Time-series forecasting in real-world applications such as finance and energy often faces challenges due to limited training data and complex, noisy temporal dynamics. Existing deep forecasting models typically supervise predictions using full-length temporal windows, which include substantial high-frequency noise and obscure long-term trends. Moreover, auxiliary variables containing rich domain-specific information are often underutilized, especially in few-shot settings. To address these challenges, we propose LoFT-LLM, a frequency-aware forecasting pipeline that integrates low-frequency learning with semantic calibration via a large language model (LLM). Firstly, a Patch Low-Frequency forecasting Module (PLFM) extracts stable low-frequency trends from localized spectral patches. Secondly, a residual learner then models high-frequency variations. Finally, a fine-tuned LLM refines the predictions by incorporating auxiliary context and domain knowledge through structured natural language prompts. Extensive experiments on financial and energy datasets demonstrate that LoFT-LLM significantly outperforms strong baselines under both full-data and few-shot regimes, delivering superior accuracy, robustness, and interpretability.

Accurate classification of healthcare time series is critical for clinical decision-making. However, existing models often struggle under real-world data shifts and lack interpretability- two key requirements for reliable medical deployment. To address these challenges, we propose SHINE, a novel endto-end framework that learns disentangled and shift-invariant representations by modeling the generative process of multivariate healthcare signals. Specifically, SHINE first introduces a genuine data representation learning that disentangles healthcare signals into trend, seasonality, and noise components, reflecting distinct temporal dynamics of healthcare series. Then, we inject several inductive biases into each component to encourage latent representations to be invariant to data shifts and aligned with their corresponding semantic units. Extensive experiments on six healthcare benchmarks spanning ECG, EEG, and continuous glucose monitoring (CGM) domains-under a variety of simulated real-world shift scenarios-demonstrate that SHINE consistently outperforms state-of-the-art baselines, providing robust performance and clinically meaningful interpretations grounded in the estimated components.

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.

Over the past decade, smartphones have become indispensable personal mobile devices, experiencing a remarkable surge in software apps. These apps empower users to seamlessly connect with various internet services, such as social communication and online shopping. Accurately predicting smartphone app usage can effectively improve user experience and optimize resource utilization. However, existing models often treat app usage prediction as a classification problem, which suffers from issues of app usage imbalance and out-of-distribution (OOD) during deployment. To address these challenges, this paper proposes a novel click-through rate (CTR) ranking-based method for predicting app usage. By transforming the classification problem into a CTR problem, we can eliminate the negative impact of the app usage imbalance issue. To address the OOD issue during deployment, we generate the app click sequence and three types of discriminative features, which enable generalization on unseen apps. The app click sequence and the three types of features serve as inputs for training a CTR estimation model in the cloud, and the trained model is then deployed on the user's smartphone to predict the CTR for each installed app. The decision-making process involves ranking these CTR values and selecting the app with the highest CTR as the final prediction. Our method has been extensively tested with large-scale app usage data. The results demonstrate that our approach is able to outperform state-of-the-art methods, with improvements over 4.93% in top-3 accuracy and 6.64% in top-5 accuracy. It achieves approximately twice the accuracy in predicting apps with low usage frequencies in comparison to baseline methods. Our method has been successfully deployed on the app recommendation system of a leading smartphone manufacturer.

Urban grid prediction can be applied to many classic spatial-temporal prediction tasks such as air quality prediction, crowd density prediction, and traffic flow prediction, which is of great importance to smart city building. In light of its practical values, many methods have been developed for it and have achieved promising results. Despite their successes, two main challenges remain open: (a) how to well capture the global dependencies and (b) how to effectively model the multi-scale spatial-temporal correlations? To address these two challenges, we propose a novel method— DeepMeshCity , with a carefully-designed Self-Attention Citywide Grid Learner ( SA-CGL ) block comprising of a self-attention unit and a Citywide Grid Learner ( CGL ) unit. The self-attention block aims to capture the global spatial dependencies, and the CGL unit is responsible for learning the spatial-temporal correlations. In particular, a multi-scale memory unit is proposed to traverse all stacked SA-CGL blocks along a zigzag path to capture the multi-scale spatial-temporal correlations. In addition, we propose to initialize the single-scale memory units and the multi-scale memory units by using the corresponding ones in the previous fragment stack, so as to speed up the model training. We evaluate the performance of our proposed model by comparing with several state-of-the-art methods on four real-world datasets for two urban grid prediction applications. The experimental results verify the superiority of DeepMeshCity over the existing ones. The code is available at https://github.com/ILoveStudying/DeepMeshCity.

The rapid growth of location-based services (LBS) has yielded massive amounts of data on human mobility. Effectively extracting meaningful representations for user-generated check-in sequences is pivotal for facilitating various downstream services. However, the user-generated check-in data are simultaneously influenced by the surrounding objective circumstances and the user's subjective intention. Specifically, the temporal uncertainty and spatial diversity exhibited in check-in data make it difficult to capture the macroscopic spatial-temporal patterns of users and to understand the semantics of user mobility activities. Furthermore, the distinct characteristics of the temporal and spatial information in check-in sequences call for an effective fusion method to incorporate these two types of information. In this paper, we propose a novel Spatial-Temporal Cross-view Contrastive Representation (STCCR) framework for check-in sequence representation learning. Specifically, STCCR addresses the above challenges by employing self-supervision from “spatial topic” and “temporal intention” views, facilitating effective fusion of spatial and temporal information at the semantic level. Besides, STCCR leverages contrastive clustering to uncover users’ shared spatial topics from diverse mobility activities, while employing angular momentum contrast to mitigate the impact of temporal uncertainty and noise. We extensively evaluate STCCR on three real-world datasets and demonstrate its superior performance across three downstream tasks.

Accurate traffic forecasting is critical in improving safety, stability, and efficiency of intelligent transportation systems. Despite years of studies, accurate traffic prediction still faces the following challenges, including modeling the dynamics of traffic data along both temporal and spatial dimensions, and capturing the periodicity and the spatial heterogeneity of traffic data, and the problem is more difficult for long-term forecast. In this paper, we propose an Attention based Spatial-Temporal Graph Neural Network (ASTGNN) for traffic forecasting. Specifically, in the temporal dimension, we design a novel self-attention mechanism that is capable of utilizing the local context, which is specialized for numerical sequence representation transformation. It enables our prediction model to capture the temporal dynamics of traffic data and to enjoy global receptive fields that is beneficial for long-term forecast. In the spatial dimension, we develop a dynamic graph convolution module, employing self-attention to capture the spatial correlations in a dynamic manner. Furthermore, we explicitly model the periodicity and capture the spatial heterogeneity through embedding modules. Experiments on five real-world traffic flow datasets demonstrate that ASTGNN outperforms the state-of-the-art baselines.

Geographical topic models have been used to mine geo-tagged documents for topical region and geographical topics, and also have applications in recommendations, user mobility modeling, event detection, etc. Existing studies focus on learning effective geographical topic models while ignoring the efficiency issue. However, it is very expensive to train geographical topic models — it may take days to train a geographical topic model of a small scale on a collection of documents with millions of word tokens. In this paper, we propose the first distributed solution, called${\sf PGeoTopic}$, for training geographical topic models. The proposed solution comprises several novel technical components to increase parallelism, reduce memory requirement, and reduce communication cost. Experiments show that our approach for mining geographical topic models is scalable with both model size and data size on distributed systems.

In this work, we propose a robust road network representation learning framework called Toast, which comes to be a cornerstone to boost the performance of numerous demanding transport planning tasks. Specifically, we first propose a traffic context aware skip-gram module to incorporate auxiliary tasks of predicting the traffic context of a target road segment. Furthermore, we propose a trajectory-enhanced Transformer module that utilizes trajectory data to extract traveling semantics on road networks. Apart from obtaining effective road segment representations, this module also enables us to obtain the route representations. With these two modules, we can learn representations which can capture multi-faceted characteristics of road networks to be applied in both road segment based applications and trajectory based applications. Last, we design a benchmark containing four typical transport planning tasks to evaluate the usefulness of Toast and comprehensive experiments verify that Toast consistently outperforms the state-of-the-art baselines across all tasks.

Recommender systems play a vital role in modern web services. In a typical recommender system, we are given a set of observed user-item interaction records and seek to uncover the hidden behavioral patterns of users from these historical interactions. By exploiting these hidden patterns, we aim to discover users’ personalized tastes and recommend them new items. Among various types of recommendation methods, the latent factor collaborative filtering models have dominated the field. In this paper, we develop a unified view for the existing latent factor models from a probabilistic perspective. The unified framework enables us to discern the underlying connections of different latent factor models and deepen our understandings of their advantages and limitations. In particular, we observe that the loss functions adopted by the existing models are oblivious to the geometry induced by the item-similarity. To address this, we propose a novel model—SinkhornCF—based on Sinkhorn divergence. To address the challenge of the expensive computational cost of Sinkhorn divergence, we also propose new techniques to enable the resulting model to be able to scale to large datasets. Its effectiveness is verified on two real-world recommendation datasets.

As a fundamental component in location-based services, inferring the relationship between points-of-interests (POIs) is very critical for service providers to offer good user experience to business owners and customers. Most of the existing methods for relationship inference are not targeted at POI, thus failing to capture unique spatial characteristics that have huge effects on POI relationships. In this work we propose PRIM to tackle POI relationship inference for multiple relation types. PRIM features four novel components, including a weighted relational graph neural network, category taxonomy integration, a self-attentive spatial context extractor, and a distance-specific scoring function. Extensive experiments on two real-world datasets show that PRIM achieves the best results compared to state-of-the-art baselines and it is robust against data sparsity and is applicable to unseen cases in practice.

Similar trajectory search is a crucial task that facilitates many downstream spatial data analytic applications. Despite its importance, many of the current literature focus solely on the trajectory’s spatial similarity while neglecting the temporal information. Additionally, the few papers that use both the spatial and temporal features based their approach on a traditional point-to-point comparison. These methods model the importance of the spatial and temporal aspect of the data with only a single, pre-defined balancing factor for all trajectories, even though the relative spatial and temporal balance can change from trajectory to trajectory. In this article, we propose the first spatio-temporal, deep-representation-learning-based approach to similar trajectory search. Experiments show that utilizing both features offers significant improvements over existing point-to-point comparison and deep-representation-learning approach. We also show that our deep neural network approach is faster and performs more consistently compared to the point-to-point comparison approaches.

In this paper, we study the problem of predicting the most likely traveling route on the road network between two given locations by considering the real-time traffic. We present a deep probabilistic model-DeepST-which unifies three key explanatory factors, the past traveled route, the impact of destination and real-time traffic for the route decision. DeepST explains the generation of next route by conditioning on the representations of the three explanatory factors. To enable effectively sharing the statistical strength, we propose to learn representations of K-destination proxies with an adjoint generative model. To incorporate the impact of real-time traffic, we introduce a high dimensional latent variable as its representation whose posterior distribution can then be inferred from observations. An efficient inference method is developed within the Variational Auto-Encoders framework to scale DeepST to large-scale datasets. We conduct experiments on two real-world large-scale trajectory datasets to demonstrate the superiority of DeepST over the existing methods on two tasks: the most likely route prediction and route recovery from sparse trajectories. In particular, on one public large-scale trajectory dataset, DeepST surpasses the best competing method by almost 50% on the most likely route prediction task and up to 15% on the route recovery task in terms of accuracy.

Travel time estimation of a given route with respect to real-time traffic condition is extremely useful for many applications like route planning. We argue that it is even more useful to estimate the travel time distribution, from which we can derive the expected travel time as well as the uncertainty. In this paper, we develop a deep generative model - DeepGTT - to learn the travel time distribution for any route by conditioning on the real-time traffic. DeepGTT interprets the generation of travel time using a three-layer hierarchical probabilistic model. In the first layer, we present two techniques, amortization and spatial smoothness embeddings, to share statistical strength among different road segments; a convolutional neural net based representation learning component is also proposed to capture the dynamically changing real-time traffic condition. In the middle layer, a nonlinear factorization model is developed to generate auxiliary random variable i.e., speed. The introduction of this middle layer separates the statical spatial features from the dynamically changing real-time traffic conditions, allowing us to incorporate the heterogeneous influencing factors into a single model. In the last layer, an attention mechanism based function is proposed to collectively generate the observed travel time. DeepGTT describes the generation process in a reasonable manner, and thus it not only produces more accurate results but also is more efficient. On a real-world large-scale data set, we show that DeepGTT produces substantially better results than state-of-the-art alternatives in two tasks: travel time estimation and route recovery from sparse trajectory data.

As a fundamental problem in social influence propagation analysis, learning influence parameters has been extensively investigated. Most of the existing methods are proposed to estimate the propagation probability for each edge in social networks. However, they cannot effectively learn propagation parameters of all edges due to data sparsity, especially for the edges without sufficient observed propagation. Different from the conventional methods, we introduce a novel social influence embedding problem, which is to learn parameters for nodes rather than edges. Nodes are represented as vectors in a low-dimensional space, and thus social influence information can be reflected by these vectors. We develop a new model Inf2vec, which combines both the local influence neighborhood and global user similarity to learn the representations. We conduct extensive experiments on two real-world datasets, and the results indicate that Inf2vec significantly outperforms state-of-the-art baseline algorithms.

Trajectory similarity computation is fundamental functionality with many applications such as animal migration pattern studies and vehicle trajectory mining to identify popular routes and similar drivers. While a trajectory is a continuous curve in some spatial domain, e.g., 2D Euclidean space, trajectories are often represented by point sequences. Existing approaches that compute similarity based on point matching suffer from the problem that they treat two different point sequences differently even when the sequences represent the same trajectory. This is particularly a problem when the point sequences are non-uniform, have low sampling rates, and have noisy points. We propose the first deep learning approach to learning representations of trajectories that is robust to low data quality, thus supporting accurate and efficient trajectory similarity computation and search. Experiments show that our method is capable of higher accuracy and is at least one order of magnitude faster than the state-of-the-art methods for k-nearest trajectory search.

Air quality is one of the most important environmental concerns in the world, and it has deteriorated substantially over the past years in many countries. For example, Chinese Academy of Social Sciences reports that the problem of haze and fog in China is hitting a record level, and China is currently suffering from the worst air pollution. Among the various causal factors of air quality, particulate matter with a diameter of 2.5 micrometers or less (i.e., PM2.5) is a very important factor; governments and people are increasingly concerned with the concentration of PM2.5. In many cities, stations for monitoring PM2.5 concentration have been built by governments or companies to monitor urban air quality. Apart from monitoring, there is a rising demand for finding pollution sources of PM2.5 and discovering the transmission of PM2.5 based on the data from PM$_{2.5}$ monitoring stations.