Using the time-series synthesized data expansion method for classifying data
Data augmentation using synthetic data for time series classification with deep residual networks
Using the time-series synthesized data expansion method for classifying data
Abstract
Data enhancement technology is widely used in computer vision, for a small number of sample data sets for model fit easily reached using data enhancement technology can effectively seen the degree of fit. However, in the time-series data classification enhancement technology is very limited, the authors proposed based on data from DTW enhancement technology to fill the gaps in this regard.
Introduction
For image enhancement technology for data achieved good results, but the application in time sequence from the effect is not ideal. This may be because, for the picture is enhanced after the data does not change the picture category, such as a picture of a cat, by the picture translation, rotation, scaling transformation does not take a cat into a dog, and for the time sequences, one can not easily control the impact of this particular conversion time series properties.
This paper proposes a time series-based enhancement DTW technique, by carrying out experiments on the UCR data set, the data show that enhanced results can greatly improve the accuracy of the neural network model for certain data sets, while other smaller data sets negative impact. Finally, these two kinds of decision-making training model combining instructions on how to ensure that the data set while high gain accuracy, effectively reduce the negative impact of rare data increase brings.
Method
- Architecture
Residual network using the network structure, the network is the input of a length lof a single variable time series. Output from the data set Cconsisting of the class probability distribution. The core network 3 comprises a residual block, followed by a global layer and a mean cell classification softmax layer. Each block contains a residual three layers 1-D convolution, convolution layer size are 8,5,3. After each layer is a convolution is then batch normalization Relu activation unit. Number 3 convolution filter blocks are 64,128,128.
All the network parameters are initialized using the initialization method Glorot's Uniform optimizer uses Adam, the learning rate is set to 0.001, the exponential decay rate of the first-order moment and the second-order moment estimation value estimated values are set to 0.999 and 0.9, and finally using the cross-entropy as the cost of network functions.
- Data augmentation
Use of a time series to form a weighted average of the center of gravity DTW (DBA) Technical data enhancement, by simply changing the weight of the weights, the method may create an infinite number of new time series given from a set of time series. A method of using a weighted average of selected method called three weighting process.
- From the initial training set randomly select a time series began, we gave it an equal weight of 0.5. The randomly selected time sequence as a time series of first test the DBA.
- The DTW distance to find the time series of the first test and DBA Recent time series. And randomly selecting two of these five, and the two weights are set to 0.15.
- In order to make the sum of the weights equals 1, the right sequence of the remainder of the training set and the weight is 1 - 0.15 * 2 - 0.5 = 0.2 1-0.15 2-0.5 * = 0.2, the average remaining time series allocates the weights 0.2 .
- DBA algorithm generated using a weighted average of the sequence.
Other average sequence generation algorithm will be used in future research author.
Result
The results show that in the worst case, data enhancement can significantly improve the accuracy of the depth learning model, while producing less negative impact on some of the data sets. To make the data more enhanced generalization properties, the use of two integrated networks residuals (data not enhanced and increased data), in fact, is the posterior probability of each of the two class classifier output averaged value, then the maximum average probability distribution of labels for each time series, so as to generate a time series sample outside provides a more robust approach. The results show that the most benefit from the data set the data expansion, the accuracy almost no change.
[Original Address] [ https://arxiv.org/abs/1808.02455 ]
Data augmentation using synthetic data for time series classification with deep residual networks
Using the time-series synthesized data expansion method for classifying data
Abstract
Data enhancement technology is widely used in computer vision, for a small number of sample data sets for model fit easily reached using data enhancement technology can effectively seen the degree of fit. However, in the time-series data classification enhancement technology is very limited, the authors proposed based on data from DTW enhancement technology to fill the gaps in this regard.
Introduction
For image enhancement technology for data achieved good results, but the application in time sequence from the effect is not ideal. This may be because, for the picture is enhanced after the data does not change the picture category, such as a picture of a cat, by the picture translation, rotation, scaling transformation does not take a cat into a dog, and for the time sequences, one can not easily control the impact of this particular conversion time series properties.
This paper proposes a time series-based enhancement DTW technique, by carrying out experiments on the UCR data set, the data show that enhanced results can greatly improve the accuracy of the neural network model for certain data sets, while other smaller data sets negative impact. Finally, these two kinds of decision-making training model combining instructions on how to ensure that the data set while high gain accuracy, effectively reduce the negative impact of rare data increase brings.
Method
- Architecture
Residual network using the network structure, the network is the input of a length lof a single variable time series. Output from the data set Cconsisting of the class probability distribution. The core network 3 comprises a residual block, followed by a global layer and a mean cell classification softmax layer. Each block contains a residual three layers 1-D convolution, convolution layer size are 8,5,3. After each layer is a convolution is then batch normalization Relu activation unit. Number 3 convolution filter blocks are 64,128,128.
All the network parameters are initialized using the initialization method Glorot's Uniform optimizer uses Adam, the learning rate is set to 0.001, the exponential decay rate of the first-order moment and the second-order moment estimation value estimated values are set to 0.999 and 0.9, and finally using the cross-entropy as the cost of network functions.
- Data augmentation
Use of a time series to form a weighted average of the center of gravity DTW (DBA) Technical data enhancement, by simply changing the weight of the weights, the method may create an infinite number of new time series given from a set of time series. A method of using a weighted average of selected method called three weighting process.
- From the initial training set randomly select a time series began, we gave it an equal weight of 0.5. The randomly selected time sequence as a time series of first test the DBA.
- The DTW distance to find the time series of the first test and DBA Recent time series. And randomly selecting two of these five, and the two weights are set to 0.15.
- In order to make the sum of the weights equals 1, the right sequence of the remainder of the training set and the weight is 1 - 0.15 * 2 - 0.5 = 0.2 1-0.15 2-0.5 * = 0.2, the average remaining time series allocates the weights 0.2 .
- DBA algorithm generated using a weighted average of the sequence.
Other average sequence generation algorithm will be used in future research author.
Result
The results show that in the worst case, data enhancement can significantly improve the accuracy of the depth learning model, while producing less negative impact on some of the data sets. To make the data more enhanced generalization properties, the use of two integrated networks residuals (data not enhanced and increased data), in fact, is the posterior probability of each of the two class classifier output averaged value, then the maximum average probability distribution of labels for each time series, so as to generate a time series sample outside provides a more robust approach. The results show that the most benefit from the data set the data expansion, the accuracy almost no change.
[Original Address] [ https://arxiv.org/abs/1808.02455 ]