Pre-training universal models across multiple domains to enhance downstream tasks is a prevalent learning paradigm. However, there has been minimal progress in pre-training transferable models across domains for time series representation. This dilemma is incurred by two key factors: the limited availability of training set within each domain and the substantial differences in data characteristics between domains. To address these challenges, we present a novel framework, namely CrossTimeNet, designed to perform cross-domain self-supervised pre-training to benefit target tasks. Specifically, to address the issue of data scarcity, we utilize a pre-trained language model as the backbone network to effectively capture the sequence dependencies of the input time series. Meanwhile, we adopt the recovery of corrupted region inputs as a self-supervised optimization objective, taking into account the locality of the time series. To address discrepancies in data characteristics, we introduce a novel tokenization module that converts continuous time series inputs into discrete token sequences using vector quantization techniques. This approach facilitates the learning of transferable time series models across different domains. Extensive experimental results on diverse time series tasks, including classification and forecasting, demonstrate the effectiveness of our approach. Our codes are publicly available at https://github.com/Mingyue-Cheng/CrossTimeNet.

For the advancement of time series classification, we can summarize that most existing methods adopt a common learning-to-classify paradigm - a classifier model tries to learn the relation between sequence inputs and target label encoded by one-hot distribution. Although effective, this paradigm conceals two inherent limitations: (1) one-hot distribution fails to reflect the comparability and similarity between labels, and (2) it is difficult to learn transferable representation across domains. In this work, we propose InstructTime, a novel attempt to reshape time series classification as a learning-to-generate paradigm. Relying on the generative capacity of the pre-trained language model, the core idea is to formulate the classification of time series as a multimodal understanding task. Specifically, firstly, a time series discretization module is designed to convert continuous inputs into a sequence of discrete tokens to solve the inconsistency issue across modality data. Secondly, we introduce an alignment projected layer before feeding the transformed token of time series into language models. Thirdly, prior to fine-tuning the language model for the target domain, it is essential to emphasize the necessity of auto-regressive pre-training across various modality inputs. Finally, extensive experimentation are conducted on several prevalent public benchmark datasets, indicating the superior performance of the InstructTime. Our code is at https://github.com/Mingyue-Cheng/InstructTime.

Nowadays, couriers are still the main solution to address the "last mile" problem in logistics. They are usually required to record the delivery time of each parcel manually, which is essential for delivery insurances, delivery performance evaluations, and customer available time discovery. Stay points extracted from couriers' trajectories provide a chance to fill the delivery time automatically to ease their burdens. However, it is challenging due to inaccurate delivery locations and various stay scenarios. To this end, we propose the improved Delivery Time Inference (DTInf+), to infer the delivery time of waybills based on couriers' trajectories. Our solution is composed of three steps: 1) Data Pre-processing, which organizes waybills and stay points by delivery trips, 2) Delivery Location Mining, which obtains the delivery location for each address and each Geocoded waybill location by mining historical delivery caused stay points, and 3) Delivery Event-based Matching, which jointly selects the best-matched stay point for waybills at the same delivery location based on Pointer Network-like model SPSelector to infer the delivery time. Extensive experiments and case studies based on real-world datasets from JD Logistics confirm the effectiveness of our approach. Finally, a system powered by DTInf+ is deployed in JD Logistics.

Delivery locations are fundamental data source for intelligent logistics, which can be used in route planning, arrival time estimation, parcel allocation, etc. Using the Geocoded way-bill location of an address as the delivery location is not sufficient, due to wrong address parsing, coarse-grained POI database, or different preferences of customers. To mitigate the insufficiency of Geocoding, some methods have been proposed, which utilize couriers' locations when waybills are confirmed to be delivered for delivery location inference. Nevertheless, these methods highly rely on the quality of couriers' annotations and fail when couriers confirm deliveries with delays. We propose to infer actual delivery locations of addresses from couriers' trajectories. This idea lies on an observation that the semantics of delivering a parcel are well captured by couriers' trajectories (e.g., a stay point would be generated when a delivery occurs), which holds even couriers confirm deliveries with delays. Specifically, we design Delivery Location Inference under Mis-Annotation (DLInfMA), which (1)generates location candidates from stay points in couriers' trajectories; (2) extracts features from both an address and its location candidates; and (3) uses an attention-based neural network model LocMatcher to predict the delivery location for each address. Experiments on two real-world datasets from JD Logistics as well as synthetic datasets demonstrate the effectiveness, robustness and scalability of DLInfMA. We also present a deployed system along with two applications based on DLInfMA.

Service time is a part of time cost in the last-mile delivery, which is the time spent on delivering parcels at a certain location. Predicting the service time is fundamental for many downstream logistics applications, e.g., route planning with time windows, courier workload balancing and delivery time prediction. Nevertheless, it is non-trivial given the complex delivery circumstances, location heterogeneity, and skewed observations in space. The existing solution trains a supervised model based on aggregated features extracted from parcels to deliver, which cannot handle above challenges well. In this paper, we propose MetaSTP, a meta-learning based neural network model to predict the service time. MetaSTP treats the service time prediction at each location as a learning task, leverages a Transformer-based representation layer to encode the complex delivery circumstances, and devises a model-based meta-learning method enhanced by location prior knowledge to reserve the uniqueness of each location and handle the imbalanced distribution issue. Experiments show MetaSTP outperforms baselines by at least 9.5% and 7.6% on two real-world datasets. Finally, an intelligent waybill assignment system based on MetaSTP is deployed and used internally in JD Logistics.

The rapid development of e-commerce requires efficient and reliable logistics services. Nowadays, couriers are still the main solution to address the "last mile" problem in logistics. They are usually required to record the accurate delivery time of each parcel manually, which provides vital information for applications like delivery insurances, delivery performance evaluations, and customer available time discovery. Couriers' trajectories generated by their PDAs provide a chance to infer the delivery time automatically to ease the burdens on the couriers. However, directly using the nearest stay point to infer the delivery time is under satisfactory due to two challenges: 1) inaccurate delivery locations, and 2) various stay scenarios. To this end, we propose Delivery Time Inference (DTInf), to automatically infer the delivery time of waybills based on couriers' trajectories. Our solution is composed of three steps: 1) Data Pre-processing, which detects stay points from trajectories, and separates stay points and waybills by delivery trips, 2) Delivery Location Correction, which infers true delivery locations of waybills by mining historical deliveries, and 3) Delivery Event-based Matching, which selects the best-matched stay point for waybills in the same delivery location to infer the delivery time. Extensive experiments and case studies based on large scale real-world waybill and trajectory data from JD Logistics confirm the effectiveness of our approach. Finally, we introduce a system based on DTInf, which is deployed and used internally in JD Logistics.