Compositional data which contain relative or structure information of a whole occur commonly in many disciplines and practical scenarios. Yet relatively few works are available for multivariate compositional data classification with different numbers of parts using machine learning. This is because compositional data is inherently constrained to unit sum, resulting in the existing methods cannot be directly applied. Particularly, the multivariate analysis methods for compositional data variables with unequal sizes of parts are not sufficiently investigated. Moreover, to design a good classification model is indeed a complicated work. Except for the learning algorithm, data quality is also an essential determinant, which is rarely been concerned. In this paper, we propose an effective framework for multivariate compositional data classification. Specifically, the Dirichlet feature embedding is proposed to implement on the original compositional data features with the goal of removing the constraint and obtaining high quality training data, as well as reducing the dimension. Support vector machine is then used to build the classification model. Results of simulation study and real-world dataset show our proposed method can achieve good performances.

包含整体的相对或结构信息的成分数据通常出现在许多学科和实际情况中。然而，相对较少的作品可用于使用机器学习对零件数量不同的多元成分数据进行分类。这是因为成分数据固有地限于单位和，导致现有方法无法直接应用。特别是，对于零件尺寸不相等的成分数据变量的多元分析方法尚未得到充分研究。而且，设计一个好的分类模型确实是一项复杂的工作。除学习算法外，数据质量也是至关重要的决定因素，这一点很少受到关注。在本文中，我们提出了一个有效的多元组成数据分类框架。具体地，提出了狄利克雷特征嵌入以在原始组成数据特征上实现，以消除约束并获得高质量的训练数据，以及减小尺寸。然后，使用支持向量机构建分类模型。仿真研究和实际数据集的结果表明，本文提出的方法可以取得良好的性能。