基于最大熵的深度模糊聚类方法研究

Research on Deep Fuzzy Clustering Based on Maximum Entropy

摘要: 高维数据聚类是数据挖掘和模式识别研究领域的一项关键且具有挑战性的任务。深度聚类方法借助神经网络高效地特征提取能力，往往比传统聚类方法具有更好的性能。因此，本文提出了一种基于最大熵的深度模糊聚类算法(DFMEC)。该算法通过构建神经网络来表示模糊聚类，具有算法模型的可解释性。联合深度自动编码器模型，DFMEC通过梯度下降实现了深度特征学习和聚类中心的同步更新，解决了硬聚类由于其离散性而不能更新梯度的问题。此外，在所提出方法的目标函数的优化中，添加了基于模糊分配的最大熵正则项来提高聚类模型的鲁棒性。在各个高维数据集上的综合实验表明，与其他先进的深度聚类方法相比，该方法在重构表示和聚类质量方面都有更好的性能，大量实验的深入分析证明了这一点。

Abstract: Clustering of high-dimensional data is a crucial but challenging task in data mining and pattern recognition. The deep clustering usually have better performance than traditional methods due to more efficient feature extraction capability of neural networks. Therefore, this paper proposes a deep fuzzy clustering algorithm based on maximum entropy (DFMEC). The method represents fuzzy clustering by constructing neural networks and has the interpretability of the model. Jointly the deep autoencoder model, DFMEC through the network gradient realizes deep feature extraction and updating the clustering centers simultaneously, solving the problem that hard clustering cannot update gradients due to the discrete. Moreover, in the optimization of the objective function of the proposed method, a maximum entropy regularity term based on fuzzy assignment is added to improve the robustness of the clustering model. Comprehensive experiments on various high-dimensional datasets show that, in contrast to other advanced deep clustering methods, the method has better performance in terms of reconstruction representation and clustering quality, it is demonstrated by extensive experimental analyses.