【摘要】 项集通常具备两个特点:(1)观测样本来自于不同的实体,这些实体间存在着相似关系;(2)样本量稀少,导致模式发掘不完整。本文考虑如何在这类数据上有效地发掘精简的模式集合。首先,通过定义一个扩散核函数,可将每个节点下的小样本扩展至图中的所有节点,并通过权重来标识他们与当前节点的相似度:继而这一权重值,又可以自然地引入到精简模式的搜索与评估过程中。这样我们不仅从理论上给出了图结构对MDL评估的影响,并且在实现上也相对简单,只需对现有算法添加一个预处理过程,并进行少量修改即可。实验表明,这一方案的挖掘效果,比通常的独立挖掘、全局挖掘方式均具备明显的优势。而且,由于只有一个额外的预处理过程,计算代价也较低。更多还原

【Abstract】 Itemset( or binary data tables) has the following two characteristics:( 1) observation samples come from different entities,and there are similarities between these entities;( 2) the sample size is scarce,resulting in pattern excavation is incomplete. This article further discusses how to discover patterns on structural itemsets. First,by defining a diffusion kernel function,we can extend a small sample under each node to all nodes in the graph and identify their similarity with the current node by weight; and this weight value can be naturally introduced into the pipeline of the search and evaluation process. In this way,we not only give theoretically the influence of the graph structure on the MDL evaluation,but also the relatively simple implementation,where only a small amount of modifications are needed based on the original MDL evaluation logic. Experiments show that the effect of this scheme has obvious advantages over the usual independent mining and global mining methods. Moreover,since there is only one additional pre-processing phase,the computational cost is also low.更多还原