基于相似度衡量的决策树自适应迁移

Self-adaptive Transfer for Decision Trees Based on Similarity Metric

如何解决迁移学习中的负迁移问题并合理把握迁移的时机与方法,是影响迁移学习广泛应用的关键点.针对这个问题,提出一种基于相似度衡量机制的决策树自适应迁移方法 (Self-adaptive transfer for decision trees based on a similarity metric,STDT).首先,根据源任务数据集是否允许访问,自适应地采用成分预测概率或路径预测概率对决策树间的相似性进行判定,其亲和系数作为量化衡量关联任务相似程度的依据.然后,根据多源判定条件确定是否采用多源集成迁移,并将相似度归一化后依次分配给待迁移源决策树作为迁移权值.最后,对源决策树进行集成迁移以辅助目标任务实现决策.基于UCI机器学习库的仿真结果说明,与多源迁移加权求和算法(Weighted sum rule,WSR)和MS-TrAdaBoost相比,STDT能够在保证决策精度的前提下实现更为快速的迁移.

Negative transfer, transfer opportunity and transfer method are the most key problems affecting the learning perfor- mance of transfer learning. In order to solve these problems, a self-adaptive transfer for decision trees based on a similarity metric （STDT） is proposed. At first, according to whether the source task datasets to be allowed to access, a prediction prob- ability based on constituents or paths is adaptively used to cal- culate the affinity coefficient between decision trees, which can quantify the similarity degree of related tasks. Secondly, a judg- ment condition of multi-sources is used to determine whether the multi-source integrated transfer is adopted. If do, the sim- ilarity degrees are normalized, which can be viewed as transfer weights assigned to source decision trees to be transferred. At last, the source decision trees are transferred to assist the tar- get task in making decisions. Simulation results on UCI and text classification datasets illustrate that, compared with multi- source transfer algorithms, i.e., weighted sum rule （WSR） and MS-TrAdaBoost, the proposed STDT has a faster transfer speed with the assurance of high decision accuracy.