Time Series Foundation Models (TSFMs) have emerged as a promising approach in time series analysis. Due to the large-scale parameters of TSFMs and pretraining cost, how to adapt TDFMs in streaming data is always the key factor constraining their application effectiveness. Because streaming data often experiences data distribution and task drifts, which cannot be learnt by offline training. Existing methods typically address streaming data modeling with continuous learning through model fine-tuning or model editing. However, fine-tuning incurs significant computational costs, while editing methods can lead to shifts in the original feature space during streaming updates. To address these limitations, we propose a novel Orthogonal Rotation Transformation-based Continuous Learning method, called ORTCL, for TSFMs. Our key insight is to apply orthogonal matrix rotations to the input and output feature spaces of the TSFMs during model editing. This preserves the metric structure of the original feature space and enables new data to be directly mapped into the existing feature space of the TSFMs. Specifically, we obtain the orthogonal matrix for the input layer via singular value decomposition and derive the corresponding transformation matrix for the output layer through least squares optimization. Extensive experimental results demonstrate that ORTCL outperforms existing methods in both single-domain and cross-domain streaming time series forecasting tasks, effectively mitigating catastrophic forgetting.

Customer value prediction is an essential task for effective customer relationship management, particularly for business-to-business (B2B) service platforms. In B2B scenarios, customer value is jointly shaped by customer preferences and platform service characteristics. Prior research mainly focuses on modeling pairwise relationships between individual customers, failing to capture the multifaceted factors influencing customer value on B2B platforms. To fill this gap, we propose a Multi-view Hypergraph Convolutional Network with Counterfactual Optimization (MHCC) to achieve key-factor-aware customer value prediction. First, to capture hierarchical high-order relationships among business customers, we propose a residual multi-view hypergraph convolutional module. It leverages distinct hypergraphs to model customers' industry categories and service interactions. An attention-based residual fusion layer integrates these relational features with historical behavioral features. Second, we propose a counterfactual pruning optimization module that employs counterfactual perturbation and reasoning to quantify the causal influence of individual connections in the hypergraph. This enables the pruning of redundant connections and the identification of key factors driving customer value. Extensive experiments on a large-scale dataset from a real-world B2B platform demonstrate the superior performance of our model, achieving an 8.59 % improvement in MAE over state-of-the-art methods. We further identify distinctive key factors for customers at different value tiers, thereby offering actionable insights for data-driven customer relationship management.

The rapid development of e-commerce systems has led to the accumulation of substantial customer data. Effectively utilizing this data for marketing and customer management is crucial for enterprises. Customer value prediction, a key task, aims to predict the revenue value of customers for the platform. Although various customer value prediction models have been proposed in existing studies, challenges remain in dealing with the complexity of heterogeneous interaction data and quantifying customer value. This is particularly because the process of data collection and obtaining customer value labels heavily relies on extensive manual annotation. To address these challenges, we propose SPHGraph, a self-supervised pre-training framework tailored for heterogeneous graphs, driven by pseudo labels automatically extracted from the data. SPHGraph integrates three core components: a pseudo label initialization module to provide structure-aware supervision, a representation learning module to capture semantic and structural information from heterogeneous interactions, and a label refinement module that iteratively improves learning signals through attention-guided propagation. This design allows the model to learn robust and transferable representations under label-scarce conditions. Extensive experiments across three real-world datasets consistently demonstrate the superior performance of SPHGraph compared to state-of-the-art baselines.

In logistic systems, demand prediction is an essential task providing the basis for improving the quality of terminal services, such as pick-up and delivery efficiency. However, the geographical scope of operations across multiple cities brings challenges due to the sparsity of user behavior data, hindering accurate predictions. Despite cross-city prediction methods potentially solving this problem by relying on the label of overlapping users in different cities, annotating these overlapping users is expensive. Additionally, the dynamic and diverse nature of user behaviors complicates feature transfer between cities. In this work, we define the logistics demand prediction problem as forecasting pick-up and delivery demand for zones, the smallest operational units in logistics systems, in different cities. To address the challenge, we propose TSTL, a Transferable Spatio-Temporal Learning framework for cross-city logistics prediction with sparse user data. TSTL advances existing methods from two aspects: (1) User-level invariant representation module extracts consistent user representations for overlapping and non-overlapping users across cities. (2) User-zone graph aggregation module enhances user embeddings by integrating dynamic interactions, such as logistics behaviors, into inherent user relations. Finally, the multi-city transfer module fine-tunes model parameters for city-invariant knowledge adoption and predicts future logistics demand. We implement and evaluate TSTL on one of the largest logistics systems. Extensive offline experiments and real-world deployment demonstrate the effectiveness of TSTL.

The analysis of express trajectory is essential for delivery service optimization and courier management. Express trajectory consists of a large amount of community paths which is different from the trajectory of transportation tools such as taxis which run in the main road of a city. Due to the complexity of the community road networks and the energy constraint of devices, trajectories are collected under a low sampling rate and accuracy. These trajectories need to be recovered and mapped to the road networks for further analysis. Current recovery approaches utilize two-stage or end-to-end models to generate the upsampled trajectories. These methods fail to fully exploit the periodicity relation in historical data, which leads to accuracy losses for complex community road networks. In this paper, we design PATR, a periodicity-aware Seq2Seq model, to recover the couriers' trajectories by leveraging the periodical patterns of the historical data. We conduct experimental evaluations based on real-world GPS data of JD logistics consisting of 836,959 points and 25,726 road segments, and our model outperforms the state-of-the-art baselines by reducing recovery error by 2.4%-21.9%.