基于多图的交替优化图直推方法

Graph transduction via alternating minimization method based on multi-graph

针对基于单图的半监督学习(GSSL)算法的性能受单个图质量的影响,且在单视图数据下,大多数基于多图的GSSL算法难以使用的问题,提出了一种基于多图的交替优化图直推方法(MG-GTAM)。首先,使用不同的图构建参数来构建单视图数据下的多个图,利用多个图来表达数据间关系;然后,借助交替迭代方式综合多个图的信息,选择置信度高的未标记样本进行伪标记并通过权重权衡各图的重要程度,以优化多图上的预测函数的一致性和平滑性;最后通过组合每个图的预测函数完成对所有未标记样本的标记。仿真实验表明,与经典的局部和全局一致(LGC)、高斯随机场和调和函数(GFHF)、交替优化直推(GTAM)、组合图拉普拉斯(CGL)算法相比,在COIL20目标物体数据集和NEC Animal数据集上,MG-GTAM的分类错误率比这些经典算法均有下降,表明了该方法具有良好的性能。实验结果表明,MG-GTAM能有效地利用多个图来表达数据之间的关系,获得更低的分类错误率。

The performance of the Graph-based Semi-Supervised Learning （GSSL） method based on one graph mainly depends on a well-structured single graph and most algorithms based on multiple graphs are difficult to be applied while the data has only single view. Aiming at the issue, a Graph Transduction via Alternating Minimization method based on Multi- Graph （MG-GTAM） was proposed. Firstly, using different graph construction parameters, multiple graphs were constructed from data with one single view to represent data point relation. Secondly, the most confident unlabeled examples were chosen for pseudo label assignment through the integration of a plurality of map information and imposed higher weights to the most relevant graphs based on alternating optimization, which optimized agreement and smoothness of prediction function over multiple graphs. Finally, more accurate labels were given over the entire unlabeled examples by combining the predictions of all individual graphs. Compared with the classical algorithms of Local and Global Consistency （LGC）, Gaussian Fields and Harmonic Functions （GFHF）, Graph Transduction via Alternation Minimization （GTAM）, Combined Graph Laplacian （CGL）, the classification error rates of MG-GTAM decrease on data sets of COIL20 and NEC Animal. The experimental results show that the proposed method can efficiently represent data point relation with multiple graphs, and has lower classification error rate.