基于深度自编码的多视图子空间聚类网络

传统子空间浅层聚类模型对于多视图和非线性数据的聚类性能不佳。为此,提出一种基于深度自编码器的多视图子空间聚类网络模型,通过在深度自编码器中引入子空间聚类中的“自我表示”特性以及加权稀疏表示,提升了多视图子空间聚类算法的学习能力。推导的深度自编码多视图子空间聚类算法能够聚类具有复杂结构的数据点。通过多视图数据集验证了提出算法的有效性。结果表明,该方法能够有效地挖掘数据固有的多样性聚类结构,并利用多个视图之间互补信息,在性能上与现有方法相比有较大的提升。     

基于超图和样本自表征的谱聚类算法*

针对传统谱聚类算法仅考虑数据点对点间的相互关系而未考虑数据间可能隐藏的复杂的相关性的问题,提出一种基于超图和自表征的谱聚类方法。首先,建立数据的超图,得到超图的拉普拉斯矩阵表示;然后,利用L2,1-范数对样本进行行稀疏自表征,同时融入超图来描述数据间多层次的相互关系;最后,利用生成的自表征系数进行谱聚类。利用基于超图的样本自表征技术考虑了样本之间复杂的相关性。通过在Hopkins155等数据集上的实验表明,在聚类错误率评判标准下,算法优于现有基于普通图的谱聚类算法SSC、SRC等。     

基于Schatten-p范数和特征自表示的无监督特征选择

特征选择是去除不相关和冗余特征,找到具有良好泛化能力的原始特征的紧凑表示,同时,数据中含有的噪声和离群点会使学习获得的系数矩阵的秩变大,使得算法无法捕捉到高维数据中真实的低秩结构。因此,利用Schatten-p范数逼近秩最小化问题和特征自表示重构无监督特征选择问题中的系数矩阵,建立一个基于Schatten-p范数和特征自表示的无监督特征选择（SPSR）算法,并使用增广拉格朗日乘子法和交替方向法乘子法框架进行求解。最后在6个公开数据集上与经典无监督特征选择算法进行实验比较,SPSR算法的聚类精度更高,可以有效地识别代表性特征子集。     

Self-concept differentiation across the adult life span.

This study used 2 measures to examine 158 adults' (80 men, 78 women; ages 20 to 88 years) self-concept differentiation (SCD) across 5 role-specific self-representations. Findings revealed that the 2 measures did not assess SCD in similar ways and that they showed different associations with age. Specifically, the 1st measure was not significantly related to age, whereas the 2nd measure showed a curvilinear, U-shaped association with age. The 2nd SCD index also showed significant associations with several measures of emotional adjustment and 6 dimensions of psychological well-being. Additional analyses showed that age moderated the associations between SCD and positive and negative psychological well-being. A high level of SCD was associated with lower positive and higher negative psychological well-being for both young and older adults. However, this effect was significantly more pronounced in older adults. (PsycINFO Database Record (c) 2010 APA, all rights reserved)     

Supporting children's online identity in international communities

Increasingly, adults and children socialise and communicate online. Children's safe and secure online communication with people from all over the world can increase their understanding of other cultures, which is an important goal in today's multicultural world. Our research studied such interactions between children from three countries, Hungary, Mexico, and the USA. The goals of our research were to study how children represent their identity online and what the implications are for the design of children's online communities. We used qualitative methods to derive a deep understanding of children's behaviours and motivations. Our results show that children exposed their true and complete identities online. They focused on sharing and learning about personal, ethnic, and gender identities via online media and largely ignored cultural identity. They learned about children from other countries and developed positive attitudes towards them. Based on our results we describe design guidelines for children's online identity tools.     

Self-representation and brain development.

This study examined the relation between self-representation and brain development in infants and young children. Self-representation was assessed by mirror recognition, personal pronoun use, and pretend play. Structural brain images were obtained from magnetic resonance imaging (MRI). Brain development was assessed by a quantitative measure of maturation of the temporo-parietal junction, temporal pole, medial frontal cortex, and occipital cortex. Fifteen children (15 to 30 months of age; 3 girls) without MRI abnormalities were assessed. Only maturation of the left temporo-parietal junction was related to self-representation after controlling for age. These findings provide some evidence that brain maturation is related to the emergence of a representation of self in the human child. (PsycINFO Database Record (c) 2010 APA, all rights reserved)     

Semiotic perspectives on interactive languages for life on the screen

Cross-disciplinary research involving semiotics and computer science is rare. With the Web 2.0, contemporary activities of users can be properly described as real ‘life on the screen’. One of the challenges for the design of interactive languages is to support these activities and to express the much wider variety of meanings that users want to exchange through and with software. As the discipline whose aim is to investigate meanings, through representation and interpretation processes, semiotics is remarkably well-positioned to contribute with new knowledge in our field. This viewpoint article examines the reasons why in spite of this positioning, semiotics remains unpopular among researchers interested in interactive computer languages. In particular, it proposes that a semiotic approach can help us think about computer languages to represent our individual and collective ‘selves’ on the screen.     

基于图学习正则判别非负矩阵分解的人脸识别

基于图正则非负矩阵分解（NMF）算法充分利用了高维数据通常位于一个低维流形空间的假设从而构造拉普拉斯矩阵,但该算法的缺点是构造出的拉普拉斯矩阵是提前计算得到的,并没有在乘性更新过程中对它进行迭代。为了解决这个问题,结合子空间学习中的自表示方法生成表示系数,并进一步计算相似性矩阵从而得到拉普拉斯矩阵,而且在更新过程中对拉普拉斯矩阵进行迭代。另外,利用训练集的标签信息构造类别指示矩阵,并引入两个不同的正则项分别对该类别指示矩阵进行重构。该算法被称为图学习正则判别非负矩阵分解（GLDNMF）,并给出了相应的乘性更新规则和目标函数的收敛性证明。在两个标准数据集上的人脸识别实验结果显示,和现有典型算法相比,所提算法的人脸识别的准确率提升了1% ~ 5%,验证了其有效性。     

基于特征选择与鲁棒图学习的多视图聚类

现有的多视图聚类方法大多直接在原始数据样本上构建各视图的相似图,而原始数据中的冗余特征和噪声会导致聚类精度下降。针对该问题,基于特征选择和鲁棒图学习提出多视图聚类算法FRMC。在自适应选择不同视图特征时降低数据维度,减少冗余特征,同时利用自表示学习获取数据的表示系数,滤除噪声影响并得到数据样本的全局结构,从而去除样本中的噪声和离群点。在此基础上,通过自适应近邻学习构造样本鲁棒图,利用鲁棒图矩阵的加权和构建最终的亲和图矩阵,提出一种基于增广拉格朗日乘子的交替迭代算法对目标函数进行优化。在6个不同类型的标准数据集上进行实验,与SC、RGC、AWP等算法的对比结果表明,FRMC算法能够有效提升聚类精度且具有较好的收敛性与鲁棒性。     

特征自表达和图正则化的鲁棒无监督特征选择

为了在揭示数据全局结构的同时保留其局部结构,本文将特征自表达和图正则化统一到同一框架中,给出了一种新的无监督特征选择(unsupervised feature selection,UFS)模型与方法。模型使用特征自表达,用其余特征线性表示每一个特征,以保持特征的局部结构;用基于 ${L_{2, 1}}$ 范数的图正则化项,在保留数据的局部几何结构的同时可以降低噪声数据对特征选择的影响;除此之外,在权重矩阵上施加了低秩约束,保留数据的全局结构。在6个不同的公开数据集上的实验表明,所给算法明显优于其他5个对比算法,表明了所提出的UFS框架的有效性。