An Improved Self-Organizing Map (SOM) Based on Virtual Winning Neurons

• Published in: Symmetry (Basel) Vol. 17; no. 3; p. 449
• Main Authors: Fan, Xiaoliang, Zhang, Shaodong, Xue, Xuefeng, Jiang, Rui, Fan, Shuwen, Kou, Hanliang
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 17.03.2025
• Subjects: Accuracy; Algorithms; Anomalies; Classification; Clustering; Complexity; Computational linguistics; Data processing; Datasets; Fault diagnosis; Feature extraction; Graph representations; Labeling; Language processing; Natural language interfaces; Neural networks; Neurons; Principal components analysis; Probability distribution; Real time; Scientific visualization; Self organizing maps; Support vector machines; Waveforms
• ISSN: 2073-8994; 2073-8994
• DOI: 10.3390/sym17030449

Self-Organizing Map (SOM) neural networks can project complex, high-dimensional data onto a two-dimensional plane for data visualization, enabling an intuitive understanding of the distribution and symmetric structures of such data, thereby facilitating the clustering and anomaly detection of complex high-dimensional data. However, this algorithm is sensitive to the initial weight matrix and suffers from insufficient feature extraction. To address these issues, this paper proposes an improved SOM based on virtual winning neurons (virtual-winner SOMs, vwSOMs). In this method, the principal component analysis (PCA) is utilized to generate the initial weight matrix, allowing the weights to better capture the main features of the data and thereby enhance clustering performance. Subsequently, when new input sample data are mapped to the output layer, multiple neurons with a high similarity in the weight matrix are selected to calculate a virtual winning neuron, which is then used to update the weight matrix to comprehensively represent the input data features within a minimal error range, thus improving the algorithm’s robustness. Multiple datasets were used to analyze the clustering performance of vwSOM. On the Iris dataset, the S is 0.5262, the F1 value is 0.93, the ACC value is 0.9412, and the VA is 0.0012, and the experimental result with the Wine dataset shows that the S is 0.5255, the F1 value is 0.93, the ACC value is 0.9401, and the VA is 0.0014. Finally, to further demonstrate the performance of the algorithm, we use the more complex Waveform dataset; the S is 0.5101, the F1 value is 0.88, the ACC value is 0.8931, and the VA is 0.0033. All the experimental results show that the proposed algorithm can significantly improve clustering accuracy and have better stability, and its algorithm complexity can meet the requirements for real-time data processing.