Graph analysis can be done at scale by using Spark GraphX which loading data into memory and running graph analysis in parallel.In this way,we should take data out of graph databases and put it into memory.Considering...Graph analysis can be done at scale by using Spark GraphX which loading data into memory and running graph analysis in parallel.In this way,we should take data out of graph databases and put it into memory.Considering the limitation of memory size,the premise of accelerating graph analytical process reduces the graph data to a suitable size without too much loss of similarity to the original graph.This paper presents our method of data cleaning on the software graph.We use SEQUITUR data compression algorithm to find out hot code path and store it as a whole paths directed acyclic graph.Hot code path is inherent regularity of a program.About 10 to 200 hot code path account for 40%-99%of a program’s execution cost.These hot paths are acyclic contribute more than 0.1%-1.0%of some execution metric.We expand hot code path to a suitable size which is good for runtime and keeps similarity to the original graph.展开更多
为克服依靠图像数据进行识别的局限,使用航迹数据和深度学习方法是当前船型识别的热门方案。针对船型识别任务中常用的长短时记忆(Long Short Term Memory,LSTM)网络对航迹数据特征提取的性能饱和问题,提出了一种基于“Four-hot”编码和...为克服依靠图像数据进行识别的局限,使用航迹数据和深度学习方法是当前船型识别的热门方案。针对船型识别任务中常用的长短时记忆(Long Short Term Memory,LSTM)网络对航迹数据特征提取的性能饱和问题,提出了一种基于“Four-hot”编码和Transformer-LSTM神经网络模型的船型识别方法。首先将航迹数据编码为“Four-hot”向量形式;然后构建由Transformer编码模块和LSTM网络级联的Transformer-LSTM神经网络模型,用LSTM网络对Transformer输出的高层表示进行二次学习。在船舶自动识别系统(Automatic Identification System,AIS)数据集上的实验结果显示,所提出的方法在测试集上的加权平均F1分数(Weighted Average F1,WAF1)比未编码航迹数据经5类实验模型直接训练和测试得到的值高3.09百分点以上。展开更多
基金This research work is supported by Hunan Provincial Education Science 13th Five-Year Plan(Grant No.XJK016BXX001)Social Science Foundation of Hunan Province(Grant No.17YBA049)+2 种基金Hunan Provincial Natural Science Foundation of China(Grant No.2017JJ2016)The work is also supported by Open foundation for University Innovation Platform from Hunan Province,China(Grand No.16K013)the 2011 Collaborative Innovation Center of Big Data for Financial and Economical Asset Development and Utility in Universities of Hunan Province.National Students Platform for Innovation and Entrepreneurship Training(Grand No.201811532010).
文摘Graph analysis can be done at scale by using Spark GraphX which loading data into memory and running graph analysis in parallel.In this way,we should take data out of graph databases and put it into memory.Considering the limitation of memory size,the premise of accelerating graph analytical process reduces the graph data to a suitable size without too much loss of similarity to the original graph.This paper presents our method of data cleaning on the software graph.We use SEQUITUR data compression algorithm to find out hot code path and store it as a whole paths directed acyclic graph.Hot code path is inherent regularity of a program.About 10 to 200 hot code path account for 40%-99%of a program’s execution cost.These hot paths are acyclic contribute more than 0.1%-1.0%of some execution metric.We expand hot code path to a suitable size which is good for runtime and keeps similarity to the original graph.
文摘为克服依靠图像数据进行识别的局限,使用航迹数据和深度学习方法是当前船型识别的热门方案。针对船型识别任务中常用的长短时记忆(Long Short Term Memory,LSTM)网络对航迹数据特征提取的性能饱和问题,提出了一种基于“Four-hot”编码和Transformer-LSTM神经网络模型的船型识别方法。首先将航迹数据编码为“Four-hot”向量形式;然后构建由Transformer编码模块和LSTM网络级联的Transformer-LSTM神经网络模型,用LSTM网络对Transformer输出的高层表示进行二次学习。在船舶自动识别系统(Automatic Identification System,AIS)数据集上的实验结果显示,所提出的方法在测试集上的加权平均F1分数(Weighted Average F1,WAF1)比未编码航迹数据经5类实验模型直接训练和测试得到的值高3.09百分点以上。
