稀疏标签传播:一种鲁棒的领域适应学习方法

Sparse Label Propagation: A Robust Domain Adaptation Learning Method

稀疏表示因其所具有的鲁棒性,在模式分类领域逐渐得到关注.研究了一种基于稀疏保留模型的新颖领域适应学习方法,并提出一种鲁棒的稀疏标签传播领域适应学习(sparse label propagation domain adaptation learning,简称SLPDAL)算法.SLPDAL通过将目标领域数据进行稀疏重构,以实现源领域数据标签向目标领域平滑传播.具体来讲,SLPDAL算法分为3步:首先,基于领域间数据分布均值差最小化准则寻求一个优化的核空间,并将领域数据嵌入到该核空间;然后,在该嵌入核空间,基于l1-范最小化准则计算各领域数据的核稀疏重构系数;最后,通过保留领域数据间核稀疏重构系数约束,实现源领域数据标签向目标领域的传播.最后,将SLPDAL算法推广到多核学习框架,提出一个SLPDAL多核学习模型.在鲁棒人脸识别、视频概念检测和文本分类等领域适应学习任务上进行比较实验,所提出的方法取得了优于或可比较的学习性能.

Sparse representation has received an increasing amount of interest in pattern classification due to its robustness. In this paper a domain adaptation learning （DAL） approach is explored based on a sparsity preserving model, which assumes that each data point can be sparsely reconstructed. The proposed robust DAL algorithm, called sparse label propagation domain adaptation learning （SLPDAL）, propagates the labels from labeled points in the source domain to the unlabeled dataset in the target domain using those sparsely reconstructed objects with sufficient smoothness. SLPDAL consists of three steps. First, it finds an optimal kernel space in which all samples from both source and target domains can be embedded by minimizing the mean discrepancy between these two domains. Then, it computes the best kernel sparse reconstructed coefficients for each data point in the kernel space by using ll-norm minimization. Finally, it propagates the labels of source domain to the target domain by preserving the kernel sparse reconstructed coefficients. The paper also derives an easy way to extend SLPDAL to out-of-sample data and multiple kernel learning respectively. Promising experimental results have been obtained for several DAL problems such as face recognition, visual video detection and text classification tasks.