting a tiny fraction of expensive spatio-temporal data with abundant, plain time-series data and distilling knowledge from pre-trained time-series models, T2ST allows models trained on just 10% of a dataset to match or even outperform those trained on the full 100%.

What Problem Does This Paper Address?

Spatio-temporal foundation models (STFMs) are designed to predict complex patterns across urban environments, such as traffic flow, energy consumption, and air quality. To work well, they require massive "corpora"—huge datasets of synchronized sensor readings paired with precise spatial metadata (like road network graphs or coordinates).

The bottleneck is that this data is incredibly expensive to collect and curate. In contrast, "plain" time-series data (sequences of values without spatial context) is abundant. While a typical STFM might be trained on a few hundred million data points, Time-Series Foundation Models (TSFMs) like Google’s TimesFM are trained on hundreds of billions (see Figure 4).

The authors ask a compelling question: Can we use the "easy-to-get" temporal knowledge from TSFMs to supplement the "hard-to-get" spatio-temporal data needed for STFMs?

Comparison of training data scales between STFMs and TSFMs

Methods

T2ST bridges the data gap through a two-stage framework: Temporal Data Selection and Knowledge Distillation.

1. Temporal Data Selection

Instead of just grabbing random time-series data, T2ST uses a shared encoder to map both the limited spatio-temporal data and a massive auxiliary time-series pool into a common representation space. It then selects two specific types of data:

  • Matched Sequences: Data that looks similar to the existing spatio-temporal patterns, reinforcing what the model has already seen.
  • Diverse Sequences: Novel patterns that aren't in the original dataset but are "trusted" by a pre-trained teacher model (TSFM). This expands the model's "vocabulary" of temporal patterns.

2. Knowledge Distillation

The framework uses a frozen TSFM (the teacher) to guide the STFM (the student). Because the teacher and student often have different architectures, T2ST doesn't try to match their internal features directly. Instead, it uses:

  • Temporal Relation Distillation: It aligns the relationships between different time patches. If the teacher thinks patch A is highly related to patch B, the student is encouraged to learn that same relationship.
  • Multi-scale Prediction Distillation: It aligns forecasts at multiple resolutions (e.g., 5-minute intervals vs. 40-minute averages). This ensures the model captures both high-frequency noise and long-term trends.
Fig. 1. Overview of the T2ST framework. (a) Temporal data selection: a shared encoder maps spatio-temporal and auxiliary time-series windows into a common representation space. A calibrated spatio-temporal reference guides matched selection, while teacher reliability filtering and novel pattern discovery guide diversity selection. Selected embeddings identify the source windows retrieved by index to construct DM and DD. (b) Knowledge distillation: a frozen TSFM guides a STFM student through temporal relation distillation and multi-scale prediction distillation. The student is trained on DST, DM, and DD, sharing temporal modeling parameters across both input types and bypassing spatial interactions for temporal-only inputs. Snowflakes and flames indicate frozen and trainable parameters, respectively.

Datasets

The authors evaluated T2ST across a wide array of benchmarks:

  • Long-term Forecasting: Evaluated on 6 datasets including CAD3, CAD5, PEMS07M, TrafficSH, CHI-TAXI, and NYC-BIKE.
  • Short-term Forecasting: Evaluated on 12 datasets from the EvalST collection, covering flow, occupancy, speed, and grid-based urban observations (e.g., PEMS-03, METR-LA, Tdrive-BJ).
  • Auxiliary Data: The GiftEvalPretrain corpus was used as the source for auxiliary time-series data.

Key Insights and Ideas

  • Decoupling Time and Space: The "Aha!" moment here is the realization that temporal patterns (trends, seasonality, cycles) are somewhat universal. You can learn "how things change over time" from a generic electricity load dataset and apply that logic to traffic sensors, as long as the spatial dependencies are still grounded in real spatio-temporal observations.
  • The Power of "Diverse" Data: Simply picking data similar to your training set isn't enough. By using a TSFM to filter for reliable but novel patterns, the model becomes much more robust to out-of-distribution scenarios.
  • Relational Alignment over Feature Alignment: In distillation, trying to force a student's hidden vectors to match a teacher's is often a losing battle due to dimension mismatches. Aligning the relational structure of the data is a much more flexible and effective way to transfer knowledge across different model architectures.

Key Results

The results are striking. Using only 10% of the original spatio-temporal data, T2ST achieved:

  • Competitive Performance: On 5 out of 6 long-term datasets, T2ST matched or outperformed the same model trained on 100% of the data.
  • Error Reduction: On the PEMS07M dataset, the Mean Absolute Error (MAE) dropped from 4.53 (full data) to 4.16 (10% data + T2ST), an 8.17% improvement.
  • Short-term Gains: In grid-based urban forecasting (Tdrive-BJ), the Mean Absolute Percentage Error (MAPE) was slashed from roughly 34% to 19% compared to the baseline.
Performance of T2ST across different data budgets

Limitations and Future Work

  • Offline Selection: Currently, the data selection process is an offline "pre-processing" step. The authors suggest that an online, dynamic selection process during training could be even more effective, though computationally heavier.
  • Teacher Dependency: The quality of the STFM is inherently capped by the quality of the TSFM teacher. As time-series foundation models continue to improve, T2ST will likely see "free" performance gains.
  • Spatial Metadata: While T2ST reduces the need for observations, it still requires spatial metadata (like the graph structure) for the target domain. Future work could explore ways to synthesize or transfer spatial knowledge as well.

Summary updated

Abstract

Spatio-temporal foundation models (STFMs) require large-scale training corpora, yet synchronized observations and spatial metadata are costly to acquire and curate, limiting the availability of suitable data. By contrast, time-series data are abundant and easier to collect. Since observations at each location form a time series, this motivates a natural question: can abundant time-series data enhance STFMs when spatio-temporal observations are scarce? To bridge this gap, we propose T2ST, a data-efficient training framework that augments limited spatio-temporal observations with selected auxiliary time series and transfers knowledge from a frozen time-series foundation model. Our key insight is that temporal pattern learning can benefit from abundant external time series, while spatial dependency learning remains grounded in real spatio-temporal observations. Guided by this insight, T2ST retrieves matched sequences to reinforce sparsely observed temporal patterns and selects reliable complementary sequences to broaden temporal coverage. Temporal relation distillation aligns relationships among historical patches, while multi-scale prediction distillation transfers forecasting behavior at fine and coarse temporal resolutions, enabling knowledge transfer across heterogeneous architectures. Using only 10\% of the original spatio-temporal data, T2ST enables STFMs to achieve overall forecasting performance competitive with or superior to their full-data counterparts across six long-term and twelve short-term forecasting datasets. These results establish a new paradigm for data-efficient training of spatio-temporal foundation models. Code is available at https://anonymous.4open.science/r/T2ST-65F3/.
BibTeX
@misc{xu2026t,
  author = {Yue Xu and Kaiyun Wei and Wenying Duan and Zichen Xu and Shuai Ma and Xiaoxi He},
  title = {{T2ST: Data-Efficient Spatio-Temporal Foundation Models via Temporal Data and Knowledge Transfer}},
  year = {2026},
  eprint = {2611.18199},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.18199},
}
Figure