The World Wide Web increasingly relies on intelligent services that require accurate time series forecasting, from urban mobility platforms to adaptive web-based decision systems. In practice, building effective forecasting models typically requires abundant high-quality data, which may not always be available in all cities due to sensing limitations or data sparsity. To address this challenge, transfer learning methods aim to transfer knowledge from data-rich source cities to data-scarce target cities. However, source and target data distributions are often not identical: while some patterns from source cities may be beneficial, others can be irrelevant or even misleading. Existing transfer learning methods generally train the target model using all available source data without explicitly distinguishing between useful and non-useful knowledge, which may hinder performance. In this work, we propose xRAG4TS, a novel framework that integrates Retrieval-Augmented Generation (RAG) with Large Language Models (LLMs) for cross-city time series forecasting. xRAG4TS introduces a Cross-City Selective Retriever Module that filters semantically relevant historical patterns and documents from data-rich source cities, and incorporates them as structured prompts in an LLM Inference Module to guide forecasting in data-scarce target cities. By enabling selective, interpretable, and context-aware knowledge transfer, our method enhances robustness and scalability in web-oriented spatio-temporal applications. Extensive experiments on real-world cross-city datasets demonstrate that xRAG4TS significantly outperforms state-of-the-art baselines, highlighting its potential for powering adaptive and trustworthy web services under severe data scarcity.

Large Language Models (LLMs) have achieved prominent success in various applications, driven by their foundational capabilities and strong generalization potential. While their impact on natural language processing is well-established, recent works highlight their significant promise in other applications as well. One such area is traffic forecasting, where LLMs demonstrated the ability to generate powerful analytical insights, offering new opportunities for advancing Intelligent Transportation Systems (ITS). This survey summarizes 160 related works, providing a comprehensive review of methods, applications, and challenges. Specifically, we summarize the efforts to bridge the gap between time series data and language models, exploring their design principles and applications in various traffic scenarios, including traffic forecasting, traffic recommendation, mobility forecasting, urban management, signal control, and safety analysis. For deeper insights, we discuss existing challenges and highlight future research directions. This survey identifies three key limitations of current works: (1) LLM-based traffic models face practical deployment hurdles; (2) there lacks a unified benchmark across various applications; (3) resident privacy protection demands attention in their real-world applications. By providing a foundational understanding of LLMs for traffic forecasting, this survey aims to benefit not only the traffic mining community but also contribute to the broader advancement of Artificial General Intelligence (AGI).

Time series forecasting fundamentally relies on accurately modeling complex interdependencies and shared patterns within time series data. Recent advancements, such as Spatio-Temporal Graph Neural Networks (STGNNs) and Time Series Foundation Models (TSFMs), have demonstrated promising results by effectively capturing intricate spatial and temporal dependencies across diverse real-world datasets. However, these models typically require large volumes of training data and often struggle in data-scarce scenarios. To address this limitation, we propose a framework named Few-shot Spatio-Temporal Large Language Models (FSTLLM), aimed at enhancing model robustness and predictive performance in few-shot settings. FSTLLM leverages the contextual knowledge embedded in Large Language Models (LLMs) to provide reasonable and accurate predictions. In addition, it supports the seamless integration of existing forecasting models to further boost their predicative capabilities. Experimental results on real-world datasets demonstrate the adaptability and consistently superior performance of FSTLLM over major baseline models by a significant margin. Our code is available at: https://github.com/JIANGYUE61610306/FSTLLM.

The widespread adoption of location-based services has led to the generation of vast amounts of mobility data, providing significant opportunities to model user movement dynamics within urban environments. Recent advancements have focused on adapting Large Language Models (LLMs) for mobility analytics. However, existing methods face two primary limitations: inadequate semantic representation of locations (i.e., discrete IDs) and insufficient modeling of mobility signals within LLMs (i.e., single templated instruction fine-tuning). To address these issues, we propose QT-Mob, a novel framework that significantly enhances LLMs for mobility analytics. QT-Mob introduces a location tokenization module that learns compact, semantically rich tokens to represent locations, preserving contextual information while ensuring compatibility with LLMs. Furthermore, QT-Mob incorporates a series of complementary fine-tuning objectives that align the learned tokens with the internal representations in LLMs, improving the model's comprehension of sequential movement patterns and location semantics. The proposed QT-Mob framework not only enhances LLMs' ability to interpret mobility data but also provides a more generalizable approach for various mobility analytics tasks. Experiments on three real-world dataset demonstrate the superior performance in both next-location prediction and mobility recovery tasks, outperforming existing deep learning and LLM-based methods.

Missing values pose a formidable obstacle in multivariate time series analysis. Existing imputation methods rely on entangled representations that struggle to simultaneously capture multiple orthogonal time-series patterns, leading to suboptimal performance and limited interpretability. Meanwhile, requiring the entire data span as input renders these models impractical for long time series. To address these issues, we propose${\sf TIDER}$and its enhanced version,${\sf AdaTIDER}$.${\sf TIDER}$employs low-rank matrix factorization and disentangled temporal representations to model intricate dynamics like trend, seasonality, and local bias. However,${\sf TIDER}$is limited to single-period modeling and does not explicitly capture dependencies between channels. To overcome these limitations,${\sf AdaTIDER}$incorporates adaptive cross-channel dependency modeling and multi-period seasonality representations. These advancements enable it to dynamically capture variable relationships and complex multi-period patterns, significantly enhancing imputation accuracy and interpretability, while maintaining${\sf TIDER}$'s scalability. Extensive experiments on real-world datasets validate the superiority of our models in imputation accuracy, scalability, interpretability, and robustness.

Time series forecasting is essential for our daily activities and precise modeling of the complex correlations and shared patterns among multiple time series is essential for improving forecasting performance. Spatial-Temporal Graph Neural Networks (STGNNs) are widely used in multivariate time series forecasting tasks and have achieved promising performance on multiple real-world datasets for their ability to model the underlying complex spatial and temporal dependencies. However, existing studies have mainly focused on datasets comprising only a few hundred sensors due to the heavy computational cost and memory cost of spatial-temporal GNNs. When applied to larger datasets, these methods fail to capture the underlying complex spatial dependencies and exhibit limited scalability and performance. To this end, we present a Scalable Adaptive Graph Diffusion Forecasting Network (SAGDFN) to capture complex spatial-temporal correlation for large-scale multivariate time series and thereby, leading to exceptional performance in multivariate time series forecasting tasks. The proposed SAGDFN is scalable to datasets of thousands of nodes without the need of prior knowledge of spatial correlation. Extensive experiments demonstrate that SAGDFN achieves comparable performance with state-of-the-art baselines on one real-world dataset of 207 nodes and outperforms all state-of-the-art baselines by a significant margin on three real-world datasets of 2000 nodes.

In online complex systems such as transportation system, an important work is real-time traffic prediction. Due to the data shift, data model inconsistency, and sudden change of traffic patterns (like transportation accident), the prediction result derived from an offline-built model would be unreliable. Retraining the model is usually not time affordable for online prediction, especially when the prediction model is very complex and costs a lot of training time (for example, deep neural networks). A real-time prediction correction strategy would be of great value under this situation. Traditionally, the prediction correction usually relies on the prediction error in several previous time intervals. They assume that the error pattern is similar in the current time interval, so that it is time-delayed to some extent. In this article, we propose the prediction correction strategy using the reconstruction error in the deep neural network. The reconstruction error can reflect the model’s ability on feature representation and then determine the fitness of an input data to the model. We first build the relationship between reconstruction error and prediction error. From the perspective of the prediction interval, we demonstrate that the reconstruction error is in positive relation with the prediction interval. Thus the prediction result is more reliable when the reconstruction error is smaller. Then we propose two mechanisms of real-time prediction correction using the reconstruction error. The data driven prediction correction approach selects several training instances with similar reconstruction errors to the current instance and using their average prediction error in correcting the prediction result. The model-driven approach builds several component deep neural networks in training. The component training set for each network is selected according to the reconstruction error of training instances. For a predicting instance, it first computes the reconstruction error of the sample in each component network and then averages the results by the reconstruction error and prediction interval. The model-driven approach is actually a reconstruction error-based deep neural network ensemble approach. Finally, a series of experiments demonstrated that reconstruction error based prediction correction approaches are effective in several prediction problems in transportation including traffic flow prediction on road, traffic flow prediction in entrance and exit station and travel time prediction. Besides the high overall accuracy, our approach can also provide many observations of using the reconstruction error in transportation prediction.