Deep clustering algorithm with fusion selective clustering ensemble

With the advent of the 5G era, rapid advancements in computer technology, big data technologies and applications have become a hot topic. Nowadays, big data technologies are widely applied across various industries, particularly in business, healthcare, finance, and other fields. In the business sector, big data is utilized in sales forecasting, customer behavior analysis, and aiding enterprises in making more informed business decisions. In the healthcare sector, big data is employed for clinical case analysis, drug development, and assisting medical institutions in enhancing medical standards and efficiency. In the financial sector, big data is used for risk management, fraud detection, enabling financial institutions to better manage risks and enhance profitability. However, the enormous volume, complex structures, and lack of labeled data in today’s big data pose significant challenges. The intricate correlations among data further complicate the task of efficiently extracting meaningful and concise information.

Clustering analysis is a machine learning algorithm that discovers underlying knowledge patterns and structures in unlabeled data without any prior knowledge. Clustering is often used as an effective tool for understanding and analyzing raw data, particularly for problems that lack prior knowledge or where acquiring such knowledge is costly. Common clustering algorithms include k-means, Gaussian mixture models, spectral clustering, among others, which have shown decent performance on many datasets. However, when dealing with structurally complex big data, traditional clustering algorithms may yield suboptimal or even failed clustering results. This is because such data often exhibit higher complexity, sparse distributions, and originate from diverse data sources with varied formats and types, making them challenging to segment. On the other hand, most traditional clustering algorithms can only utilize shallow features of the data and fail to explore the interdependencies of complex data features in latent spaces.

The emergence and development of deep clustering algorithms provide a new approach to address these challenges. By combining unsupervised deep learning networks with powerful representation learning capabilities with clustering, deep clustering algorithms can be applied to extract low-dimensional feature representations conducive to clustering from various complex data types. Existing deep clustering methods can be broadly categorized into three types based on the unsupervised deep networks they employ: autoencoder-based, variational autoencoder-based, and generative adversarial network-based deep clustering. Each type of unsupervised deep network has its own limitations; for instance, the generative adversarial network-based deep clustering algorithm faces issues such as convergence difficulties and mode collapse. Furthermore, nearly all these deep clustering algorithms share a common challenge: they heavily rely on the low-dimensional representations extracted through unsupervised deep network training, making them susceptible to fluctuations in clustering results due to network parameter initialization and hyperparameters, resulting in poor robustness.

To address the challenges faced by deep clustering algorithms and achieve more stable and higher-quality clustering results when dealing with various types of complex data, a new deep clustering algorithm has been proposed. This algorithm integrates clustering ensembles with deep clustering, incorporating selective ensemble strategies to fuse selective clustering ensembles for improved performance.

The proposed deep clustering with fusion selective clustering ensemble algorithm could be divided into two stages: the deep base clustering generation stage and the selective clustering ensemble stage. In the deep base clustering generation stage, different initialized autoencoder unsupervised deep networks were used to train non-clustering losses and apply k-means to initialize cluster centroids. The similarity between the deep features extracted by the autoencoder and the cluster centers was calculated using the Student t-distribution as soft cluster assignments. The relative entropy loss was then used as the clustering loss for both soft cluster assignments and auxiliary distributions, where the auxiliary distribution was obtained by square-normalizing the soft cluster assignments. Finally, joint iterative optimization of the clustering loss was performed to obtain multiple deep clustering results. In the selective clustering ensemble stage, an evaluation of different deep base clustering results was conducted. Considering that both the quality and diversity of base clusters in the ensemble could impact performance, measures for evaluating the ensemble quality and diversity of base clusters are constructed. Base clusters with higher quality and richer diversity were selected as candidates for the clustering ensemble. When introducing the clustering ensemble strategy, the uncertainty of clusters was considered to evaluate the reliability of the clustering. Ultimately, a weighted graph partition consensus function was utilized to establish a weighted graph, and the Tcut algorithm was employed for graph partitioning to obtain the final clustering results.

A series of experiments was conducted on six datasets with eleven comparative clustering ensemble algorithms. The experimental results indicate that the deep clustering algorithm integrated with selective clustering ensemble not only improved robustness but also outperformed existing clustering ensemble algorithms in achieving better clustering results.

This paper proposes a deep clustering algorithm, which integrates selective clustering ensembles. By combining selective strategies and clustering ensemble techniques with deep clustering algorithms, this approach effectively addresses the challenges faced by existing deep clustering algorithms, enabling them to achieve more stable and higher-quality clustering results when dealing with various types of complex data. Comparative experiments on different types of data with various ensemble clustering algorithms demonstrate that the DCFSCE algorithm exhibits stronger robustness, better clustering performance, and practical utility. It can provide a foundational support for the advancement of clustering technology.