摘要
针对传统标记分布学习算法借助标记的全局相关性信息,忽略仅存于部分样本范围内标记局部相关性的问题,提出了一种基于样本稀疏表达的标记分布学习算法。借助样本点的自表达性质,建立稀疏表达优化模型,挖掘样本局部相关性信息。通过设计的标记分布目标函数约束,将得到的稀疏系数引入标记空间中,并将其作为隐含的标记空间局部相关性预测值,帮助标记分布模型的训练。使用交替方向乘子法求解样本稀疏系数,使用有限内存拟牛顿法求解标记分布目标函数,通过最大熵模型生成实例的标记分布预测值。在11个真实数据集上进行实验,并与7个现有标记分布学习算法进行对比。结果表明:所提算法在不同评价指标下的55次对比实验中取得了1.52的平均排名;面部表情数据集SBU-3DFE上,以相对熵衡量的表情判别准确度较标记分布学习问题转换算法PT-SVM、适应性算法AA-kNN及专用算法LDLLC的分别提高了3.10%、2.53%、2.48%;与传统标记分布学习算法相比,所提算法能够有效挖掘并利用标记局部相关性,具有良好的标记分布预测精度,且在不同类型的真实数据集上均能表现稳定。
The traditional label distribution learning algorithms utilize the label correlation in a global way but ignore the local correlation.An algorithm called label distribution learning via sample sparse representation(LDL-SR)is proposed.Particularly,to capture the sample local correlation,LDL-SR uses the self-expressiveness property of samples to establish a sparse representation optimization model.Then,by the well-designed objective function,the obtained sparse coefficient is introduced into the label space to predict the correlation of the labels to help the training of the label distribution model.The sparse representation optimization model can be solved with alternating direction multiplier method(ADMM),and limited memory quasi-Newton method(L-BFGS)is chosen to optimize the target function of the label distribution model.Finally,LDL-SR calculates the predicted label distribution via the maximum entropy model.Compared with 7 existing label distribution learning algorithms,the experimental results on 11 real-world datasets show that this algorithm achieves an average ranking of 1.52 in 55 comparative experiments under different evaluation measures.Compared with the PT-SVM,AA-kNN,and the LDLLC algorithms on facial expression dataset SBU-3 DFE,the accuracy of expression discrimination is enhanced by 3.10%,2.53%and 2.48%respectively.It is found that LDL-SR can achieve good prediction accuracy and stable performance on different real-world datasets.
作者
邵佳鑫
原盛
刘新媛
刘睿馨
SHAO Jiaxin;YUAN Sheng;LIU Xinyuan;LIU Ruixin(School of Software Engineering,Xi’an Jiaotong University,Xi’an 710049,China)
出处
《西安交通大学学报》
EI
CAS
CSCD
北大核心
2020年第11期139-148,共10页
Journal of Xi'an Jiaotong University
基金
国家自然科学基金资助项目(61902310)。
关键词
标记分布学习
稀疏表达
最大熵模型
label distribution learning
sparse representation
maximum entropy model
