基于NR-Transformer的集群作业运行时间预测

高性能集群的作业调度通常使用作业调度系统来实现，准确填写作业运行时间能在很大程度上提升作业调度效率。现有的研究通常使用机器学习的预测方式，在预测精度和实用性上还存在一定的提升空间。为了进一步提高集群作业运行时间预测的准确率，考虑先对集群作业日志进行聚类，将作业类别信息添加到作业特征中，再使用基于注意力机制的NR-Transformer网络对作业日志数据建模和预测。在数据处理上，根据与预测目标的相关性、特征的完整性和数据的有效性，从历史日志数据集中筛选出7维特征，并按作业运行时间的长度将其划分为多个作业集，再对各作业集分别进行训练和预测。实验结果表明，相比于传统机器学习和BP神经网络，时序神经网络结构有更好的预测性能，其中NR-Transformer在各作业集上都有较好的性能。     

基于人脸卡口数据的行人共现关系图谱构建

随着人脸识别技术的不断进步以及人脸卡口的大范围且密集的部署, 本文针对团伙犯罪案件侦察这一应用场景, 对人脸卡口数据进行深入挖掘, 探究其中行人间的共现关系, 获取所关注的嫌疑人的现实社交网络, 锁定团伙其余人员. 经过实验比对和论证, 本文使用Chinese Whispers聚类算法对行人节点进行识别, 通过Faiss加速邻接边的构建, 加速图的初始化步骤, 解决其聚类效率低下的问题. 在此基础上, 使用共现频次和Apriori算法中的置信度挖掘行人间的共现关系, 构建行人共现关系图谱.     

针对细微特征进行K-means聚类的电台分选识别技术

现代战场中的无线通信设备日益增多，精准获取个体信息已成为研究热点，但也是难点。针对通信电台，提出了一种分选识别技术。该技术从电台物理层特性出发，对其辐射信号的细微特征进行K-means聚类以实现分选，分选的同时提取各个个体的特征属性值，未知信号通过与特征属性值相关运算实现个体识别。该技术无需先验知识，无需训练运算，通过实验验证，其可行、高效，易于工程实现。     

Various density light field image coding based on distortion minimization interpolation

In recent years, the light field (LF) as a new imaging modality has attracted wide interest. The large data volume of LF images poses great challenge to LF image coding, and the LF images captured by different devices show significant differences in angular domain. In this paper we propose a view prediction framework to handle LF image coding with various sampling density. All LF images are represented as view arrays. We first partition the views into reference view (RV) set and intermediate view (IV) set. The RVs are rearranged into a pseudo sequence and directly compressed by a video encoder. Other views are then predicted by the RVs. To exploit the four dimensional signal structure, we propose the linear approximation prior (LAP) to reveal the correlation among LF views and efficiently remove the LF data redundancy. Based on the LAP, a distortion minimization interpolation (DMI) method is used to predict IVs. To robustly handle the LF images with different sampling density, we propose an Iteratively Updating depth image based rendering (IU-DIBR) method to extend our DMI. Some auxiliary views are generated to cover the target region and then the DMI calculates reconstruction coefficients for the IVs. Different view partition patterns are also explored. Extensive experiments on different types LF images also valid the efficiency of the proposed method.     

A Tradeoff Between Accuracy and Speed for K-Means Seed Determination

With a sharp increase in the information volume, analyzing and retrieving this vast data volume is much more essential than ever. One of the main techniques that would be beneficial in this regard is called the Clustering method. Clustering aims to classify objects so that all objects within a cluster have similar features while other objects in different clusters are as distinct as possible. One of the most widely used clustering algorithms with the well and approved performance in different applications is the k-means algorithm. The main problem of the k-means algorithm is its performance which can be directly affected by the selection in the primary clusters. Lack of attention to this crucial issue has consequences such as creating empty clusters and decreasing the convergence time. Besides, the selection of appropriate initial seeds can reduce the cluster’s inconsistency. In this paper, we present a new method to determine the initial seeds of the k-mean algorithm to improve the accuracy and decrease the number of iterations of the algorithm. For this purpose, a new method is proposed considering the average distance between objects to determine the initial seeds. Our method attempts to provide a proper tradeoff between the accuracy and speed of the clustering algorithm. The experimental results showed that our proposed approach outperforms the Chithra with 1.7% and 2.1% in terms of clustering accuracy for Wine and Abalone detection data, respectively. Furthermore, achieved results indicate that comparing with the Reverse Nearest Neighbor (RNN) search approach, the proposed method has a higher convergence speed.     

Reconfigurable topology synthesis for application-specific NoC on partially dynamically reconfigurable systems

In this paper, a four-stage method for synthesizing reconfigurable ASNoC topology is proposed for partially dynamically reconfigurable systems, where the topology is reconfigured dynamically at run-time along with the application's execution. Firstly, a simulated annealing based topology-aware integrated optimization framework is proposed to generate the proper schedule and floorplan of task modules. Secondly, based on the schedule and floorplan of task modules, an Integer Linear Programming (ILP)-based method and a heuristic method, are proposed to partition the communication requirements of the application into $T$ time intervals. Thirdly, we explore the proper positions of switches in the floorplan for global communications. Finally, considering the reconfiguration costs between adjacent time intervals, the routing path allocation problem is solved for time intervals in an iterative procedure to generate fine-grained dynamically reconfigurable ASNoC topologies. Experimental results show that, compared to the random partition of communication requirements, the proposed heuristic method and ILP-based method can achieve 5.4% and 10.0% power consumption improvement, respectively. And, the reconfigurable ASNoC can achieve 31.6% power consumption improvement when compared with static ASNoC.     

基于Word2Vec模型特征扩展的Web搜索结果聚类性能的改进

对于用户泛化和模糊的查询，将Web搜索引擎返回的列表内容聚类处理，便于用户有效查找感兴趣的内容。由于返回的列表由称为片段（snippet）的短文本组成，而传统的单词频率-逆文档频率（TF-IDF）特征选择模型不能适用于稀疏的短文本，使得聚类性能下降。一个有效的方法就是通过一个外部的知识库对短文本进行扩展。受到基于神经网络词表示方法的启发，提出了通过词嵌入技术的Word2Vec模型对短文本扩展，即采用Word2Vec模型的TopN个最相似的单词用于对片段（snippet）的扩展，扩展文档使得TF-IDF模型特征选择得到聚类性能的提高。同时考虑到通用性单词造成的噪声引入，对扩展文档的TF-IDF矩阵进行了词频权重修正。实验在两个公开数据集ODP239和SearchSnippets上完成，将所提方法和纯snippet无扩展的方法、基于Wordnet的特征扩展方法和基于Wikipedia的特征扩展方法进行了对比。实验结果表明，所提方法在聚类性能方面优于对比方法。     

ATSP:时态网络可视化的自适应时间片划分方法

现有的时态网络可视化方法大多采用等量时间片来可视化网络的演变，不利于时态模式的快速挖掘和发现。为此，根据时态网络固有的特征提出自适应时间片划分方法（Adaptive Time Slice Partition method，ATSP）。在时态网络的两种表示方式（基于事件的表示方式和基于快照的表示方式）的基础上，构建了ATSP的基础模型，同时提出了一种改进模型用来描述事件间隔时间服从长尾分布的时态网络。为了实现时间片的不等量划分，针对探索任务的不同提出了基于时态模式的ATSP规则和基于中心节点的ATSP规则，并提出了实现算法--层次划分算法（Hierarchical Partition algorithm，HP）和增量划分算法（Incremental Partition algorithm，IP）。实验结果表明，ATSP方法比传统的时间片划分方法更能准确地表示网络的时态特征，且该方法应用于可视化时，能有效归纳并展示网络的特征，明显提高了视觉分析的效率。