The co-training algorithm addresses the limitations of single-view learning by utilizing multi-view features to train two classifiers that complement each other and enhance learning performance. However, the lack of labeled samples and the problem of unlabeled samples not being effectively utilized restrict the further improvement of the performance of the co-training algorithm. In order to solve the above problems, a co-training algorithm combining local density sample synthesis and inter-class interpolation was presented. The algorithm first utilizes the local density sample synthesis method to expand the labeled sample set to improve the spatial structure of the data and obtain two initial classifiers with better performance. Then, the K-means clustering algorithm is used to cluster the unlabeled samples with inconsistent categories predicted by the two classifiers, and the samples of different categories are randomly selected to interpolate to generate new samples. These new samples are then added to each classifier, which pushes the decision boundary further from the class boundary to achieve a larger margin. This process is repeated until the final classifiers are obtained. Experimental results on twelve UCI datasets verify the effectiveness of the proposed algorithm.
面对工业样本中有标记样本量不足的情形,Zhu等[15]提出了一种可生成虚拟样本的协同训练方法以增加有标记样本数,该方法利用稀疏指数识别特征空间的稀疏程度,并在该空间上生成虚拟样本,再利用双K近邻(K-Nearest Neighbor, KNN)模型赋予标签。该方法虽能合成新样本,但在数据集不平衡时,双KNN模型对生成的样本赋予的标签会存在较大误差。考虑到不平衡数据集的样本扩充和数据本身的空间结构信息,本文提出一种结合局部密度的样本合成(Sample Synthesis with Local Density, SSLD)方法,用其合成新样本扩充有标记样本集合,解决上述问题1)。
综上所述,本文提出一种结合局部密度样本合成与类间插值的协同训练(Co-Training based on Sample Synthesis with Local Density and Inter-class Interpolation, CTSSLDII)算法。首先该算法对有标记样本进行扩充后训练两个初始分类器;其次两个分类器对无标记样本进行预测,保留预测不一致样本;最后对预测不一致样本进行K-means聚类,选取不同类别中的样本进行插值,用插值产生的样本推进决策边界的移动。本文主要贡献有:
④ 文献[10]提出的CTALDPC算法(Co-Training algorithm combined with Active Learning and Density Peak Clustering,CTALDPC);
⑤ 文献[12]提出的基于熵和多准则的协同训练方法(Co-Training method based on Entropy and Multi-Criteria,CTEMC);
⑥ 文献[13]提出的结合改进密度峰值聚类和共享子空间的协同训练算法(Co-Training algorithm combining Improved Density Peak Clustering and Shared Subspace,CTIDPCSS);
⑦ 文献[8]提出的基于预测未标记样本邻居置信度的协同训练算法(Co-Training algorithm by predicting Confidence of Unlabeled Neighbors,CTPCUN);
⑧ 文献[14]提出的一种无参数单步无标记数据选择策略的自然邻居协同训练方法(Co‑Training method based on parameter‑free and single‑step unlabeled data selection strategy with Natural Neighbors,CTNaN)。
VAN ENGELENJ E, HOOSH H. A survey on semi-supervised learning[J]. Machine Learning, 2020, 109(2): 373-440. DOI: 10.1007/s10994-019-05855-6 .
[2]
NINGX, WANGX R, XUS H, et al. A review of research on co-training[J]. Concurrency and Computation: Practice and Experience, 2021: e6276. DOI: 10.1002/cpe.6276 .
QIANT Y, LIUB, CHENL, et al. Tri-Training for authorship attribution with limited training data: A comprehensive study[J]. Neurocomputing, 2016, 171(1): 798-806. DOI: 10.1016/j.neucom.2015.07.064 .
[5]
NGUYENC M, LIX, BLANTONR D, et al. Partial Bayesian co-training for virtual metrology[J]. IEEE Transactions on Industrial Informatics, 2020, 16(5): 2937-2945. DOI: 10.1109/TII.2019.2903718 .
[6]
NANF T, TANGY H, YANGP, et al. A novel sub-Kmeans based on co-training approach by transforming single-view into multi-view[J]. Future Generation Computer Systems, 2021, 125: 831-843. DOI: 10.1016/j.future.2021.07.019 .
[7]
WANGR Y, WANGZ Z, YANGX T, et al. Co-training semi-supervised learning algorithm for classification of seismic facies of carbonate tidal channels[C]//International Field Exploration and Development Conference. Singapore: Springer, 2022: 1417-1434. DOI: 10.1007/978-981-19-2149-0_132 .
[8]
SONGX L, LIUH A, FENGY L, et al. A co-training algorithm by predicting confidence of unlabeled neighbors[J]. International Journal on Artificial Intelligence Tools, 2022, 31(6): 2240023. DOI: 10.1142/s0218213022400231 .
[9]
WUD, SHANGM S, LUOX, et al. Self-training semi-supervised classification based on density peaks of data[J]. Neurocomputing, 2018, 275(5): 180-191. DOI: 10.1016/j.neucom.2017.05.072 .
GONGY L, LÜJ. Co-training algorithm with combination of active learning and density peak clustering[J]. Journal of Computer Applications, 2019, 39(8): 2297-2301. DOI: 10.11772/j.issn.1001-9081.2019010075(Ch ).
[12]
VALEK M O, GORGÔNIOF L, ARAÚJOY N, et al. A co-training-based algorithm using confidence values to select instances[C]//2020 International Joint Conference on Neural Networks (IJCNN). New York: IEEE Press, 2020: 1-7. DOI: 10.1109/IJCNN48605.2020.9206621 .
[13]
LUJ, GONGY L. A co-training method based on entropy and multi-criteria[J]. Applied Intelligence, 2021, 51(6): 3212-3225. DOI: 10.1007/s10489-020-02014-6 .
LÜJ, XIANY. Co-training algorithm combining improved density peak clustering and shared subspace[J]. Journal of Computer Applications, 2021, 41(3): 686-693. DOI: 10.11772/j.issn.1001-9081.2020071095(Ch ).
[16]
GONGY L, WUQ W, CHENGD D. A co-training method based on parameter-free and single-step unlabeled data selection strategy with natural neighbors[J]. International Journal of Machine Learning and Cybernetics, 2023, 14(8): 2887-2902. DOI: 10.1007/s13042-023-01805-w .
[17]
ZHUQ X, ZHANGH T, TIANY, et al. Co-training based virtual sample generation for solving the small sample size problem in process industry[J]. ISA Transactions, 2023, 134(3): 290-301. DOI: 10.1016/j.isatra.2022.08.021 .
[18]
VERMAV, KAWAGUCHIK, LAMBA, et al. Interpolation consistency training for semi-supervised learning[J]. Neural Networks, 2022, 145(1): 90-106. DOI: 10.1016/j.neunet.2021.10.008 .
[19]
ZHANGC Y, WANGJ M, LIX Y, et al. Clustering centroid selection using a K-means and rapid density peak search fusion algorithm[C]//2020 IEEE 11th International Conference on Software Engineering and Service Science (ICSESS). New York: IEEE Press, 2020: 201-207. DOI: 10.1109/ICSESS49938.2020.9237746 .
[20]
AHMADM F, ISAN A M, LIMW H, et al. Differential evolution: A recent review based on state-of-the-art works[J]. Alexandria Engineering Journal, 2022, 61(5): 3831-3872. DOI: 10.1016/j.aej.2021.09.013 .