Urban mobility data has significant connections with economic growth and plays an essential role in various smart-city applications. Due to privacy concerns and substantial data collection costs, fine-grained human mobility trajectories are challenging to make publicly available on a large scale. A promising solution to address this issue is trajectory synthesizing, which generates synthetic trajectories that preserve aggregate spatiotemporal distributions. However, existing works often neglect the road network structural constraints or rely on instance-level external supervision, thus limiting their scalability in generating fine-grained and high-fidelity trajectories. In this paper, we propose Cardiff, a coarse-to-fine Cascaded hybrid diffusion-based framework for fine-grained and structure-plausible trajectory generation. By leveraging the hierarchical nature of urban mobility, Cardiff decomposes the generation process into two cascaded levels, i.e., discrete road segment-level and continuous fine-grained GPS-level: (i) At the segment level, to reduce computational costs and redundancy in raw trajectories, we first encode the discrete road segments into low-dimensional latent embeddings and design a diffusion transformer-based latent denoising network for segment-level synthesis. (ii) Taking the first stage of generation as conditions, we then design a fine-grained GPS-level conditional denoising network with a noise augmentation mechanism to achieve road-network-constrained and fine-grained generation. The cascaded progressive generation yields high-fidelity fine-grained trajectories while adhering to road geometry and topology constraints. Experimental results on three large real-world trajectory datasets demonstrate that our method outperforms state-of-the-art baselines in various metrics. The code is available at~ https://github.com/urban-mobility-generation/Cardiff.

Points of Interest (POIs) are crucial for urban planning and location-based services. However, the high dynamics and expensive updating costs of POIs pose a key roadblock for their urban applications. This is especially true for developing countries, where active economic activities lead to frequent POI updates (e.g., merchants closing down and new ones opening). Therefore, POI updating, i.e., detecting new POIs and different names of the same POIs (alias) to update the POI database, has become an urgent but challenging problem to address. In this paper, we attempt to answer the research question of how to detect and update large-scale POIs via a low-cost approach. To do so, we propose a novel mobile sensing framework called UrbanPOI, which consists of three key modules: (i) a hierarchical POI candidate generation module based on the POINet model that detects POIs from shipping addresses; (ii) a new POI detection module based on the Siamese Attention Network that models multi-modal data and crowd intelligence; and (iii) a POI profiling model to infer the semantic category and the spatial location of updated POIs. Extensive results on real-world logistics delivery datasets show that our model outperforms state-of-theart models in Beijing City by 26.2% in precision and 10.7

Human behavior prediction is an essential AI-based task, which has inspired many real-world applications. In last-mile logistics, predicting couriers' behavior can benefit the couriers' preference learning and workflow optimization. In this paper, we devote to the behavioral prediction of courier workload and quantify their workload by the working time spent at each area of interest (AOI). Considering the behavior interpretability of inverse reinforcement learning (IRL), existing studies have applied IRL to some real-world transportation prediction scenarios. However, in last-mile logistics, the platform assigns multiple orders to each courier, and couriers also receive new tasks in real-time, which additionally influence the couriers' subsequent decisions. The uncertainty in decision spaces and dynamic the workflow distribution make it more challenging to predict the couriers' working time. In this paper, we propose CourIRL, a practical IRL-based framework leveraging cross-attention to integrate Couriers' historical and spatio-temporal features to predict their future working time. CourIRL formulates the couriers' pick-up and delivery tour as a sequential decision-making process and designs a model-free IRL to learn decision-making preference vectors. A multi-head cross-attention mechanism-based deep regression model is proposed for fine-grained working-time prediction. The results of extensive experiments on two real-world datasets demonstrate that the proposed CourIRL surpasses the state-of-the-art baselines by an average of 6.11% across settings, showing the efficacy and potential contributions of CourIRL in last-mile logistics.

In on-demand delivery,online orders are delivered by couriers from merchants to customers within a short time (e.g., 45 minutes). An important task is to provide an efficient order dispatching solution. Existing studies focus on scenarios with stable routing behavior using pre-determined courier-order matching before delivery while ignoring real-time dynamics during delivery. In this work, we leverage courier-courier encounter events as an opportunity to enable cooperative order dispatching (i.e., conducting order transfers among couriers during delivery) for better delivery efficiency. However, it is non-trivial to conduct encounter-aware cooperative order dispatching in real-time dynamics due to two major challenges: (i) the dynamic nature of encounters in diverse real-world scenarios, and (ii) global delivery efficiency optimization by local order transfers. To address the above challenges, we design a detection-driven cooperative dispatching framework, called DECO. Specifically, we design (i) a Received Signal Strength Indicator (RSSI) variance-based state encoder to model encounter dynamics, (ii) an encounter event selector to choose encounter scenarios, (iii) a time-constrained order mask module to filter unsuitable orders, and (iv) an encounter-aware order transfer scheduler to make detailed order transfer decisions. Extensive experiments on real-world data from two large companies (i.e., JD Logistics, Eleme) show that DECO outperforms other baselines.Real-world deployment results at JD Logistics show that DECO improves the order overdue rate by 4.8%.

Recent years have witnessed significant advancements in O2O logistics, which require predicting the volume of shipments generated by customers, commonly referred to as customer value. The essence of accurately predicting customer value in O2O logistics involves analyzing both online buying habits and offline logistics operations. Existing customer value prediction efforts focus solely on online or offline features, making them unsuitable for O2O scenarios. In this paper, we investigate the integration of both online and offline features for customer value prediction, which faces challenges including (i) data silos issues between logistics platforms and e-commerce platforms, and (ii) feature heterogeneity between online and offline data. To address these challenges, we propose a Periodic Asynchronous Vertical Federated Learning framework with adaptive Feature Selection (PAVFL-FS), enabling logistics platforms to efficiently and accurately predict customer value in collaboration with the e-commerce platforms. PAVFL-FS consists of two components: (i) a periodic asynchronous vertical federated learning algorithm to handle data silos problem and enable efficient model training; (ii) a local feature selection algorithm based on stochastic gates to address the cross-platform feature heterogeneity. We have conducted experimental validation on a large-scale real-world dataset collected from a major logistics company in China, including over 2.4 million waybill records and over 6 million online sales records from more than 3,500 merchants. The experimental results demonstrate that PAVFL-FS achieves a mean absolute error of 0.52 in O2O logistics customer value prediction, outperforming 24.6% to the baseline.

Logistics advertising is an emerging task in online-to-offline logistics systems, where logistics companies expand parcel shipping services to new users through advertisements on shopping websites. Compared to existing online e-commerce advertising, logistics advertising has two significant new characteristics: (i) the complex factors in logistics advertising considering both users' offline logistics preference and online purchasing profiles; and (ii) data sparsity and mutual relations among multiple steps due to longer advertising conversion processes. To address these challenges, we design MCAC, a Multi-task Conditional Attention network-based logistics advertising Conversion prediction framework, which consists of (i) an offline shipping preference extraction model to extract the user's offline logistics preference from historical shipping records, and (ii) a multi-task conditional attention-based conversion rate prediction module to model mutual relations among multiple steps in logistics advertising conversion processes. We evaluate and deploy MCAC on one of the largest e-commerce platforms in China for logistics advertising. Extensive offline experiments show that our method outperforms state-of-the-art baselines in various metrics. Moreover, the conversion rate prediction results of large-scale online A/B testing show that MCAC achieves a 15.22% improvement compared to existing industrial practices, which demonstrates the effectiveness of the proposed framework.

Geocoding, associating textual addresses with corresponding GPS coordinates, is vital for many location-based services (e.g., logistics, ridesharing, and social networks). One of the most common Geocoding solutions is using commercial map services such as Google Maps. However, this is typically not practical for some location-based service providers due to real-world challenges like commercial competition and high costs (recurring fees). In this paper, we design a new cost-effective Geocoding framework to automatically infer the geographic coordinates from textual addresses. To achieve this, we take the E-Commerce logistics service as a concrete scenario and designCoMiner, an unsupervised coordinate inference framework based on textual address data, delivery event data, and courier trajectory data.CoMinerincludes three main components, (1) A POI-level clustering model, (2) A Delivery Mobility Graph (DMG), and (3) A behavior-driven address ranking model. Furthermore, we designCoMiner-W, a coordinates mining algorithm based on WiFi data, to further enhance the effectiveness ofCoMiner. We conduct extensive experiments on three large-scale datasets whereCoMineroutperforms the state-of-the-art methods by 20.3%. Moreover, we have designed an abnormal delivery event detection system based onCoMinerand deployed it at JD Logistics, which brings a significant reduction in abnormal delivery event rates.

On-demand delivery has experienced rapid growth in recent years, revolutionizing people's lifestyles with its timeliness and convenience. The order dispatch process in on-demand delivery isconcurrent, wherein couriers continuously accept new orders and deliver them to customers within strict time constraints and dynamic demand and supply. Most of the existing order dispatch mechanisms are designed for independent dispatch or concurrent dispatch without strict deadlines, rendering them unsuitable for real-time concurrent dispatch in on-demand delivery. To address the challenge, we propose aTime-ConstrainedActor-Critic Reinforcement learning based concurrent dispatch system calledTCAC-Dispatchto reduce the overdue rate and enhance the long-term revenue. Specifically, we first design a deep matching network (DMN) with a variable action space, which integrates both states embedding (including route behaviors encoding) and actions' embedding into a long-term value for dispatching decisions. Additionally, we design a time-constrained action pruning module to ensure compliance with time constraints. Then we utilize the Actor-Critic framework to tackle the concurrent dispatch considering strict time constraints and stochastic demand-supply. To further optimize the efficiency and delivery resource utilization, we propose an extension of TCAC (i.e., TCAC+), which consists of (i) a learning-based order service time prediction module to determine whether to relax the deadline of some orders; and (ii) a multi-critic framework to optimize concurrent order dispatch with both tight deadlines and relaxed deadlines using dynamic weighting mechanism. We evaluate the TCAC-Dispatch with one-month data involved with 36.48 million orders and 42,000 couriers collected from Eleme, one of the largest on-demand delivery companies in China. Experiments are conducted on a data-driven emulator deployed on the development environment of Eleme and the results demonstrate that our method outperforms state-of-the-art baselines with various metrics in both tight deadline and mixed deadline scenarios.

Route prediction in instant delivery is still challenging due to the unique characteristics compared with conventional delivery services, such as strict deadlines, overlapped delivery time of multiple orders, and diverse individual preferences on delivery routes. Recently, development in the mobile Internet of Things (IoT) offers the opportunity to collect multi-sensor data with rich real-time information. Therefore, this study proposes a route prediction model called Roupid, which leverages multi-sensor data to improve the accuracy of route prediction in instant delivery. Specifically, we design a 3-Conversion Network-based route prediction framework to take full advantage of various information provided by multi-sensor data, including the encounter data sensed by Bluetooth low energy (BLE) beacons, active site data reported by smart handheld devices, and trajectory data detected by GPS. The 3-Conversion Network we propose is based on a deep neural network framework, which integrates an improved relational graph attention network with edge features (RGATE) to encode global information that couriers typically consider when planning routes. We evaluate our Roupid with real-world data collected from one of the largest instant delivery companies in the world, i.e., Eleme. Experimental results show that our Roupid outperforms other state-of-the-art baselines and offers up to 85.51% of the route prediction precision.

The popularity of online e-commerce has promoted the rapid development of last-mile logistics in recent years. In last-mile services, to ensure delivery efficiency and enhance user experience, the delivery zone is proposed to perform delivery task assignment, which is a fundamental part of last-mile delivery. Each courier is responsible for one delivery zone. Couriers will collect orders belonging to their delivery zones from the delivery station and deliver orders to customers. Existing delivery zone partition practices in last-mile logistics consist of manual experience-based and static optimization-based methods, which perform order amount balancing among different zone but suffer from dissatisfaction and inefficiency because of two limitations: (i) using order amount is not always a good balancing metric considering deliveries' various difficulties (e.g., residence or industrial park, with or without elevators); (ii) less considering couriers' familiarity and preference behaviors. To generate delivery zone partition with equitable workload assignment, in this paper, we propose E-partition, a data-driven delivery zone partition framework to achieve equitable workload assignment in last-mile logistics. We first design a learning-based workload prediction model to estimate service time given a partition plan that consists of unseen courier-zone matching scenarios. Then, a delivery zone partition algorithm is proposed to iterative optimize couriers' core-AOI (i.e., area of interest) generation and AOI assignment process. Extensive offline experimental results show that our model outperforms baselines in working time prediction and workload balancing performances. Real-world deployment results at JD Logistics also verify the effectiveness of equitable-assignment aware delivery zone partition, with a 2.2% increase in service on-time rate compared to state-of-practice partition solutions.

In recent years, we have been witnessing a rapid prevalence of instant delivery services (e,g., UberEats, Instacart, and Eleme) due to their convenience and timeliness. A unique characteristic of instant delivery services is the concurrent dispatch mode, where (i) one courier usually simultaneously delivers multiple orders, especially during rush hours, and (ii) couriers can receive new orders when delivering existing orders. Most existing concurrent dispatch systems are efficiency-oriented, which means they usually dispatch a group of orders that have a similar delivery route to a courier. Although this strategy may achieve high overall efficiency, it also potentially causes a huge disparity of earnings between different couriers. To address the problem, in this paper, we design a Fairness-aware Concurrent dispatch system called FairCod, which aims to optimize the overall operation efficiency and individual fairness at the same time. Specifically, in FairCod, we design a Dynamic Advantage Actor-Critic algorithm with Fairness constrain (DA2CF). The basic idea is that it includes an Actor network to make dispatch decisions based on dynamic action space and a Critic network to evaluate the dispatch decisions from the fairness perspective. More importantly, we extensively evaluate our FairCod system based on one-month real-world data consisting of 36.38 million orders from 42,000 couriers collected by one of the largest instant delivery companies in China. Experimental results show that our FairCod improves courier fairness by 30.3% without sacrificing the overall system benefit compared to state-of-the-art baselines.

On-demand delivery has become prevailing for people to order meals and groceries online, especially during the pandemic. It is essential to dispatch massive orders to limited couriers to satisfy on-demand delivery users, especially during peak hours. Existing studies mainly focus on order dispatching within a region, and they are challenging to be applied to the cross-region courier displacement problem due to (1) unique practical factors, including regional spatial-temporal demand-supply dynamics and strict delivery time constraints, and (2) the large-scale setting and high-dimensional decision space given massive couriers in on-demand delivery. To address these challenges, in this work, we propose an efficient cross-region courier displacement framework, i.e.,CourierDisplacementReinforcementLearning (short forCDRL) based on centralized multi-agent actor-critic, which first design the actor-critic network with a time-varying displacement intensity control module to capture demand-supply dynamics and utilize the centralized training and decentralized execution multi-agent framework to address the large-scale coordination. One-month real-world order records collected from one of the biggest on-demand delivery services in the world are utilized to show the performance of our design. The extensive results show that our method offers a 47.97% of increase in balancing supply and demand and reduces idle ride time by 24.62% simultaneously.

Most commercial players adopt adaptive bitrate (ABR) algorithms to dynamically decide each chunk's bitrate based on the perceived network bandwidth and buffer occupancy. However, current ABR algorithms are agnostic of audio bitrate selection since they deem it has negligible influence on video bitrate selection due to small size of audio chunks. Nevertheless, with the development of audio technologies, the bitrate of audio content increases dramatically in recent years. Thus, inappropriate audio selection can significantly affect video selection and deteriorate the viewing experience. To tackle these inefficiencies, we propose a deepReinforcement learning-based ABR algorithm that takesAudio andVideo quality into account (RAV) to circumvent a series of suboptimal performances, like low playback quality, frequent playback interruptions, poor playback smoothness, and undesirable combinations of video and audio chunks. Furthermore, RAV trains a neural network model that automatically outputs the bitrates for future audio and video chunks without relying on any presumptions about the environment, achieving good robustness to a broad spectrum of conditions. By conducting trace-driven and real-world experiments, we demonstrate that RAV significantly ameliorates the average overall viewing quality by 37.96%-118.20% over the state-of-the-art ABR algorithms. In addition, we also conduct subjective experiments by inviting 32 volunteers, and 27/32 users strongly agree that RAV provides them a better viewing experience than existing ABR solutions.

The popularity of e-commerce has promoted the rapid development of the logistics industry in recent years. As an important step in logistics, last-mile delivery from delivery stations to customers' addresses is now mainly finished by couriers, which requires accurate workload assessment based on actual efforts. However, the state-of-the-practice assessment methods neglect a vital factor that orders with the same customer's address (i.e., Homogeneous orders) can be delivered in a group (i.e., in a single trip) or separately (i.e., in multiple trips). It would cause unfair assessment among couriers if following the same rule. Thus, grouping homogeneous order accurately in the workload assessment is significant for achieving fair courier's workload assessment. To this end, we design, implement, and deploy a nationwide homogeneous order grouping system called FHOG for improving the accuracy of homogeneous order grouping in last-mile delivery for fair courier's workload assessment. FHOG utilizes the courier's reporting behavior for order inspection, collection, and delivery to identify homogeneous orders in the delivery station simultaneously for homogeneous order grouping. Compared with the state-of-the-practice method, our evaluation shows FHOG can effectively reduce order amounts with the higher and lower assessed courier's workload. We further deploy FHOG online in 8336 delivery stations to provide homogeneous order grouping service for more than 120 thousand couriers and 12 million daily orders. The results of the two surveys show that the couriers' acceptance rate is improved by 67% with FHOG after the promotion.

Geocoding, associating textual addresses with corresponding GPS coordinates, is vital for many location-based services (e.g., logistics, ridesharing, and social networks). One of the most common Geocoding solutions is using commercial map services (e.g., Google Maps) by uploading textual addresses to obtain corresponding coordinates. However, this is typically not practical for some location-based service providers due to real-world challenges like commercial competition and high costs (recurring fees). In this paper, we design a new cost-effective Geocoding framework to automatically infer the geographic coordinates from textual addresses for service providers. To achieve this, we take the E-Commerce logistics service as a concrete scenario and design CoMiner, an unsupervised coordinate inference framework based on textual address data, delivery event data, and courier trajectory data. There are three main components in CoMiner. (1) A POI-level clustering model by modeling customers' shopping patterns at different spatial granularities; (2) A Delivery Mobility Graph (DMG) by modeling couriers' delivery events and geographic coordinates; (3) A behavior-driven address ranking model by mining couriers' uncertain reporting behaviors to further infer coordinates on DMG. We extensively verify the performance of CoMiner with a three-phase evaluation from data-driven experiments to real-world deployment. (i) We conduct extensive experiments on three large-scale datasets where CoMiner achieves an average accuracy of 95.1%, which outperforms the state-of-the-art methods by 20.3%. (ii) We deploy CoMiner in JD Logistics, inferring coordinates for over 30 million addresses with an average accuracy of 93.3%. (iii) We utilize CoMiner for two Geocoding-based applications, i.e., parcel re-routing optimization and abnormal delivery event detection.

Region-level sales prediction is an essential task in the e-commerce industry. Accurate demand prediction enables enterprises to respond to future sales changes in time, thereby saving resources and improving efficiency. Existing regional sales prediction solutions focus on utilizing single-source temporal features (e.g., date types and historical sales data) to conduct time-series forecasting, which have not yet modeled the correlations of features among intra and inter regions explicitly (e.g., the regional population, age structure, sex ratio, and spatial correlations of multi-regions). In this paper, we propose a novel long-term regional sales prediction framework to take the Intra- and Inter-region features' correlations into consideration while making long-term region sales prediction. Specifically, the Intra- and Inter-region features are first encoded by the graph convolutional network (GCN), which simultaneously considers static features and contextual information of similar regions during the encoding process. Then, we design a transformer-based model to conduct sales prediction considering both Intra- and Inter-region features embeddings and long-term temporal features. To show the effectiveness of our model, we evaluate our method based on 8-month real-world data consisting of 389 regions and 312 million sales records from one of the largest e-commerce platforms in China. The experimental results show that the prediction results are significantly improved by an average of 20% compared with other state-of-the-art baselines.

The emergence of Online-to-Offline (O2O) stores based on delivery platforms (e.g., Uber Eats, DoorDash, and Eleme) provides great convenience to people's lives. In the O2O model, one of the essential problems for merchants is to select a suitable store site, i.e., store site recommendation problem. We argue that the existing works for the traditional brick-and-mortar stores cannot address this problem due to two unique factors in the O2O model including (i) dynamic supply caused by courier capacity and dispatching strategies and (ii) various customer demands caused by delivery distance and customer preferences. To incorporate these new factors, we design$O^{2}$SiteRec, a store site recommendation method under the O2O model via multi-graph attention networks, which consists of (i) a courier capacity model based on a multi-semantic relation graph attention network to capture courier capacity; (ii) a heterogeneous multi-graph based recommendation model, where the courier capacity, customer preferences, and context features are fused. We evaluate our method based on one-month real-world data consisting of 39,465 stores and 23.6 million orders from one of the largest O2O platforms in China. Experimental results demonstrate that our method outperforms state-of-the-art baselines in various metrics.

Real-time Vehicle-of-Interest (VoI) detection is becoming a core application to smart cities, especially in areas with high accident rates. With the increasing number of surveillance cameras and the advanced developments in edge computing, video tasks prefer to run on edge devices close to cameras due to the constraints of bandwidth, latency, and privacy concerns. However, resource-constrained edge devices are not competent for dynamic traffic loads with resource-intensive video analysis models. To address this challenge, we propose RT-VeD, a real-time VoI detection system based on the limited resources of edge nodes. RT-VeD utilizes multi-granularity computer vision models with different resource-accuracy trade-offs. It schedules vehicle tasks based on a traffic-aware actor-critic framework to maximize the accuracy of VoI detection while ensuring an inference time-bound. To evaluate the proposed RT-VeD, we conduct extensive experiments based on a real-world vehicle dataset. The experiment results demonstrate that our model outperforms other competitive methods.

Postoperative pain cause discomfort to the patient, and even postoperative complications in severe cases, which suggests there is a severe need for predicting the postoperative pain. A number of studies have investigated the correlation between different physiological parameters and nociception, and developed indicators for evaluating the degree of intraoperative nociception. However, these technologies require additional monitoring equipment, which increases the difficulty of deployment and popularization of postoperative pain prediction. In this paper, We propose DoseGuide, a graph-based dynamic time-aware prediction system based on the patient data collected from existing standard infrastructure. DoseGuide takes as input the static physical data and the dynamic intraoperative data of the patient, and output the prediction of postoperative pain level for the certain patient, in which the two types of features are fused via a hybrid feature encoder. Additionally, a graph attention mechanism is introduced to utilize the similarity relationships between patients, which promoted the accuracy of prediction further. We evaluate the system with the medical records of 999 patients undergoing cardiothoracic surgery in the Fourth Affiliated Hospital of Zhejiang University School of Medicine. The Experimental results show that our model achieves 78% accuracy for postoperative pain, and has the best comprehensive performance in comparison with baselines.

Compared with conventional delivery services, instant delivery usually provides a stricter constraint on delivery time (e.g., 30 minutes). To guarantee the quality of time constraint service, precisely predicting the courier’s actual route plays an important role in order dispatching. Most of the existing studies on route prediction are based on single-source data-set such as GPS trajectories or order waybills information, and are not significant to accurately predict the courier’s route. This paper focuses on fully leveraging multi-source data to improve the accuracy of route prediction, including the encounter data, active site report data and GPS trajectories. To achieve this, we propose a multi-source data fusion framework for route prediction. It consists of (i) a multi-source features extracting and fusion module to address the challenge of the heterogeneity of multisource data; (ii) a prediction module taking full advantage of features with different aspects of information containing noise. We evaluate our approach with real-world data collected from one of the largest instant delivery companies in China, i.e., Eleme. Experimental results show that the performance of our multisource data fusion-based prediction model outperforms other state-of-the-art baselines, and achieves a precision of 83.08% for route prediction.

Instant delivery has developed rapidly in recent years and significantly changed the lifestyle of people due to its timeliness and convenience. In instant delivery, the order dispatch process is concurrent. Couriers take new orders continuously and deliver multiple orders in a delivery trip (i.e., a batch). The delivery time of orders in a batch is often overlapped and interlinked with each other. The pickup and delivery sequence of the existing orders in a batch changes dynamically due to time constraints and real-time overdue possibility (i.e., the rate of deliveries that are not finished in promised time). Most of existing order dispatch mechanisms are designed for independent order dispatch or concurrent delivery without strict time constraints, hence are incapable of handling real-time concurrent dispatch with strict time constraints in on-demand instant delivery. To address the challenge, we propose a Time-Constrained Actor- Critic Reinforcement learning based concurrent dispatch system called TCAC-Dispatch to enhance the long-term overall revenue and reduce the overdue rate. Specifically, we design a deep matching network (DMN) with a variable action space, which integrates the state embedding (including route behaviors encoding) and actions embedding features into a long-term matching value. Then the Actor-Critic model tackles the concurrent order dispatch problem considering strict time constraints and stochastic demand-supply in instant delivery. An estimated time-based action pruning module is designed to ensure time constraints guarantee and accelerate the training as well as dispatching processes. We evaluate the TCAC-Dispatch with one-month data involved with 36.48 million orders and 42,000 couriers collected from one of the largest instant delivery companies in China, i.e., Eleme. Empirical experiments are conducted on a data-driven emulator deployed on the development environment of Eleme and results show that our method achieves 22% of the increase in total revenue and reduces the overdue rate by 21.6%.