Enhancing the Effectiveness of Interactive Case-Based Reasoning with Clustering and Decision Forests

In interactive case-based reasoning, it is important to present a small number of important cases and problem features to the user at one time. This goal is difficult to achieve when large case bases are commonplace in industrial practice. In this paper we present our solution to the problem by highlighting the interactive user- interface component of the CaseAdvisor system. In CaseAdvisor, decision forests are created in real time to help compress a large case base into several small ones. This is done by merging similar cases together through a clustering algorithm. An important side effect of this operation is that it allows up-to-date maintenance operations to be performed for case base management. During the retrieval process, an information-guided subsystem can then generate decision forests based on users' current answers obtained through an interactive process. Possible questions to the user are carefully analyzed through information theory. An important feature of the system is that case-base maintenance and reasoning are integrated in a seamless whole. In this article we present the system architecture, algorithms as well as empirical evaluations.     

一种大数据交互式挖掘框架与实现

在传统的数据挖掘过程中,用户需根据专业知识对数据进行预处理,为模型设定参数后构建模型,通过评估指标判断模型是否可行。该过程的不便性体现在模型以黑盒的方式构建,用户不可见其中间过程,模型产生的结果也不易被理解。在海量数据的环境下,传统数据挖掘过程在预处理时异常数据的定位和模型生成后知识的表达方面都有不便。为了解决传统数据挖掘过程存在的问题,本文提出了一种大数据环境下的交互式数据挖掘框架。该框架使交互贯穿整个数据挖掘的过程,使得用户可以轻松定位异常输入源数据,参与模型训练过程,对模型生成的结果溯源。本文还基于Spark对该框架进行了实现,并在食源性疾病爆发预测场景下验证了其可行性。     

一种基于空间映射及尺度变换的聚类框架

传统聚类算法通常建立在显式的模型之上,很少考虑泛化模型以适应不同的数据,由此导致了模型不匹配问题。针对此问题,该文提出了一种基于空间映射(Mapping)及尺度变换(Rescaling)的聚类框架(简称M-R框架)。具体而言,M-R框架首先将语料映射到一组具有良好区分度的方向所构建的坐标系中,以统计各个簇的分布特性,然后根据这些分布特性对各个坐标轴进行尺度变换,以归一化语料中各个类簇的分布。如上两步操作伴随算法迭代执行,直至算法收敛。该文将M-R框架应用到K-means算法及谱聚类算法上以验证其性能,在国际标准评测语料上的实验表明,应用了M-R框架的K-means及谱聚类在所有语料集上获得了全面的性能提升。     

改进的加权t-SNE算法及在故障诊断中的应用

对随机邻域嵌入算法(stochastic neighbor embedding, SNE)中的距离进行改进,提出一种基于Manhattan距离的加权t-SNE(Mwt-SNE)算法。使用受空间维数影响较小的Manhattan距离作为度量方式,使用k均值聚类算法将高维空间数据样本点距离分为三类,基于表格法进行权重参数寻优与加权,以加权相对Manhattan距离代替欧式绝对距离计算相似度条件概率,从而增大数据对象之间的区分度,提升降维效果,增强分类显著性。提出基于Mwt-SNE算法的在线故障诊断模型,使用核密度估计(KDE)确定控制限并进行在线监控。TE化工过程实验表明Mwt-SNE算法能有效降低误报率和漏报率,从而提高故障诊断稳定性和准确性。     

A Case-Based Framework for Interactive Capture and Reuse of Design Knowledge

Aerospace design is a complex task requiring access to large amounts of specialized information. Consequently, intelligent systems that support and amplify the abilities of human designers by capturing and presenting relevant information can profoundly affect the speed and reliability of design generation. This article describes research on supporting aerospace design by integrating a case-based design support framework with interactive tools for capturing expert design knowledge through concept mapping. In the integrated system, interactive concept mapping tools provide crucial functions for generating and examining design cases and navigating their hierarchical structure, while CBR techniques facilitate retrieval and aid interactive adaptation of designs. Our goal is both to provide a useful design aid and to develop general interactive techniques to facilitate case acquisition and adaptation. Experiments illuminate the performance of the system's context-sensitive retrieval during interactive case adaptation and the conditions under which it provides the most benefit.     

一种特征加权的聚类算法框架

为了考虑数据各维特征对聚类的不同贡献,并把有监督特征评价方法应用到无监督分类问题中,提出一种特征加权的聚类算法框架.该框架首先通过某种聚类算法对数据聚类,然后,根据聚类结果,采用有监督特征评价方法学习各维特征的权值,再根据特征权值重新聚类,之后再次学习特征权值,该过程反复迭代,直至算法收敛或达到指定的迭代次数.欧几里德空间内基于距离、基于密度的聚类算法均适用于本框架.基于本框架,采用模糊C均值聚类算法(FCM)、密度聚类算法(DBSCAN),并通过信息增益特征评价、ReliefF特征评价方法,对多个UCI数据集进行了实验,验证了该框架的有效性.     

t-SNE LDA算法在仿生嗅觉中的应用研究*

将流形学习中的t-SNE算法引入仿生嗅觉领域中,本研究提出一种基于t-SNE与LDA算法相结合的气味分类鉴定新方法。由PEN3电子鼻获取物质气味特征信息,通过t-SNE算法将非线性、高维度的气味响应数据降维到低维空间,并利用LDA算法对低维数据进行分类和识别。本研究利用五种不同成分的纺织品材料气味信息,通过t-SNE、PCA LDA和t-SNE LDA三种算法作对比实验。实验结果表明,相较于其他两种方法,t-SNE LDA算法对常见的不同成分纺织品材料拥有更好的分类和识别效果,而且t-SNE LDA算法得到的结果具有较小的类内离散度和较大的类间离散度。因此,t-SNE LDA算法是仿生嗅觉中气味分类和识别的一个新方法。     

基于WPT和t-SNE的直升机桨叶损伤特征提取

旋翼桨叶的损坏可能会导致直升机坠落损毁,开展桨叶健康状态的在线监测评估对保障飞行安全至关重要。提出一种将小波包变换(WPT)与t-分布随机近邻嵌入(t-SNE)相结合的桨叶损伤识别方法。首先利用振动台模拟直升机服役时的真实振动,用传感器获取不同故障桨叶模型在振动环境下的输出响应。然后对信号进行小波包分解,提取小波包能量作为原始特征向量,接着用流形学习对特征向量进行维数约简,最后输入到K近邻分类器进行故障识别。实验结果表明:首先,在原始特征选取方面,小波包能量特征优于时域特征与小波包能量组合成的混合特征;其次,t-SNE的降维效果优于PCA、Sammon映射、LTSA、HLLE、SNE这5种方法,且不受嵌入维数的制约。研究结果证明了所提出的方法能提高桨叶损伤评估的准确性。     

基于HLA框架的新一代分布交互仿真

本语文从ＨＬＡ产生的背景入手,介绍了ＨＬＡ的组成及特点,详细分析了ＬＡ核心ＲＴＩ的构成和葳功能,以及ＨＬＡ框架下的分布交互仿真的执行过程及特点。     

分簇结构谓词以及编译优化框架研究

谓词执行是在控制流存在的条件下可以有效挖掘指令级并行性的硬件机制。而在分簇结构上实现谓词机制,可以提高分簇结构上条件的执行效率。本文针对分簇结构展开谓词体系体系结构的研究,提出了分簇结构部分谓词的高效实现方法,以及基于循环展开的分簇结构部分谓词支持框架。实验表明,本文提出的分簇结构部分谓词及编译框架可以很好地提高条件执行程序的执行效率。     

面向签到日志的用户行为模式交互探索

签到日志记录了用户对于某类设施的使用情况,从中发现用户行为模式,在很多领域如精确广告投放、犯罪团伙发现等方面都具有非常广泛的应用价值.但是,发现过程却较为困难,主要因为：（1）日志数据体现为长时间序列且含有噪声,导致数据在高维空间分布较为稀疏,影响模式提取的准确性;（2）行为模式往往与不同的时间尺度相关;（3）多样的参数选择空间以及数据处理方式使得传统的机器学习方法很难获得可信且易于理解的行为分析结果.提出一种面向签到日志的用户行为模式交互探索的方法,该过程采用动态子空间策略,动态改变用于分析相似行为模式的时间片,从而减少人为设定参数对于分析结果的影响.方法集成了一个可视分析工具以支持该过程,利用该工具,分析人员可以实时了解方法每一步发现的模式,及时调整分析过程、直观理解和验证分析结论.包含了一个基于真实数据集的案例分析和一个来自不同领域专家的评审,其结果验证了方法的有效性.     

基于集成聚类的流量分类架构

流量分类是优化网络服务质量的基础与关键.机器学习算法利用数据流统计特征分类流量,对于识别加密私有协议流量具有重要意义.然而,特征偏置和类别不平衡是基于机器学习的流量分类研究所面临的两大挑战.特征偏置是指一些数据流统计特征在提高部分应用识别准确率的同时也降低了另外一部分应用识别的准确率.类别不平衡是指机器学习流量分类器对样本数较少的应用识别的准确率较低.为解决上述问题,提出了基于集成聚类的流量分类架构（traffic classification framework based on ensemble clustering,简称TCFEC）.TCFEC由多个基于不同特征子空间聚类的基分类器和一个最优决策部件构成,能够提高流量分类的准确率.具体而言,与传统的机器学习流量分类器相比,TCFEC的平均流准确率最高提升5%,字节准确率最高提升6%.     

Compiling CSPs into Tree-Driven Automata for Interactive Solving

Constraint programming techniques are widely used to model and solve decision problems and many algorithms have been developed to solve automatically and efficiently families of CSPs; nevertheless, they do not help solve interactive decision support problems, like product configuration. In such problems, the user chooses the values of the variables, and the role of the system is not to solve the CSP, but to help the user in this task. Dynamic global consistency maintaining is one of the most useful functionalities that should be offered by such a CSP platform. Unfortunately, this task is intractable in the worst case. Since interactivity requires short response times, intractability must be circumvented some way. To this end, compilation methods have been proposed that transform the original problem into a data structure allowing a short response time. In this paper, we extend the work of Amilhastre et al. [1] and Vempaty [15] by the use of a new structure, tree-driven automata, that takes advantage of the structural characteristics of configuration problems (decomposition of the components into independent subcomponents). Tree-driven automata can be far more compact than classical automata while keeping their good properties, especially a tractable complexity for the maintenance of global consistency.     

基于t-SNE的脑网络状态观测矩阵降维方法研究

针对基于功能核磁共振重构的脑网络状态观测矩阵维数过高和无特征的特点,对其降维方法展开研究,给出了基于t-SNE的脑网络状态观测矩阵降维算法,并且利用Python实现了降维及可视化平台。实验结果表明,与目前主流的其他降维算法相比较,使用该方法得到的脑网络状态观测矩阵低维空间的映射点有明显的聚类表现,并且在多个样本上的降维结果显现出一定的规律性,从而证明了该算法的有效性和普适性。     

Interactive Depth-of-Field Rendering with Secondary Rays

This paper presents an efficient method to trace secondary rays in depth-of-field (DOF) rendering, which significantly enhances realism. Till now, the effects by secondary rays have been little addressed in real-time/interactive DOF rendering, because secondary rays have less coherence than primary rays, making them very difficult to handle. We propose novel measures to cluster secondary rays, and take a virtual viewpoint to construct a layered image-based representation for the objects that would be intersected by a cluster of secondary rays respectively. Therefore, we can exploit coherence of secondary rays in the clusters to speed up tracing secondary rays in DOF rendering. Results show that we can interactively achieve DOF rendering effects with reflections or refractions on a commodity graphics card.     

一种两阶段的限制层次聚类算法

基于数据对象间的关联限制定义了类间关联系数,本文提出了两阶段的限制层次聚类算法TCCL.算法分为两个阶段,第一阶段主要依据数据对象的自然分布,基于数据对象间的距离把它们合并入一个个小类;在第二阶段,依据背景知识,基于类间关联系数来实现小类的进一步合并.一些实际数据集的实验结果表明,TCCL可以比较有效地利用所给关联限制来改善聚类效果.     

A Formal Framework for Interactive Agents

This paper proposes a formal framework and architecture for specification and analysis of interactive agents. The framework can be used to explore the design space, study features of different points in the design space, and to develop executable specifications of specific agents and study their interactions with the environment. A long term goal is development of reasoning principles specialized to different regions of the design space.     

A Framework for Interactive Hypertexture Modelling

Hypertexturing can be a powerful way of adding rich geometric details to surfaces at low memory cost by using a procedural three‐dimensional (3D) space distortion. However, this special kind of texturing technique still raises a major problem: the efficient control of the visual result. In this paper, we introduce a framework for interactive hypertexture modelling. This framework is based on two contributions. First, we propose a reformulation of the density modulation function. Our density modulation is based on the notion of shape transfer function. This function, which can be easily edited by users, allows us to control in an intuitive way the visual appearance of the geometric details resulting from the space distortion. Second, we propose to use a hybrid surface and volume‐point‐based representation in order to be able to dynamically hypertexture arbitrary objects at interactive frame rates. The rendering consists in a combined splat‐ and raycasting‐based direct volume rendering technique. The splats are used to model the volumetric object while raycasting allows us to add the details. An experimental study on users shows that our approach improves the design of hypertextures and yet preserves their procedural nature.     

ICEAGE: Interactive Clustering and Exploration of Large and High-Dimensional Geodata

The unprecedented large size and high dimensionality of existing geographic datasets make the complex patterns that potentially lurk in the data hard to find. Clustering is one of the most important techniques for geographic knowledge discovery. However, existing clustering methods have two severe drawbacks for this purpose. First, spatial clustering methods focus on the specific characteristics of distributions in 2- or 3-D space, while general-purpose high-dimensional clustering methods have limited power in recognizing spatial patterns that involve neighbors. Second, clustering methods in general are not geared toward allowing the human-computer interaction needed to effectively tease-out complex patterns. In the current paper, an approach is proposed to open up the black box of the clustering process for easy understanding, steering, focusing and interpretation, and thus to support an effective exploration of large and high dimensional geographic data. The proposed approach involves building a hierarchical spatial cluster structure within the high-dimensional feature space, and using this combined space for discovering multi-dimensional (combined spatial and non-spatial) patterns with efficient computational clustering methods and highly interactive visualization techniques. More specifically, this includes the integration of: (1) a hierarchical spatial clustering method to generate a 1-D spatial cluster ordering that preserves the hierarchical cluster structure, and (2) a density- and grid-based technique to effectively support the interactive identification of interesting subspaces and subsequent searching for clusters in each subspace. The implementation of the proposed approach is in a fully open and interactive manner supported by various visualization techniques.