Qiyuan Chen, Hongsen Huang, Jiahe Chen, Qian Shao, Jintai Chen, Hongxia Xu, Renjie Hua, Ren Chuan, Jian Wu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Incremental few-shot semantic segmentation (iFSS) aims to incrementally acquire new knowledge from limited labeled samples while retaining previously learned concepts, without relying on large-scale manual annotations. Despite its practical importance, effective iFSS methods remain limited. In this paper, we propose a multimodal framework to address this challenge. First, textual features are leveraged to supervise the classifier weights, mitigating forgetting of base classes and overfitting to novel classes. Second, class prototype images are generated from textual features to support low-cost replay of previous knowledge. Finally, a momentum-based updating strategy is introduced to decouple the background and novel class weights within the classifier of the old model used for knowledge distillation. Extensive experiments and ablation studies validate the effectiveness of the approach. The method achieves strong performance in continual learning of novel classes while preserving knowledge of old ones, closely mirroring human-like few-shot learning over time.

Semi-supervised community detection seeks to find a specified community type when only few communities are labeled. Existing “select-then-refine” pipelines often start from mis-aligned cores and rely on Reinforcement-Learning or Gen-erative Adversarial Network, increasing computational cost and limiting scalability. We address these issues with a unified energy framework under crystallization kinetics that jointly models energy, structure, and growth. Based on this perspective, we pro-pose CLique ANNealing (CLANN), which first employs Nucleus Proposer to select candidate clique as community core under four physics-inspired criteria. A learning-free Transitive Annealer then iteratively merges neighboring cliques and repositions the nucleus, enabling spontaneous, scalable community growth. Evaluated on diverse real-world and synthetic networks, CLANN surpasses state-of-the-art baselines by a wide margin while running faster on large graphs, demonstrating that the energy-driven crystallization kinetics framework is both princi-pled and practical for semi-supervised community detection.

Qiyuan Chen, Hongsen Huang, Qian Shao, Jiahe Chen, Jintai Chen, Hongxia Xu, Renjie Hua, Ren Chuan, Jian Wu. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

In response to the burgeoning cryptocurrency sector and its associated financial risks, there is a growing focus on detecting fraudulent activities and malicious addresses. Traditional studies are limited by their reliance on comprehensive historical data and address-wise manipulation, which are not available for early malice detection and fail to identify addresses controlled by the same fraudulent entity. We thus introduceEvolve Path Tracer, a novel solution designed for early malice detection in cryptocurrency. This system innovatively incorporates Asset Transfer Paths and corresponding path graphs in an evolve model, which effectively characterize rapidly evolving transaction patterns. First, for the target address, theClustering-based Path Selectorweight each Asset Transfer Path by finding sibling addresses along the Asset Transfer Paths.Evolve Path Encoder LSTMandEvolve Path Graph GCNthen encode the asset transfer path and path graph within a dynamic structure. Additionally, ourHierarchical Survival Predictorefficiently scales to predict the address labels, demonstrating high scalability and efficiency. We rigorously testedEvolve Path Traceron three real-world datasets of malicious addresses, where it consistently outperformed existing state-of-the-art methods. Our extensive scalability tests further confirmed the model's robust adaptability in dynamic prediction environments, highlighting its potential as a significant tool in the realm of cryptocurrency security.

Since its advent in 2009, Bitcoin (BTC) has garnered increasing attention from both academia and industry. However, due to the massive transaction volume, no systematic study has quantitatively measured the asset decentralization degree specifically from a network perspective.In this paper, by conducting a thorough analysis of the BTC transaction network, we first address the significant gap in the availability of full-history BTC graph and network property dataset, which spans over 15 years from the genesis block (1st March, 2009) to the 845651-th block (29, May 2024). We then present the first systematic investigation to profile BTC’s asset decentralization and design several decentralization degrees for quantification. Through extensive experiments, we emphasize the significant role of network properties and our network-based decentralization degree in enhancing Bitcoin analysis. Our findings demonstrate the importance of our comprehensive dataset and analysis in advancing research on Bitcoin’s transaction dynamics and decentralization, providing valuable insights into the network’s structure and its implications. The whole transaction data is available at dataset link.

Complex planning and scheduling problems have long been solved using various optimization or heuristic approaches. In recent years, imitation learning that aims to learn from expert demonstrations has been proposed as a viable alternative to solving these problems. Generally speaking, imitation learning is designed to learn either the reward (or preference) model or directly the behavioral policy by observing the behavior of an expert. Existing work in imitation learning and inverse reinforcement learning has focused on imitation primarily in unconstrained settings (e.g., no limit on fuel consumed by the vehicle). However, in many real-world domains, the behavior of an expert is governed not only by reward (or preference) but also by constraints. For instance, decisions on self-driving delivery vehicles are dependent not only on the route preferences/rewards (depending on past demand data) but also on the fuel in the vehicle and the time available. In such problems, imitation learning is challenging as decisions are not only dictated by the reward model but are also dependent on a cost-constrained model. In this paper, we provide multiple methods that match expert distributions in the presence of trajectory cost constraints through (a) Lagrangian-based method; (b) Meta-gradients to find a good trade-off between expected return and minimizing constraint violation; and (c) Cost-violation-based alternating gradient. We empirically show that leading imitation learning approaches imitate cost-constrained behaviors poorly and our meta-gradient-based approach achieves the best performance.

Semi-Supervised Learning (SSL) has become a preferred paradigm in many deep learning tasks, which reduces the need for human labor. Previous studies primarily focus on effectively utilising the labelled and unlabeled data to improve performance. However, we observe that how to select samples for labelling also significantly impacts performance, particularly under extremely low-budget settings. The sample selection task in SSL has been under-explored for a long time. To fill in this gap, we propose a Representative and Diverse Sample Selection approach (RDSS). By adopting a modified Frank-Wolfe algorithm to minimise a novel criterion $\alpha$-Maximum Mean Discrepancy ($\alpha$-MMD), RDSS samples a representative and diverse subset for annotation from the unlabeled data. We demonstrate that minimizing $\alpha$-MMD enhances the generalization ability of low-budget learning. Experimental results show that RDSS consistently improves the performance of several popular SSL frameworks and outperforms the state-of-the-art sample selection approaches used in Active Learning (AL) and Semi-Supervised Active Learning (SSAL), even with constrained annotation budgets. Our code is available at [RDSS](https://github.com/YanhuiAILab/RDSS).

We present a multi-layer network propagation model including both offline and online social networks, which aims to imitate the interaction between the offline actual individuals and the online social networking ones throughout the propagation process. After that, we utilize the improved SIR model to simulate the transmission in social networks under the impact of actual individuals. In our model, the individual layer adopts the generalized linear threshold model, while the social network layer uses the improved SIR model proposed in this paper. We apply these two propagation models into a three-layer network with inter-layer links, and use Monte Carlo simulation to analyze the model dynamic properties. Furthermore, we make use of mean-field approximations to obtain the analytical solutions to the critical thresholds regarding the epidemic process. We discuss the effect of different factors on the model results, and observe whether the model has different dynamic performance from the classical SIR model as well. The present results will provide some deeper insights into how various contagion phenomena spread within many real-world networks.