Spatial-Temporal Graph (STG) forecasting on large-scale networks has garnered significant attention. However, existing models predominantly focus on short-horizon predictions and suffer from notorious computational costs and memory consumption when scaling to long-horizon predictions and large graphs. Targeting the above challenges, we present FaST, an effective and efficient framework based on heterogeneity-aware Mixture-of-Experts (MoEs) for long-horizon and large-scale STG forecasting, which unlocks one-week-ahead (672 steps at a 15-minute granularity) prediction with thousands of nodes. FaST is underpinned by two key innovations. First, an adaptive graph agent attention mechanism is proposed to alleviate the computational burden inherent in conventional graph convolution and self-attention modules when applied to large-scale graphs. Second, we propose a new parallel MoE module that replaces traditional feed-forward networks with Gated Linear Units (GLUs), enabling an efficient and scalable parallel structure. Extensive experiments on real-world datasets demonstrate that FaST not only delivers superior long-horizon predictive accuracy but also achieves remarkable computational efficiency compared to state-of-the-art baselines. Our source code is available at: https://github.com/yijizhao/FaST.

In Location-Based Services (LBS), such as food delivery, a fundamental task is segmenting Areas of Interest (AOIs), aiming at partitioning the urban geographical spaces into non-overlapping regions. Traditional AOI segmentation algorithms primarily rely on road networks to partition urban areas. While promising in modeling the geo-semantics, road network-based models overlooked the service-semantic goals (e.g., workload equality) in LBS service. In this article, we point out that the AOI segmentation problem can be naturally formulated as a Markov Decision Process (MDP), which gradually chooses a nearby AOI for each grid in the current AOI’s border. Based on the MDP, we present the first attempt to generalize Deep Reinforcement Learning (DRL) for AOI segmentation, leading to a novel DRL-based framework called DRL4AOI. The DRL4AOI framework introduces different service-semantic goals in a flexible way by treating them as rewards that guide the AOI generation. To evaluate the effectiveness of DRL4AOI, we develop and release an AOI segmentation system. We also present a representative implementation of DRL4AOI—TrajRL4AOI—for AOI segmentation in the logistics service. It introduces a Double Deep Q-learning Network (DDQN) to gradually optimize the AOI generation for two specific semantic goals: (i) trajectory modularity, i.e., maximize tightness of the trajectory connections within an AOI and the sparsity of connections between AOIs, (ii) matchness with the road network, i.e., maximizing the matchness between AOIs and the road network. Quantitative and qualitative experiments conducted on synthetic and real-world data demonstrate the effectiveness and superiority of our method. The code and system is publicly available at https://github.com/Kogler7/AoiOpt .

Detecting anomalies in human mobility is essential for applications such as public safety and urban planning. While traditional anomaly detection methods primarily focus on individual movement patterns (e.g., a child should stay at home at night), collective anomaly detection aims to identify irregularities in collective mobility behaviors across individuals (e.g., a child is at home alone while the parents are elsewhere) and remains an underexplored challenge. Unlike individual anomalies, collective anomalies require modeling spatiotemporal dependencies between individuals, introducing additional complexity. To address this gap, we propose CoBAD, a novel model designed to capture Collective Behaviors for human mobility Anomaly Detection. We first formulate the problem as unsupervised learning over Collective Event Sequences (CES) with a co-occurrence event graph, where CES represents the event sequences of related individuals. CoBAD then employs a two-stage attention mechanism to model both the individual mobility patterns and the interactions across multiple individuals. Pre-trained on large-scale collective behavior data through masked event and link reconstruction tasks, CoBAD is able to detect two types of collective anomalies: unexpected co-occurrence anomalies and absence anomalies, the latter of which has been largely overlooked in prior work. Extensive experiments on large-scale mobility datasets demonstrate that CoBAD significantly outperforms existing anomaly detection baselines, achieving an improvement of 13%-18% in AUCROC and 19%-70% in AUCPR. All source code is available at https://github.com/wenhaomin/CoBAD.

Given the temporal GPS coordinates from a large set of human agents, how can we model their mobility behavior toward effective anomaly (e.g., bad-actor or malicious behavior) detection without any labeled data? Human mobility and trajectory modeling have been extensively studied, showcasing varying abilities to manage complex inputs and balance performance-efficiency trade-offs. In this work, we formulate anomaly detection in complex human behavior by modeling raw GPS data as a sequence of stay-point events, each characterized by spatio-temporal features, along with trips (i.e., commutes) between the stay-points. Our problem formulation allows us to leverage modern sequence models for unsupervised training and anomaly detection. Notably, we equip our proposed model USTAD (for Uncertainty-aware Spatio-Temporal Anomaly Detection) with aleatoric (i.e., data) uncertainty estimation to account for inherent stochasticity in certain individuals' behavior, as well as epistemic (i.e., model) uncertainty to handle data sparsity under a large variety of human behaviors. Together, aleatoric and epistemic uncertainties unlock a robust loss function as well as uncertainty-aware decision-making in anomaly scoring. Extensive experiments show that USTAD significantly outperforms baselines in industry-scale data. We open-source all code at https://github.com/wenhaomin/USTAD.

Regional socioeconomic indicators are critical across various domains, yet their acquisition can be costly. Inferring global socioeconomic indicators from a limited number of regional samples is essential for enhancing management and sustainability in urban areas and human settlements. Current inference methods typically rely on spatial interpolation based on the assumption of spatial continuity, which does not adequately address the complex variations present within regional spaces. In this paper, we present GeoHG, the first space-aware socioeconomic indicator inference method that utilizes a heterogeneous graph-based structure to represent geospace for non-continuous inference. Extensive experiments demonstrate the effectiveness of GeoHG in comparison to existing methods, achieving an R2 score exceeding 0.8 under extreme data scarcity with a masked ratio of 95%. The code and data are available at https://github.com/CityMind-Lab/GeoHG.

Spatio-Temporal (ST) data science, which includes sensing, managing, and mining large-scale data across space and time, is fundamental to understanding complex systems in domains such as urban computing, climate science, and intelligent transportation. Traditional deep learning approaches have significantly advanced this field, particularly in the stage of ST data mining. However, these models remain task-specific and often require extensive labeled data. Inspired by the success of Foundation Models (FM), especially large language models, researchers have begun exploring the concept of Spatio-Temporal Foundation Models (STFMs) to enhance adaptability and generalization across diverse ST tasks. Unlike prior architectures, STFMs empower the entire workflow of ST data science, ranging from data sensing, management, to mining, thereby offering a more holistic and scalable approach. Despite rapid progress, a systematic study of STFMs for ST data science remains lacking. This survey aims to provide a comprehensive review of STFMs, categorizing existing methodologies and identifying key research directions to advance ST general intelligence.

The ever-increasing sensor service, though opening a precious path and providing a deluge of earth system data for deep-learning-oriented earth science, sadly introduce a daunting obstacle to their industrial level deployment. Concretely, earth science systems rely heavily on the extensive deployment of sensors, however, the data collection from sensors is constrained by complex geographical and social factors, making it challenging to achieve comprehensive coverage and uniform deployment. To alleviate the obstacle, traditional approaches to sensor deployment utilize specific algorithms to design and deploy sensors. These methods dynamically adjust the activation times of sensors to optimize the detection process across each sub-region. Regrettably, formulating an activation strategy generally based on historical observations and geographic characteristics, which make the methods and resultant models were neither simple nor practical. Worse still, the complex technical design may ultimately lead to a model with weak generalizability. In this paper, we introduce for the first time the concept of spatio-temporal data dynamic sparse training and are committed to adaptively, dynamically filtering important sensor distributions. To our knowledge, this is the first proposal (termed DynST) of an industry-level deployment optimization concept at the data level. However, due to the existence of the temporal dimension, pruning of spatio-temporal data may lead to conflicts at different timestamps. To achieve this goal, we employ dynamic merge technology, along with ingenious dimensional mapping to mitigate potential impacts caused by the temporal aspect. During the training process, DynST utilize iterative pruning and sparse training, repeatedly identifying and dynamically removing sensor perception areas that contribute the least to future predictions.

Spatiotemporal trajectories are sequences of timestamped locations, which enable a variety of analyses that in turn enable important real-world applications. It is common to map trajectories to vectors, called embeddings, before subsequent analyses. Thus, the qualities of embeddings are very important. Methods for pre-training embeddings, which leverage unlabeled trajectories for training universal embeddings, have shown promising applicability across different tasks, thus attracting considerable interest. However, research progress on this topic faces two key challenges: a lack of a comprehensive overview of existing methods, resulting in several related methods not being well-recognized, and the absence of a unified pipeline, complicating the development of new methods and the analysis of methods. We present UniTE, a survey and a unified pipeline for this domain. In doing so, we present a comprehensive list of existing methods for pre-training trajectory embeddings, which includes methods that either explicitly or implicitly employ pre-training techniques. Further, we present a unified and modular pipeline with publicly available underlying code, simplifying the process of constructing and evaluating methods for pre-training trajectory embeddings. Additionally, we contribute a selection of experimental results using the proposed pipeline on real-world datasets.

As an important data resource containing spatial information, addresses record the geospatial information corresponding to social production activities and human behavioral activities. How to effectively encode addresses has always been a core challenge in the field of Geographic Information Systems (GIS). Pre-trained Models (PTMs) designed for Natural Language Process (NLP) have emerged as the dominant tools for encoding semantic information in text. Though promising, those NLP-based PTMs fall short of encoding geographic knowledge in addresses, which limits their application potential in geospatial tasks. To tackle the above problem, this study proposes a Geography-Graph Pre-trained model (G2PTL) that combines graph learning and text pre-training, aiming to make up for the shortcomings of traditional PTM in the geography field. Specifically, we first utilize real-world delivery data to build a large-scale heterogeneous graph of addresses, which contains abundant geographic knowledge and spatial topology information. Then, G2PTL is pre-trained with subgraphs sampled from the heterogeneous graph. Through experimental evaluation on multiple downstream tasks of GIS, including geocoding, geographic entity prediction, and geographic entity recognition, G2PTL demonstrated significant performance improvements. G2PTL has been successfully deployed in production-level GIS, such as Cainiao's logistics system, effectively improving the execution efficiency and accuracy of address-related tasks. This research not only provides a new technical path for the encoding and processing of geographical information, but also opens up a new perspective for the study of pre-training models in the geographical field. The code resources of the G2PTL model have been opened for research and application developers to access and use at https://huggingface.co/Cainiao-AI/G2PTL.

The query-POI matching task involves interpreting noisy textual addresses to retrieve corresponding Points-of-Interest (POIs), which is crucial for location-based service providers. However, existing methods typically rely on annotated user search logs, limiting their generalization. This paper address the query-POI matching problem through geographical data alignment using a contrastive learning framework. Our model, MoCo-GA (Momentum Contrastive Geographical Alignment), learns similar representations for various geographical data elements of the same POI, including query address, POI address, and geolocation. We developed a method to create a cross-modal geographical dataset from crowd-sourced data for training. Our MoCo-GA employs the momentum contrastive instance discrimination mechanism to learn representations for textual addresses. We further proposed a siamese contrastive learning module to to for geographical data alignment. Experimental results demonstrate that MoCo-GA can consistently outperform baseline methods on query-POI matching task, particularly in zero-shot scenarios. Our code is available at https://github.com/CainiaoTechAi/TextGeoAlign

Time series analysis stands as a focal point within the data mining community, serving as a cornerstone for extracting valuable insights crucial to a myriad of real-world applications. Recent advances in Foundation Models (FMs) have fundamentally reshaped the paradigm of model design for time series analysis, boosting various downstream tasks in practice. These innovative approaches often leverage pre-trained or fine-tuned FMs to harness generalized knowledge tailored for time series analysis. This survey aims to furnish a comprehensive and up-to-date overview of FMs for time series analysis. While prior surveys have predominantly focused on either application or pipeline aspects of FMs in time series analysis, they have often lacked an in-depth understanding of the underlying mechanisms that elucidate why and how FMs benefit time series analysis. To address this gap, our survey adopts a methodology-centric classification, delineating various pivotal elements of time-series FMs, including model architectures, pre-training techniques, adaptation methods, and data modalities. Overall, this survey serves to consolidate the latest advancements in FMs pertinent to time series analysis, accentuating their theoretical underpinnings, recent strides in development, and avenues for future exploration.

Real-world last-mile express datasets are crucial for research in logistics, supply chain management, and spatio-temporal data mining. Despite a plethora of algorithms developed to date, no widely accepted, publicly available last-mile express dataset exists to support research in this field. In this paper, we introduce LaDe, the first publicly available last-mile express dataset with millions of packages from the industry. LaDe has three unique characteristics: (1)Large-scale. It involves 10,677k packages of 21k couriers over 6 months of real-world operation. (2)Comprehensive information. It offers original package information, task-event information, as well as couriers' detailed trajecotries and road networks. (3)Diversity. The dataset includes data from various scenarios, including package pick-up and delivery, and from multiple cities, each with its unique spatio-temporal patterns due to their distinct characteristics such as populations. We verify LaDe on three tasks by running several classical baseline models per task. We believe that the large-scale, comprehensive, diverse feature of LaDe can offer unparalleled opportunities to researchers in the supply chain community, data mining community, and beyond. The dataset and code is publicly available at https://huggingface.co/datasets/Cainiao-AI/LaDe.

Urban region profiling from web-sourced data is of utmost importance for urban computing. We are witnessing a blossom of LLMs for various fields, especially in multi-modal data research such as vision-language learning, where text modality serves as a supplement for images. As textual modality has rarely been introduced into modality combinations in urban region profiling, we aim to answer two fundamental questions: i) Can text modality enhance urban region profiling? ii) and if so, in what ways and which aspects? To answer the questions, we leverage the power of Large Language Models (LLMs) and introduce the first-ever LLM-enhanced framework that integrates the knowledge of text modality into urban imagery, named LLM-enhanced Urban Region Profiling with Contrastive Language-Image Pretraining (UrbanCLIP ). Specifically, it first generates a detailed textual description for each satellite image by Image-to-Text LLMs. Then, the model is trained on image-text pairs, seamlessly unifying language supervision for urban visual representation learning, jointly with contrastive loss and language modeling loss. Results on urban indicator prediction in four major metropolises show its superior performance, with an average improvement of 6.1% on R2 compared to the state-of-the-art methods. Our code and dataset are available at https://github.com/StupidBuluchacha/UrbanCLIP.

Instant delivery services, such as food delivery and package delivery, have achieved explosive growth in recent years by providing customers with daily-life convenience. An emerging research area within these services is service Route&Time Prediction (RTP), which aims to estimate the future service route as well as the arrival time of a given worker. As one of the most crucial tasks in those service platforms, RTP stands central to enhancing user satisfaction and trimming operational expenditures on these platforms. Despite a plethora of algorithms developed to date, there is no systematic, comprehensive survey to guide researchers in this domain. To fill this gap, our work presents the first comprehensive survey that methodically categorizes recent advances in service route and time prediction. We start by defining the RTP challenge and then delve into the metrics that are often employed. Following that, we scrutinize the existing RTP methodologies, presenting a novel taxonomy of them. We categorize these methods based on three criteria: (i) type of task, subdivided into only-route prediction, only-time prediction, and joint route&time prediction; (ii) model architecture, which encompasses sequence-based and graph-based models; and (iii) learning paradigm, including Supervised Learning (SL) and Deep Reinforcement Learning (DRL). Conclusively, we highlight the limitations of current research and suggest prospective avenues. We believe that the taxonomy, progress, and prospects introduced in this paper can significantly promote the development of this field.

Spatio-temporal graph neural networks (STGNN) have emerged as the dominant model for spatio-temporal graph (STG) forecasting. Despite their success, they fail to model intrinsic uncertainties within STG data, which cripples their practicality in downstream tasks for decision-making. To this end, this paper focuses on probabilistic STG forecasting, which is challenging due to the difficulty in modeling uncertainties and complex ST dependencies. In this study, we present the first attempt to generalize the popular de-noising diffusion probabilistic models to STGs, leading to a novel non-autoregressive framework called DiffSTG, along with the first denoising network UGnet for STG in the framework. Our approach combines the spatio-temporal learning capabilities of STGNNs with the uncertainty measurements of diffusion models. Extensive experiments validate that DiffSTG reduces the Continuous Ranked Probability Score (CRPS) by 4%-14%, and Root Mean Squared Error (RMSE) by 2%-7% over existing methods on three real-world datasets.

Instant-logistics (e.g., food delivery and package pick-up) is increasingly calling for Route and Time Prediction (RTP), which aims to predict both future route and arrival time of a courier’s unvisited locations. Accurate RTP can greatly benefit the platform, such as optimizing order dispatching and improving user experience. Although recent years have witnessed various works for solving the RTP problem, they still suffer from the following three limitations: i) Failing to consider the high-level transfer mode of couriers between AOIs (Areas Of Interest, such as residential quarters or office buildings), which can help to build more accurate RTP. ii) Failing to simultaneously make the route and time prediction. Existing works either separately predict route/time or predict them in a two-step way. However, since route and time are strongly correlated (nearby locations in the route should have similar arrival times), jointly predicting them should be more effective. iii) The widely adopted tree-based or sequence-based architecture fails to fully encode the spatial relationship between different locations. To address the above limitations, we propose a multi-level and multi-task graph model, named M2G4RTP, for instant-logistics route and time joint prediction. Specifically, we propose a multi-level graph encoder equipped with a newly-designed GAT-e encoding module to capture couriers’ both high-level transfer modes between AOIs and low-level transfer modes between locations. Moreover, a multi-task decoder is presented to jointly predict the route and time at different levels. Finally, a loss weighting method based on homoscedastic uncertainty is designed to balance the two tasks adaptively. Extensive experiments on an industry-scale real-world dataset, as well as the online deployment on Cainiao Alibaba, demonstrate the superiority of our proposed model.

Last-mile delivery, which refers to delivering packages from the depot to customers, is a crucial step for logistics service. The Route and Time Prediction (RTP) in last-mile package delivery is beneficial to improve customers’ experience and supervise couriers’ behavior. However, the limited raw information brings great challenges to accurately predict the route and delivery time. In this paper, we propose a deep model named I2RTP, which explores the heterogeneous representation of the package’s community to help predict the delivery route and estimate the arrival time of each package. Specifically, for the entire delivery route prediction, we model the inter- and intra-community information to learn the route features from global and local perspectives. Besides, by integrating the community representation with package features, our model could make more accurate predictions of the next-delivery package and its time duration. Experiments on the offline dataset and the online deployment on Cainiao’s Delivery System demonstrate the effectiveness of our proposed method, as well as validate the rationality of the global and local prediction pipeline.

Pick-up and Delivery Route Prediction (PDRP), which aims to estimate the future service route of a worker given his current task pool, has received rising attention in recent years. Deep neural networks based on supervised learning have emerged as the dominant model for the task because of their powerful ability to capture workers' behavior patterns from massive historical data. Though promising, they fail to introduce the non-differentiable test criteria into the training process, leading to a mismatch in training and test criteria. Which considerably trims down their performance when applied in practical systems. To tackle the above issue, we present the first attempt to generalize Reinforcement Learning (RL) to the route prediction task, leading to a novel RL-based framework called DRL4Route. It combines the behavior-learning abilities of previous deep learning models with the non-differentiable objective optimization ability of reinforcement learning. DRL4Route can serve as a plug-and-play component to boost the existing deep learning models. Based on the framework, we further implement a model named DRL4Route-GAE for PDRP in logistic service. It follows the actor-critic architecture which is equipped with a Generalized Advantage Estimator that can balance the bias and variance of the policy gradient estimates, thus achieving a more optimal policy. Extensive offline experiments and the online deployment show that DRL4Route-GAE improves Location Square Deviation (LSD) by 0.9%-2.7%, and Accuracy@3 (ACC@3) by 2.4%-3.2% over existing methods on the real-world dataset.

In intelligent logistics systems, predicting the Estimated Time of Pick-up Arrival (ETPA) of packages is a crucial task, which aims to predict the courier’s arrival time to all the unpicked-up packages at any time. Accurate prediction of ETPA can help systems alleviate customers’ waiting anxiety and improve their experience. We identify three main challenges of this problem. First, unlike the travel time estimation problem in other fields like ride-hailing, the ETPA task is distinctively a multi-destination and path-free prediction problem. Second, an intuitive idea for solving ETPA is to predict the pick-up route and then the time in two stages. However, it is difficult to accurately and efficiently predict couriers’ future routes in the route prediction step since their behaviors are affected by multiple complex factors. Third, furthermore, in the time prediction step, the requirement for providing a courier’s all unpicked-up packages’ ETPA at once in real time makes the problem even more challenging. To tackle the preceding challenges, we propose RankETPA, which integrates the route inference into the ETPA prediction. First, a learning-based pick-up route predictor is designed to learn the route-ranking strategies of couriers from their massive spatial-temporal behaviors. Then, a spatial-temporal attention-based arrival time predictor is designed for real-time ETPA inference via capturing the spatial-temporal correlations between the unpicked-up packages. Extensive experiments on two real-world datasets and a synthetic dataset demonstrate that RankETPA achieves significant performance improvement against the baseline models.

Pick-up and delivery (P&D) services such as food delivery have achieved explosive growth in recent years by providing customers with daily-life convenience. Though many service providers have invested considerably in routing tools, more and more practitioners realize that significant deviations exist between workers' actual routes and planned ones. So it is not wise to feed "optimal routes" as workers' actual service routes into downstream tasks (e.g., arrival-time prediction and order dispatching), whose performances count on the accuracy of route prediction, i.e., to predict the future service route of a worker's unfinished tasks. Therefore, to meet the rising calling for route prediction models that can capture workers' future routing behaviors, in this paper, we formulate the Pick-up and Delivery Route Prediction task (PDRP task for short) from the graph perspective for the first time, then propose a dynamic spatial-temporal graph-based model, named Graph2Route. Unlike previous sequence-based models, our model leverages the underlying graph structure and features into the encoding and decoding process. Moreover, the dynamic graph-based nature can spontaneously describe the evolving relationship between different problem instances. As a result, abundant decision context information and various spatial-temporal information of node/edge can be fully utilized in Graph2Route to improve the prediction performance. Offline experiments over two real-world industry-scale datasets under different P&D services (i.e., food delivery and package pick-up) and online A/B test demonstrate the superiority of our proposed model.

Over 10 billion packages are picked up every day in China. A fundamental task raised in the emerging intelligent logistics systems is the couriers’ package pick-up route prediction, which is beneficial for package dispatching, arrival-time estimation and overdue-risk evaluation, by leveraging the predicted routes to improve those downstream tasks. In the package pick-up scene, the decision-making of a courier is affected by strict spatial-temporal constraints (e.g., package location, promised pick-up time, current time, and courier’s current location). Furthermore, couriers have different decision preferences on various factors (e.g., time factor, distance factor, and balance of both), based on their own perception of the environments and work experience. In this article, we propose a novel model, named DeepRoute+, to predict couriers’ future package pick-up routes according to the couriers’ decision experience and preference learned from the historical behaviors. Specifically, DeepRoute+ consists of three layers: (1) The representation layer produces experience- and preference-aware representations for the unpicked-up packages, in which a decision preference module can dynamically adjust the importance of factors that affects the courier’s decision under the current situation. (2) The transformer encoder layer encodes the representations of packages while considering the spatial-temporal correlations among them. (3) The attention-based decoder layer uses the attention mechanism to generate the whole pick-up route recurrently. Experiments on a real-world logistics dataset demonstrate the state-of-the-art performance of our model.

Dynamic networks are widely used in the social, physical, and biological sciences as a concise mathematical representation of the evolving interactions in dynamic complex systems. Measuring distances between network snapshots is important for analyzing and understanding evolution processes of dynamic systems. To the best of our knowledge, however, existing network distance measures are designed for static networks. Therefore, when measuring the distance between any two snapshots in dynamic networks, valuable context structure information existing in other snapshots is ignored. To guide the construction of context-aware distance measures, we propose a context-aware distance paradigm, which introduces context information to enrich the connotation of the general definition of network distance measures. A Context-aware Spectral Distance (CSD) is then given as an instance of the paradigm by constructing a context-aware spectral representation to replace the core component of traditional Spectral Distance (SD). In a node-aligned dynamic network, the context effectively helps CSD gain mainly advantages over SD as follows: (1) CSD is not affected by isospectral problems; (2) CSD satisfies all the requirements of a metric, while SD cannot; and (3) CSD is computationally efficient. In order to process large-scale networks, we develop a kCSD that computes top- k eigenvalues to further reduce the computational complexity of CSD. Although kCSD is a pseudo-metric, it retains most of the advantages of CSD. Experimental results in two practical applications, i.e., event detection and network clustering in dynamic networks, show that our context-aware spectral distance performs better than traditional spectral distance in terms of accuracy, stability, and computational efficiency. In addition, context-aware spectral distance outperforms other baseline methods.

Over 10 billion packages are picked up every day in China. Accurate prediction of couriers' pick-up routes can help the dispatch system to assign packages to couriers more intelligently, which is able to further increase the pick-up efficiency and reduce the overdue rate. In the package pick-up scene, the decision-making of a courier is quite complex since it's affected by strict spatial-temporal constraints (e.g., package location, promised pick-up time, current time and courier's current location). In this paper, we propose a novel model, named DeepRoute, to predict couriers' future package pick-up routes according to the couriers' decision experience learnt from their historical spatial-temporal behaviors. Specifically, DeepRoute consists of three layers: 1) The representation layer produces experience-aware representations for unpicked-up packages. 2) The transformer encoder layer encodes the representations of packages while considering the spatial-temporal correlations among them. 3) The attention-based decoder layer uses the attention mechanism to generate the whole pick-up route recurrently. Experiments on a real-world logistics dataset demonstrate the state-of-the-art performance of our DeepRoute model.