Huafeng Chen

Dr. Huafeng Chen


Research Focus & Metrics

Academic Career Length

8Years (2019-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
15 September 2019 in Concurrency and Computation: Practice and Experience
Verified
Journal Article
Closed access
PDF file
HCHuafeng Chen
Huafeng Chen
HLHongyang Li
Hongyang Li
ZXZengmin Xu
Zengmin Xu
YZYunhong Zhao
Yunhong Zhao
THTigang He
Tigang He

26 April 2024
Verified
Conference Paper
Open access
SZShun Zhang
Shun Zhang
WFWeiwei Fu
Weiwei Fu
YHYurong Hu
Yurong Hu
HCHuafeng Chen
Huafeng Chen
Question recognition is crucial in Q&A systems. An effective recognition method significantly enhances the accuracy and efficiency of information retrieval. Advances in natural language processing (NLP) technology have spurred the development of statement recognition techniques. However, many methods rely on extensive datasets, and numerous sectors struggle to keep pace with intelligent advancements due to data scarcity. Policy-related questions, which often involve a plethora of technical terms, exemplify this challenge. For colloquial questions with intricate structures, models require enhanced natural language understanding capabilities. These models must prioritize key information and filter out irrelevant data, particularly when data availability is limited. To address these issues, this study utilized medical insurance approval items and related questions as data sources. It employed sentence vectors and question structures as analytical frameworks and explored innovative solutions. Specifically, the twin network fine-tuning pre-training model RoBERTa was applied to classify medical insurance items, integrating syntactic structure and question word features to pinpoint question targets. This method outperformed standard text classification models in the medical insurance domain and has potential applicability in other policy-oriented Q&A systems, thereby offering precise services to the public.

1 January 2021 in Complexity
Verified
Journal Article
Open access
PDF file
HCHuafeng Chen
Huafeng Chen
MZMaosheng Zhang
Maosheng Zhang
ZGZhengming Gao
Zhengming Gao
YZYunhong Zhao
Yunhong Zhao
Current methods of chaos‐based action recognition in videos are limited to the artificial feature causing the low recognition accuracy. In this paper, we improve ChaosNet to the deep neural network and apply it to action recognition. First, we extend ChaosNet to deep ChaosNet for extracting action features. Then, we send the features to the low‐level LSTM encoder and high‐level LSTM encoder for obtaining low‐level coding output and high‐level coding results, respectively. The agent is a behavior recognizer for producing recognition results. The manager is a hidden layer, responsible for giving behavioral segmentation targets at the high level. Our experiments are executed on two standard action datasets: UCF101 and HMDB51. The experimental results show that the proposed algorithm outperforms the state of the art.

9 August 2022 in Journal of Testing and Evaluation
Verified
Journal Article
Closed access
LJLi Junmei
Li Junmei
HCHuafeng Chen
Huafeng Chen
SLSuruo Li
Suruo Li
Following an investigation undertaken by the publisher, we have determined that this paper was accepted on the basis of a compromised peer review process. We hereby retract the paper. The corresponding author has been notified of the retraction. The retraction statement can be found here: https://doi.org/10.1520/JTE20269996. Trusted network is characterized by a large amount of data, abnormal dispersion, and high complexity. Traditional methods are easily affected by trusted network environment, resulting in unreliable mining results. Therefore, a new real-time mining method of trusted network difference data is proposed. Real-time collection of trusted network difference data through history system is performed on the basis of determining the principle of trusted network difference data mining and collecting and extracting the characteristics of difference data. The process of trusted network differential data mining is designed through the artificial bee colony algorithm. According to the process, differential data mining is carried out from three aspects: constructing a trusted network differential data transmission path, updating pheromone, and establishing a differential data transmission path set. The experimental results show that the proposed method can effectively realize the real-time mining of difference data, and the mining accuracy is more accurate.

29 June 2023 in Applied Sciences
Verified
Journal Article
Open access
PDF file
ZXZengmin Xu
Zengmin Xu
XLXiangli Li
Xiangli Li
JLJiao-Fen Li
Jiao-Fen Li
HCHuafeng Chen
Huafeng Chen
RHRuimin Hu
Ruimin Hu
This study presents a new semi-supervised action recognition method via adaptive feature analysis. We assume that action videos can be regarded as data points in embedding manifold subspace, and their matching problem can be quantified through a specific Grassmannian kernel function while integrating feature correlation exploration and data similarity measurement into a joint framework. By maximizing the intra-class compactness based on labeled data, our algorithm can learn multiple features and leverage unlabeled data to enhance recognition. We introduce the Grassmannian kernels and the Projected Barzilai–Borwein (PBB) method to train a subspace projection matrix as a classifier. Experiment results show our method has outperformed the compared approaches when a few labeled training samples are available.

3 November 2026 in Preprints.org
Verified
Preprint
Open access
PDF file
WWWanli Wang
Wanli Wang
XSXiaoxue Su
Xiaoxue Su
XCXue Chen
Xue Chen
HCHuafeng Chen
Huafeng Chen
YZYangxi Zhang
Yangxi Zhang
Calibration remains a practical barrier for steady-state visual evoked potential (SSVEP) brain–computer interfaces. Task-related component analysis (TRCA) is accurate when several labelled trials per class are available, but its performance degrades when calibration is reduced to one trial and the test block follows a change in recording conditions. We asked whether a fixed modification of ensemble TRCA (eTRCA) could reduce this loss without any per-subject hyperparameter tuning. The proposed SST-eTRCA appends a three-sample delayed copy of each channel, z-scores observations before covariance accumulation, and pools 80 class filters into one projection. On the public 35-subject Tsinghua benchmark (40 classes, six posterior channels), it reached 95.6% with one calibration trial per class versus 93.4% for eTRCA (1400 decisions; McNemar p = 1.5E-5), and 95.0% versus 91.8% without extra calibration. The advantage was largest at 1 s (68.9% versus 63.9%). Under leave-one-subject-out, SST-eTRCA and eTRCA were indistinguishable (83.6% each), and self-training changed accuracy by only 0.1 points (p = 0.75). The gain is therefore specific to within-participant block shift, not to cross-subject transfer. These results define a narrow but useful operating range: a fixed, reproducible decoder for reducing recalibration across blocks, with the negative self-training result serving as a reliability check.