Arrow Research search

Author name cluster

Yiwen Cai

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.

2 papers
1 author row

Possible papers

2

EAAI Journal 2025 Journal Article

Feature selection-based rough graph classification via graph convolutional neural networks

  • Lin Sun
  • Yiwen Cai
  • Weiping Ding
  • Jiucheng Xu

In various applications, graph data is usually omnipresent. Many existing methods integrate graph classification and feature selection to select informative features; however, uncertainty in graph data is easily ignored and then the relationship between nodes has been not described fully. Additionally, when processing graph data, graph convolutional neural networks (GCNs) fall short in illustrating the structural features and inherent semantic of whole graphs, and relying on node features will lead to over-smoothing, which can affect the efficacy of graph classification. To address these flaws, this article presents a novel feature selection-based rough graph classification methodology via GCNs. Firstly, the average value of each node feature values is calculated to be sorted. A variance filtering-based feature selection scheme is constructed to remove irrelevant node features. Then, the upper approximate and lower approximate feature sets will be divided by a variance threshold. Secondly, according to the relationship between the upper and lower approximates, the upper approximate weight and lower approximate weight of nodes are provided, and the upper approximate weight and lower approximate weight of edges between nodes can be provided to remove the influence of redundant information on nodes. The weight coefficient between nodes is learned by calculating the similarity between nodes to create the upper and lower approximation graphs, and then rough graphs will be constructed to fully display the uncertainty between nodes. Thirdly, an updated self-attention pooling layer scheme is presented to evaluate the importance of nodes, GraphConv is used to extract features for an output feature matrix and obtain a sequence of important nodes. Feature matrix and adjacency matrix can be updated, the hierarchical pooling structure is studied to retain important nodes in network training, and then a simplified rough graph can be obtained. Thus, an end-to-end GCN model is designed by simultaneously considering the structure and node features of rough graphs. Finally, a novel rough graph classification algorithm via GCNs is designed. Experiments on 15 graph datasets illustrate the greater effectiveness of the designed methodology relative to other competitive algorithms.

Details DOI

NeurIPS Conference 2025 Conference Paper

FedFree: Breaking Knowledge-sharing Barriers through Layer-wise Alignment in Heterogeneous Federated Learning

  • Haizhou Du
  • Yiran Xiang
  • Yiwen Cai
  • Xiufeng Liu
  • Zonghan Wu
  • Huan Huo
  • Guodong Long

Heterogeneous Federated Learning (HtFL) enables collaborative learning across clients with diverse model architectures and non-IID data distributions, which are prevalent in real-world edge computing applications. Existing HtFL approaches typically employ proxy datasets to facilitate knowledge sharing or implement coarse-grained model-level knowledge transfer. However, such approaches not only elevate risks of user privacy leakage but also lead to the loss of fine-grained model-specific knowledge, ultimately creating barriers to effective knowledge sharing. To address these challenges, we propose FedFree, a novel data-free and model-free HtFL framework featuring two key innovations. First, FedFree introduces a reverse layer-wise knowledge transfer mechanism that aggregates heterogeneous client models into a global model solely using Gaussian-based pseudo data, eliminating reliance on proxy datasets. Second, it leverages Knowledge Gain Entropy (KGE) to guide targeted layer-wise knowledge alignment, ensuring that each client receives the most relevant global updates tailored to its specific architecture. We provide rigorous theoretical convergence guarantees for FedFree and conduct extensive experiments on CIFAR-10 and CIFAR-100. Results demonstrate that FedFree achieves substantial performance gains, with relative accuracy improving up to 46. 3% over state-of-the-art baselines. The framework consistently excels under highly heterogeneous model/data distributions and in large scale settings.

PDF Details
v2026.09.13