Author name cluster

Yi Xiong

Possible papers associated with this exact author name in Arrow. This page groups case-insensitive exact name matches and is not a full identity disambiguation profile.

6 papers

1 author row

AAAI Conference 2026 Conference Paper

PPFL: A Parameter Behavior-Driven Plug-in Personalization Engine for Federated Learning

Qianyue Cao
Zongwei Zhu
Zirui Lian
Rui Zhang
Boyu Li
Yi Xiong
Xuehai Zhou

Personalized Federated Learning (PFL) customizes models for each client to mitigate challenges from non-IID data, wherein a dominant strategy is model decoupling that partitions models into shared and personalized parts based on architectural priors (e.g., backbone vs. head). However, we reveal a critical flaw in this strategy: it induces "intrinsic drift," a performance degradation often more severe than the well-known client drift, which limits final accuracy. We trace this drift to a steep cliff of high loss emerging from the naive stitching of shared and personalized parts. To address this, we shift from architectural partitioning to a parameter behavior-driven paradigm. We introduce PPFL, an approach that employs a novel soft-fusion strategy guided by parameter-wise behavioral perception. PPFL dynamically infers each parameter's functional role—whether it behaves more like a 'personalist' or a 'generalist' in the current context—by synthesizing its multifaceted behavior observed during local training. Extensive experiments on image, text, and multimodal classification benchmarks show that PPFL outperforms eight state-of-the-art baselines by up to 5.3%. Moreover, it can function as a plug-in module, boosting the accuracy of vanilla FedAvg with a 16.82% absolute gain.

PDF Details DOI

JBHI Journal 2025 Journal Article

GRATCR: Epitope-Specific T Cell Receptor Sequence Generation With Data-Efficient Pre-Trained Models

Zhenghong Zhou
Junwei Chen
Shenggeng Lin
Liang Hong
Dong-Qing Wei
Yi Xiong

T cell receptors (TCRs) play a crucial role in numerous immunotherapies targeting tumor cells. However, their acquisition and optimization present significant challenges, involving laborious and time-consuming wet lab experimental resource. Deep generative models have demonstrated remarkable capabilities in functional protein sequence generation, offering a promising solution for enhancing the acquisition of specific TCR sequences. Here, we propose GRATCR, a framework incorporates two pre-trained modules through a novel “grafting” strategy, to de-novo generate TCR sequences targeting specific epitopes. Experimental results demonstrate that TCRs generated by GRATCR exhibit higher specificity toward desired epitopes and are more biologically functional compared with the state-of-the-art model, by using significantly fewer training data. Additionally, the generated sequences display novelty compared to natural sequences, and the interpretability evaluation further confirmed that the model is capable of capturing important binding patterns.

Details DOI

JBHI Journal 2025 Journal Article

Prediction of Drug-Target Interactions Based on Hypergraph Neural Networks With Multimodal Feature Fusion

Yufang Zhang
Jiayi Li
Shangqing Zhao
Hong Tan
Heqi Sun
Yi Xiong
Dong-Qing Wei

Accurate drug-target interaction prediction is vital for drug discovery and optimization. Traditional experimental methods, while effective, are time-intensive and costly. HyperGCN-DTI, a novel framework that explicitly advances beyond existing models such as CHL-DTI and HHDTI by leveraging hypergraph neural networks with a multimodal feature fusion strategy. While exsiting methods primarily focuses on low-order graph representations and fixed heterogeneous network structures, HyperGCN-DTI incorporates richer multimodal fused features including embeddings from pretrained language models and diverse biological networks and build robust hypergraphs that capture high-order multi-entity relationships within drug-target pairs. This dual-channel architecture effectively captures both local topological connections and higher-order structural dependencies. HyperGCN-DTI outperforms state-of-the-art DTI prediction models across multiple datasets and remains robust under imbalanced and large-scale real-world datasets, demonstrating its superior predictive power. The model demonstrates significant improvements when using multimodal features and hypergraph-based message passing, with sensitivity analysis confirming stability across hyperparameter variations. Top-ranked predictions are validated through biomedical literature and molecular docking, underscoring the reliability and practical relevance of our approach. HyperGCN-DTI is the first DTI prediction model to jointly integrate such a wide range of heterogeneous information sources with hypergraph representation, significantly enhancing accuracy and robustness, particularly in sparse or noisy settings. The proposed model offers a powerful and generalizable tool for accelerating drug development and target identification.

Details DOI

EAAI Journal 2024 Journal Article

Long-term deep reinforcement learning for real-time economic generation control of cloud energy storage systems with varying structures

Linfei Yin
Yi Xiong

Energy storage systems play a crucial role in modern power systems. Consequently, a mixed cloud energy storage (CES) system is proposed. The mixed CES system comprises consumers and prosumers. The consumers can only consume energy. The prosumers can either produce or consume energy at different time intervals. The proposed mixed CES system is designed for investigating the generation control challenges in mixed interconnected power systems. To optimize the active power balance and economic efficiency of the mixed CES system, a long-term deep reinforcement learning (LDRL) artificial intelligence approach is proposed as the real-time economic generation controller to control the mixed CES system. The LDRL consists of a reinforcement mechanism and two models: a long short-term memory model for economic dispatch and a deep neural networks model for smart generation control. The reinforcement framework updates policies for the prosumers. The efficacy of proposed method is validated across three mixed systems, i. e. , the improved IEEE 300-bus, Polish 2383-bus, and mixed systems with varying structures. The numerical simulations verify that the LDRL method can efficiently and economically control the mixed CES systems with diverse structures.

Details DOI

EAAI Journal 2024 Journal Article

More Quickly-RRT*: Improved Quick Rapidly-exploring Random Tree Star algorithm based on optimized sampling point with better initial solution and convergence rate

Xining Cui
Caiqi Wang
Yi Xiong
Ling Mei
Shiqian Wu

RRT* (Rapidly-exploring Random Tree Star), as a variant of RRT (Rapidly-exploring Random Tree), is widely used to solve path planning problems because of its asymptotic optimality. However, the algorithm is inefficient due to the high initial path cost and the slow convergence rate. In this paper, we propose a More Quickly-RRT* (MQ-RRT*) path planning algorithm based on optimized sampling points to solve the problems. A sparse sampling mechanism is proposed in MQ-RRT* to improve the global search efficiency by reducing repetitive sampling. To make the random tree oriented when expanding, a dynamic goal-biased strategy is proposed, which can reduce the sampling time. Like Q-RRT* (Quick-RRT*), MQ-RRT* expands the set of possible parent nodes in the ChooseParent and Rewire phases, which reduces the path cost. On this basis, a method for creating a new parent node close to the obstacle is proposed. The creation process can be divided into two steps: Remove-tips and CreateNodes, which further reduces the cost of path generation and makes the path smoother by using the triangle inequality principle. Finally, numerical simulations are used to compare the proposed algorithm with RRT*, Q-RRT*, GuILD (Guided Incremental Local Densification), and F-RRT* (Fast-RRT*), which verifies that the proposed algorithm has certain advantages in path cost, convergence rate, and path smoothness.

Details DOI

NeurIPS Conference 2021 Conference Paper

Regime Switching Bandits

Xiang Zhou
Yi Xiong
Ningyuan Chen
Xuefeng Gao

We study a multi-armed bandit problem where the rewards exhibit regime switching. Specifically, the distributions of the random rewards generated from all arms are modulated by a common underlying state modeled as a finite-state Markov chain. The agent does not observe the underlying state and has to learn the transition matrix and the reward distributions. We propose a learning algorithm for this problem, building on spectral method-of-moments estimations for hidden Markov models, belief error control in partially observable Markov decision processes and upper-confidence-bound methods for online learning. We also establish an upper bound $O(T^{2/3}\sqrt{\log T})$ for the proposed learning algorithm where $T$ is the learning horizon. Finally, we conduct proof-of-concept experiments to illustrate the performance of the learning algorithm.

PDF Details