嵌入式多标签分类算法的优化研究

多标签分类中如何有效处理具有许多实例和大量标签的大规模数据集、补偿训练集中缺失标签以及利用未标记实例改进预测性能等问题已成为重要研究方向。提出嵌入式多标签分类（EMC）算法，首先从伪实例参数化的高斯过程（GP）中提取两组随机变换来模拟特征向量、潜在空间表示向量和标签向量之间的非线性关系映射，其次引入一组辅助变量结合专家集成（EEOE）方法补偿缺失标签，最后利用未标记实例学习随机函数的平滑映射提高预测性能。仿真结果表明，与特征识别隐式标签空间编码的多标签分类（FaLE）算法和半监督低秩映射多标签分类（SLRM）算法相比，EMC算法优化了处理大规模数据集、补偿缺失标签及利用未标记数据的能力，从而提高了类标签的预测性能，且具有良好的可扩展性，训练时间短。

Optimization of embedding-based multi-label classification

How to deal effectively with large-scale data sets with many instances and a large number of labels, compensate for missing labels in training sets, and improve prediction performance by using unlabeled instances in multi-label classification has become an important research direction. This paper proposes an embedding-based multi-label classification (EMC) algorithm. Firstly, two sets of random transformations were extracted from the pseudo-instance parameterized Gaussian process (GP) to model the nonlinear relationship mapping between feature vectors, latent space representation vectors and label vectors, and then a set of auxiliary variables combined with an expert ensemble with an overriding expert (EEOE) was introduced. The method compensates for the missing tags, and finally uses the unlabeled instance to learn the smooth mapping of the random function to improve the prediction performance. The simulation results show that compared with the FaLE and SLRM algorithms, the EMC algorithm optimizes the ability to process large data sets, compensate for missing tags, and utilize unlabeled data, thereby improving the predictive performance of class tags, with good scalability and short training time.