Analysis and development of ensemble learning algorithms for information fusion

在这个项目中，我们开发和分析了集成学习算法。重点研究了基于ECOC框架的多类分类、有序标签分类和评级数据分析。对于基于ECOC的多类分类，我们解决了传统的基于Bradley-Terry模型的分类方法存在的问题，提出了一种新的方法，该方法比传统的方法具有更好的性能和更低的计算量。对于有序标签分类，我们提出了一种最大化AUC的Boost算法，并对其统计特性进行了分析。对于评级数据，我们使用混合模型对矩阵分解方法进行了扩展，并用大规模真实数据集验证了该方法的有效性。

In this program, we developed and analyzed ensemble learning algorithms. Especially, we focused on multi-class classification based on the framework of ECOC, ordered label classification and rating data analysis. For the multi-class classification based on ECOC, we tackled a problem of conventional Bradley-Terry model based method and proposed a novel method, which outperforms the conventional method with drastically lower computational cost. For the ordered label classification, we proposed a Boosting algorithm which can maximize AUC and analyzed its statistical properties. For the rating data, we extended the matrix factorization method using a mixture model and showed effectiveness of the proposed method using a large scale real dataset.