Predicting human mobility patterns is crucial for various applications, including urban planning, traffic management, and public health. A key challenge in this domain is identifying when and where interactions, or contacts , occur between moving individuals (i.e., objects) based on their recent positions. To address this, we represent object movement using a mobility network , where individuals establish connections when they come into close physical proximity. This network framework enables the analysis of individual interactions within a given time frame, providing valuable insights into mobility dynamics. Traditional methods for predicting future interactions typically focus on predicting static snapshots of the mobility network at future time points. These temporal link prediction methods are limited because they do not provide detailed insights into short-term mobility dynamics. In a mobility network, being able to predict both long-term and short-term contacts is crucial in many scenarios: for instance, health authorities may require daily forecasts to plan resource allocation and anticipate disease spread, whereas emergency response teams rely on hour-by-hour (or even minute-by-minute) predictions for real-time outbreak control. In this article, we present MobiNetForecast , a two‐stage framework that (i) individually predicts each user’s next \( k \) locations using a Transformer‐based decoder with constrained beam search to enforce spatial continuity and (ii) infers future contacts by identifying spatiotemporal co‐locations in the predicted trajectories. By unifying short‐ and long‐term forecasts in a single pipeline, our approach can anticipate both recurring and previously unseen interactions. Empirical evaluation on real‐world ( GeoLife ) and large‐scale synthetic ( SFCO‐3K ) trajectory datasets demonstrates that MobiNetForecast outperforms state‐of‐the‐art temporal link prediction methods (e.g., EvolveGCN , ROLAND , DyTed , AGCRN ) by up to two orders of magnitude in F1‐score, and improves over leading sequence models (e.g., Flashback++ ) by 15–20% on key metrics across multiple granularities. These results validate the effectiveness of trajectory‐driven contact prediction for accurate, multi‐scale mobility network forecasting.
Machine learning models are trained on datasets where undesired features have spurious correlations with the target variable, resulting in the bias shifts of the model output to a different class. In federated learning (FL), a common problem is data distribution skew. Since each client has a different distribution, models trained on these local datasets are likely to learn different biases. This phenomenon will degrade the performance of the aggregated global model and then prevent the global model from learning actual intrinsic decisions. Consequently, addressing and mitigating these biases in FL is critical to the overall effectiveness of the global model. Most existing works are designed for centralized machine learning using data augmentation or regularization terms. Nevertheless, these methods may not generalize effectively when directly migrated to federated learning scenarios due to the lack of direct access to the external data of other clients. Recent work for debiasing in FL is based on data augmentation in image space, which generates new bias-conflicting images for further training, leading to high costs and limited ability to create sufficiently diverse and realistic samples. To tackle these challenges, we propose a novel feature-level augmentation approach, named FedInt, for debiasing in federated learning. The feature-level augmentation in latent space allows the model to capture and smooth out different directions of the decision boundary, making it possible to synthesize diverse bias-conflicting samples for each client with lower training costs. Experiments on three datasets demonstrate the effectiveness of our proposed method in terms of debiasing performance, and it outperforms traditional data augmentation techniques and various typical FL methods. Our code is available on an anonymous GitHub repository: https://anonymous.4open.science/r/FedInt-C710889456/ .
Graph Neural Networks (GNNs) have achieved remarkable success, yet they are susceptible to encoding and amplifying societal biases. Current research in fair graph learning predominantly focuses on developing specialized debiasing techniques through input data preprocessing, training optimization, or model architecture modifications under the assumption that the GNN backbone is an interchangeable component capable of independently generating a debiased representation space. This article revisits this prevalent paradigm by demonstrating that the expressiveness of backbone GNNs critically influences fairness outcomes. Through spectral-domain theoretical analysis, we establish the link between GNN expressiveness and the elimination of linear statistical bias. We formally prove that GNN architectures lacking spectral expressiveness impose strict constraints on the representation space, so that harmful linear correlations between sensitive attributes and target prediction logits are preserved whenever a low-expressive backbone is paired with a debiasing operator acting within that backbone’s representation space. Universal spectral expressiveness is therefore a necessary condition for eliminating linear prediction-sensitive correlation in this family of methods, although it is not sufficient on its own. Driven by this finding, we propose a fair graph learning framework that pairs a universally expressive GNN backbone, utilizing powerful graph-wise filtering and node-wise signal transformation, with a lightweight linear decorrelation module operating in the logit space. Comprehensive evaluations on six real-world datasets show that FairFormer attains a strong accuracy–fairness tradeoff and ranks as the best or near-best fairness method on most benchmarks. On Bail and Pokec_n, however, a linear decorrelation module alone does not dominate every baseline on every fairness metric. Our results are consistent with the necessity of spectral expressiveness for linear decorrelation of prediction and sensitive attributes, establishing expressiveness as a key design axis for trustworthy GNN design. Code is available at https://github.com/qslim/expressive-fair-learning .
Adapting large language models (LLMs) for time series forecasting introduces unique challenges due to the mismatch between continuous time series tokens and the discrete word tokens LLMs are trained on. While LLMs have shown success on natural language tasks, aligning time series tokens with the semantic space of LLMs while preserving both accuracy and interpretability remains an open problem. Existing reprogramming and prompt-based approaches often generate word embeddings that are misaligned with or semantically irrelevant to the corresponding time series. For example, a segment with an increasing trend may be mapped to words like ”mirror” or ”keyboard”, which makes them hard to interpret. To bridge the gap, we propose MLTA-LLM, a Multi-Level Text Alignment framework for time series forecasting with LLMs. MLTA-LLM decomposes time series into trend, seasonal, and residual components, each of which is aligned with semantic anchor embeddings or text prototype embeddings. This component-specific alignment forces the LLM to process time series data through structured, interpretable representations. In addition, we integrate conformal prediction to provide statistically rigorous prediction intervals for uncertainty quantification. Experiments on multiple benchmark datasets show that MLTA-LLM not only improves forecasting accuracy over state-of-the-art baselines but also yields more interpretable representations and better-calibrated uncertainty estimates.
Graph Federated Learning (FL) allows multiple agents to collaboratively train a global graph model while preserving data privacy. However, current methods overlook the issue of low-quality or malicious data contributions, which can degrade the global model’s performance. This article addresses this gap by proposing a novel incentive mechanism for graph FL, called GFLValue. We identify two main challenges in graph FL: agents causing harm and those contributing with delays. Our framework introduces an agent valuation function based on two criteria: gradient alignment and graph diversity. To balance accuracy and fairness, motif prototypes are used to enhance the global model aggregation and support local optimization. Additionally, we incorporate Gaussian noise into GFLValue to ensure Differential Privacy (DP), creating a privacy-preserving version, GFLValue+. In the case of malicious agents, we observe that local and global motif prototypes can distinguish between malicious and honest agents. This leads to the GFLValue-Adv framework, which integrates motif prototype distances and agent values to detect malicious behavior. Experiments show GFLValue offers the best tradeoff between accuracy and fairness, while GFLValue+ enhances privacy without significantly compromising performance. GFLValue-Adv improves malicious agent detection accuracy by 11.7% and reduces training time by 6.94%.
