结合图模型和分层查询的半监督学习方法

Semi-supervised learning method combined graph model and hierarchic query

针对大规模检索时半监督学习方法难以兼顾运算复杂度和学习精度的问题,提出一种高效、高精度的半监督学习方法,在不降低学习精度的前提下降低运算复杂度。设计用于图模型重建的复杂度函数,在给定的复杂度水平下,结合径向基函数重建最优的图模型;采用轮换授权算法,对图模型进行分层聚类,构建树结构;采用分层查询策略,沿着树自上而下地遍历各节点,贪婪搜索期望误差最小的节点,扩展已标记样本集,不断降低学习误差,提高学习精度。仿真实验表明,与现有的半监督学习方法相比,该方法的学习精度更高、运算复杂度更低。

For solving the problem that traditional semi-supervised learning method does not give consideration to both computa-tional complexity and accuracy for large-scale query, a semi-supervised learning method with high efficiency and accuracy was proposed, which reduced the computational complexity on the premise of maintaining the learning accuracy. A complexity func-tion for reconstructing graph model was designed, and the optimal graph model based on radial basis function was reconstructed under a given complexity level. Authority-shift algorithm for hierarchical clustering of the graph model was executed, and a tree structure was built. Hierarchical query approach was used to greedily search the node with minimum expected error through traversing each node down the tree from top to bottom, the dataset with labeled samples was extended, and learning error was reduced and learning accuracy was improved. Experimental results show that, compared with the existing semi-supervised lear-ning methods, the proposed method has higher accuracy and lower computation complexity.