基于上下文增强的多级注意力会话推荐模型

会话推荐的目标是仅根据用户在匿名会话中有限的交互行为,来预测用户的下一次点击行为.最近几年,许多基于图神经网络的会话推荐方法取得了可喜的结果.然而,这些方法仍然存在不足之处.一方面,基于图神经网络的方法只考虑物品之间的转换模式,忽略了会话中的序列模式.另一方面,现有的大多数方法都只关注当前会话内部的信息,忽略了来自邻居会话的外部协作信息,即上下文模式.为了解决上述问题,本文提出了一种新颖的基于上下文增强的多级注意力会话推荐模型(CEMA),通过多级注意力机制分别在物品级和会话级这两个粒度上学习物品特征和建模用户偏好,以增强模型的个性化推荐能力.CEMA模型利用多层GraphSAGE来学习物品之间复杂的转换模式,以捕获用户的局部偏好.特别地,在CEMA模型中设计了一种物品级注意力机制,通过门控注意力单元计算会话中不同物品的重要性,以识别用户真正感兴趣的物品,避免噪声物品的干扰.这有助于准确地捕获会话的序列模式,以建模用户的全局偏好.此外,所提出的方法还设计了一种会话级注意力机制,通过简单的软注意力高效地计算不同会话之间的相似度,以聚焦于那些与当前会话最相似的邻居会话,并从中提取上下文模式,以帮助预测用户的下一次点击.本文在3个公开的基准数据集上进行了一系列实验,实验结果表明CEMA的推荐性能超过了现有最好的方法,充分验证了CEMA的有效性和优越性.

基于模式链分析的文本页面图像的分割与分类

为了能对复杂版式的文本图像(如包含镶嵌在文字中的形状不规则的图片区)的页面进行图文分割与分类,提出了一种新的基于模式链分析的文本页面分割与分类算法。该算法首先使用外接矩形框出图像中的所有黑像素,并且存入矩形框链表中,再组合所有相邻的矩形进而形成模式,最后依据各模式的统计特征分类,输出文字区和图片区两类图像。另外,对大图片模式周围个别不确定的模式,本文采用了上下文分类的算法进行再次分类。实验结果表明,该算法不仅运算速度快,而且能够对复杂版式的页面图像进行正确的图文分割和分类。

浅谈微软Unity设计模式研究

Unity是微软P＆amp;P推出的一个开源的IOC框架。Unity框架是微软IOC思想的体现，IOC意味控制反转。即不创建对象，但是描述创建它们的方式。在代码中不直接与对象和服务连接，但在配置文件中描述哪一个组件需要哪一项服务。容器负责将这些联系在一起。其原理是基于OO设计原则的“别找我，我会来找你的”。也就是说，所有的组件都是被动的（Passive），所有的组件初始化和调用都由容器负责。组件处在一个容器当中，由容器负责管理。简单的来讲，就是由容器控制程序之间的关系，而非传统实现中，由程序代码直接操控。这也就是所谓“控制反转”的概念所在：控制权由应用代码中转到了外部容器，控制权的转移，这就是反转。

基于Web弱指导的本体概念实例及属性的同步提取

该文提出了一种基于Web弱指导的本体概念实例和属性的同步提取方法,利用小规模的种子实例和属性集,该文从Web上自动获取实例和属性共现的上下文模式,并利用种子实例和属性的关联性来评价这些模式。进一步,根据上下文模式提取候选概念实例和属性后,该文提出两种方法来评价提取的候选实例和属性。第一,利用概念实例和属性的关联性来互相评价对方的准确度;第二,利用候选实例或候选属性与种子实例或属性在上下文模式分布上的相似度来评价准确度。在疾病类实验结果表明,人工确认候选实例的准确率在前500个结果达到94%,前1 000个结果的准确率也高达93%。

基于并列结构的概念实例和属性的同步提取方法

在概念实例和属性的提取研究中,针对基于模式的方法召回率比较低的特点,该文提出了一种基于并列结构的概念实例和属性的同步提取方法。首先利用并列结构模式去网页集合中提取同类词语集合,然后再用基于种子的弱指导方法去学习实例和属性共现的上下文模式,最后再通过模式去提取候选实例或候选属性。在此过程中,每提取出一个候选,就将该候选所在的同类词语集合合并到候选集合中。实验结果表明,该文的方法在不降低准确率的基础上,能大大提高提取结果的召回率。

Implications for Cultural Landscape in a Chinese Context: Geo-analysis of Spatial Distribution of Historic Sites

The protection of historic sites, especially their relationship with urban development, has become a worldwide issue, both in developed and developing countries. In the context of rapid urbanization in China, the realistic compatibility between urban construction and the protection of historic sites is always a key research topic. In this study, first, to comprehend their spatial distribution patterns, 828 historic sites throughout the country are selected based on certain criteria. Then, we conduct quantitative research using GIS software, adopting indicators that include Nearest Neighbor Index, Gini Coefficient, and Geographic Concentration Index to analyze the spatial characteristics of historic sites on the three levels of city, province and nation. The results indicate that the spatial distribution of the different types of historic sites is an agglomeration on the nationwide scale, most of which is located in the regions of the Pearl(Zhujiang) River Delta, Yangtze(Changjiang) River Delta and Beijing-Tianjin Region. Because the majority of historic sites are located within approximately 10 km of the downtown area, a certain pattern has emerged, showing that the larger cities own more historic areas, which are in a more incomplete state of preservation, indicating the fragmentation of heritage spaces. The formation mechanism of the historic sites' distribution pattern is based on the conditions of the cities/towns as well as the bid-rent theory.

Context-aware Workflow Model for Supporting Composite Workflows

In recent years, several researchers have applied workflow technologies for service automation on ubiquitous compating environments. However, most context-aware workflows do not offer a method to compose several workflows in order to get mare large-scale or complicated workflow. They only provide a simple workflow model, not a composite workflow model. In this paper, the autorhs propose a context-aware workflow model to support composite workflows by expanding the patterns of the existing context-aware wrY：flows, which support the basic woddlow patterns. The suggested workflow model of. fers composite workflow patterns for a context-aware workflow, which consists of various flow patterns, such as simple, split, parallel flows, and subflow. With the suggested model, the model can easily reuse few of existing workflows to make a new workflow. As a result, it can save the development efforts and time of context-aware workflows and increase the workflow reusability. Therefore, the suggested model is expected to make it easy to develop applications related to context-aware workflow services on ubiquitous computing environments.

Context-Aware Service IVlodes in the Cloud Computing Environment

This paper discussed the differences of context-aware service between the cloud computing environment and the traditional service system.Given the above differences,the paper subsequently analyzed the changes of context-aware service during preparation,organization and delivery,as well as the resulting changes in service acceptance of consumers.Because of these changes,the context-aware service modes in the cloud computing environment change are intelligent,immersive,highly interactive,and real-time.According to active and responded service,and authorization and non-authorized service,the paper drew a case diagram of context-aware service in Unified Modeling Language(UML) and established four categories of context-aware service modes.

Interactive Peer Learning In An English Immersion Context In China

English immersion in the mainland of China has started in Xi＇an since the late 1990s, and extended into other cities in the mainland of China. This study reported the findings of the students＇ interactive peer learning model in the immersion programs, such as peer prompting and waiting; non-verbal expressions; correcting errors and modulating speaking volume; translation; attending to the peer interlocutor＇ s needs; and reciprocating peer assistance.