| Title: | Distance-Based Common Spatial Patterns |
|---|---|
| Description: | A way to apply Distance-Based Common Spatial Patterns (DB-CSP) techniques in different fields, both classical Common Spatial Patterns (CSP) as well as DB-CSP. The method is composed of two phases: applying the DB-CSP algorithm and performing a classification. The main idea behind the CSP is to use a linear transform to project data into low-dimensional subspace with a projection matrix, in such a way that each row consists of weights for signals. This transformation maximizes the variance of two-class signal matrices.The dbcsp object is created to compute the projection vectors. For exploratory and descriptive purpose, plot and boxplot functions can be used. Functions train, predict and selectQ are implemented for the classification step. |
| Authors: | Itziar Irigoien [aut], Concepción Arenas [aut], Itsaso Rodríguez-Moreno [cre, aut] |
| Maintainer: | Itsaso Rodríguez-Moreno <[email protected]> |
| License: | GPL (>= 2) |
| Version: | 0.0.2.1 |
| Built: | 2024-12-04 07:29:02 UTC |
| Source: | CRAN |
dbcsp is a package which offers a way to apply Distance-Based Common Spatial Patterns (DB-CSP) techniques in different fields, both classical Common Spatial Patterns (CSP) as well as DB-CSP.
Itsaso Rodriguez Moreno [email protected]
Itziar Irigoien [email protected]
Concepción Arenas [email protected]
dbcsp, print, summary, train, selectQ, predict, plot, boxplot
# There is an example dataset called AR.data shipped with the package. # It contains the skeleton data extracted from videos of people performing six different actions, # recorded by a humanoid robot. So, it has 6 different classes. # As the whole process is performed pairwise, first two classes are selected (some instances are # saved to use later as test) handshake <- AR.data$handshake[1:15] ignore <- AR.data$ignore[1:15] # Now, the dbcsp object can be created, where q represents the number of vectors used in # the projection, the obtained filter will have 2*q dimension. # By default, euclidean distance is used. To change it, just select another distance type. # A mixture with euclidean distance and another one can be performed too, changing the mixture # parameter value. # mydbcsp <- new('dbcsp', handshake, ignore, type="minkowski", p=0.2) # A user-defined custom distance can be used too. fn <- function(x, y) mean(1 - cos(x - y)) mydbcsp <- new("dbcsp", X1 = handshake, X2 = ignore, type="fn") # Regarding the classification, train, predict and selectQ functions can be used. # An LDA model can be train with the features extracted after performing the CSP, both with the # train function or using the training=TRUE parameter when creating the dbcsp object mydbcsp <- new('dbcsp', handshake, ignore, q=10, training=TRUE, fold = 1) # Once the object is created, print and summary functions can be used to see some details print(mydbcsp) summary(mydbcsp) # The predict function returns the predicted labels for the selected test data. # And if true_labels are indicated, the obtained accuracy is also printed handshake_test <- AR.data$handshake[41:45] ignore_test <- AR.data$ignore[41:45] test_data <- c(handshake_test, ignore_test) true_labels <- c(rep('handshake',length(handshake_test)),rep('ignore',length(ignore_test))) predictions <- predict(mydbcsp, test_data,true_labels) # To help us deciding which is the best dimension to use when performing the CSP algorithm, the # selectQ function can be used. Instead of using train_size to validate with train/test split, # cross validation can be performed too. bestQ <- selectQ(mydbcsp, Q=c(2,3,5), train_size=0.8) # A plot can also be obtained, which displays the signals before and after the CSP projection # With the vectors parameter it can be decided which dimensions to show and if we just want to # plot the transformed signals, the before parameter must be set to FALSE plot(mydbcsp,class=2,index=1,before=FALSE,vectors = 1:5, legend=TRUE) # A boxplot can also be obtained to show the features achieved after the CSP (the variances of # the transformed signals) which are used to perform the classification # In the CSP algorithm the vectors work in pairs (the first q vectors maximize the variance of # one class and minimize the variance of the other, while the last q vectors do the opposite), # by default the vectors are showed in pairs, but this can be changed. # The logarithm of the variances are plotted, but the value of the variances are shown # when changing show_log parameter to FALSE boxplot(mydbcsp, vectors=c(2,4,8))# There is an example dataset called AR.data shipped with the package. # It contains the skeleton data extracted from videos of people performing six different actions, # recorded by a humanoid robot. So, it has 6 different classes. # As the whole process is performed pairwise, first two classes are selected (some instances are # saved to use later as test) handshake <- AR.data$handshake[1:15] ignore <- AR.data$ignore[1:15] # Now, the dbcsp object can be created, where q represents the number of vectors used in # the projection, the obtained filter will have 2*q dimension. # By default, euclidean distance is used. To change it, just select another distance type. # A mixture with euclidean distance and another one can be performed too, changing the mixture # parameter value. # mydbcsp <- new('dbcsp', handshake, ignore, type="minkowski", p=0.2) # A user-defined custom distance can be used too. fn <- function(x, y) mean(1 - cos(x - y)) mydbcsp <- new("dbcsp", X1 = handshake, X2 = ignore, type="fn") # Regarding the classification, train, predict and selectQ functions can be used. # An LDA model can be train with the features extracted after performing the CSP, both with the # train function or using the training=TRUE parameter when creating the dbcsp object mydbcsp <- new('dbcsp', handshake, ignore, q=10, training=TRUE, fold = 1) # Once the object is created, print and summary functions can be used to see some details print(mydbcsp) summary(mydbcsp) # The predict function returns the predicted labels for the selected test data. # And if true_labels are indicated, the obtained accuracy is also printed handshake_test <- AR.data$handshake[41:45] ignore_test <- AR.data$ignore[41:45] test_data <- c(handshake_test, ignore_test) true_labels <- c(rep('handshake',length(handshake_test)),rep('ignore',length(ignore_test))) predictions <- predict(mydbcsp, test_data,true_labels) # To help us deciding which is the best dimension to use when performing the CSP algorithm, the # selectQ function can be used. Instead of using train_size to validate with train/test split, # cross validation can be performed too. bestQ <- selectQ(mydbcsp, Q=c(2,3,5), train_size=0.8) # A plot can also be obtained, which displays the signals before and after the CSP projection # With the vectors parameter it can be decided which dimensions to show and if we just want to # plot the transformed signals, the before parameter must be set to FALSE plot(mydbcsp,class=2,index=1,before=FALSE,vectors = 1:5, legend=TRUE) # A boxplot can also be obtained to show the features achieved after the CSP (the variances of # the transformed signals) which are used to perform the classification # In the CSP algorithm the vectors work in pairs (the first q vectors maximize the variance of # one class and minimize the variance of the other, while the last q vectors do the opposite), # by default the vectors are showed in pairs, but this can be changed. # The logarithm of the variances are plotted, but the value of the variances are shown # when changing show_log parameter to FALSE boxplot(mydbcsp, vectors=c(2,4,8))
A dataset containing the skeleton data extracted from videos of people performing six different actions, recorded by a humanoid robot. Each class has several instances with 50 values (25 XY body keypoints extracted using OpenPose https://github.com/CMU-Perceptual-Computing-Lab/openpose) per frame of the video (92 frames).
AR.dataAR.data
A list of 6 different actions:
Gesture for telling the robot to come to you. 46 instances of [50x92] matrices.
Gesture of 'high five'. 45 instances of [50x92] matrices.
Gesture of handshaking with the robot. 45 instances of [50x92] matrices.
Gesture for telling hello to the robot. 44 instances of [50x92] matrices.
Ignore the robot, pass by. 46 instances of [50x92] matrices.
Stare at the robot in front of it. 46 instances of [50x92] matrices.
X1 <- AR.data$come X2 <- AR.data$five mydbcsp <- new('dbcsp',X1,X2)X1 <- AR.data$come X2 <- AR.data$five mydbcsp <- new('dbcsp',X1,X2)
This function plots the variance of the selected vectors in a boxplot.
## S3 method for class 'dbcsp' boxplot(x, vectors=1, pairs=TRUE, ordered_pairs=TRUE, show_log=TRUE,...)## S3 method for class 'dbcsp' boxplot(x, vectors=1, pairs=TRUE, ordered_pairs=TRUE, show_log=TRUE,...)
x |
object of class |
vectors |
integer or array of integers, indicating the index of the projected vectors to plot, by default |
pairs |
logical, if TRUE the pairs of the indicated vectors are also shown, by default |
ordered_pairs |
logical, if TRUE the pairs are plotted next to each other, else pairs are plotted at the end, by default |
show_log |
logical, if TRUE the logarithms of the variances are displayed, else the variances are displayed, by default |
... |
not currently used. |
A boxplot with the variances of the selected vectors.
Vectors values must lie between 1 and 2*q, being q the number of dimensions used to perform the DB-CSP
algorithm when creating the dbcsp object. The following should be
taken into account when plotting:
The first q values (1,...,q) are indicated as a1...aq
The last q values (q+1,...,2*q) are indicated as b1...bq.
If pairs=TRUE, it is recommended that vectors<q for better understanding,
since their pairs are plotted as well. In case that vectors>q, it should be noted
that the values are displayed from b1 to bq, where b1 and bq represent q+1 vector and 2*q vector,
respectively.
For example if q=15 and boxplot(object, vectors=16, pairs=FALSE), b1 (16-q=1) vector is
shown.
Among the selected boxplots, the largest whiskers are obtained and just the outliers within these whiskers are shown, the rest are not displayed. With the outliers which are outside the whiskers it is not possible to get a good visualization.
Displays the boxplot of the variances of the selected vectors.
dbcsp, print, summary, train, selectQ, predict, plot
# Read data from 2 classes x <- AR.data$come y <- AR.data$five mydbcsp <- new("dbcsp", X1 = x, X2 = y) boxplot(mydbcsp) boxplot(mydbcsp,vectors=1:4,pairs=FALSE) boxplot(mydbcsp, vectors=c(1,4,7),ordered_pairs=FALSE)# Read data from 2 classes x <- AR.data$come y <- AR.data$five mydbcsp <- new("dbcsp", X1 = x, X2 = y) boxplot(mydbcsp) boxplot(mydbcsp,vectors=1:4,pairs=FALSE) boxplot(mydbcsp, vectors=c(1,4,7),ordered_pairs=FALSE)
An object of class dbcsp. 'dbcsp' stands for Distance-Based Common Spatial Patterns. The object includes the Common Spatial Patterns filter obtained with the input lists and using the distance method indicated.
If the lists of matrices X1 or X2 contain NA values, these are automatically interpolated by a linear interpolation
using na.approx function. These new interpolated matrices are saved in the X1 and X2 slots of the object.
The supported distances for type are these ones:
Included in TSdist: infnorm, ccor, sts, lb.keogh, edr, erp, lcss, fourier, tquest, dissim, acf, pacf, ar.lpc.ceps, ar.mah,
ar.mah.statistic, ar.mah.pvalue, ar.pic, cdm, cid, cor, cort, int.per, per, mindist.sax, ncd, pred, spec.glk, spec.isd,
spec.llr, pdc, frechet, tam.
Included in parallelDist: bhjattacharyya, bray, canberra, chord, divergence, dtw, euclidean, fJaccard, geodesic, hellinger,
kullback, mahalanobis, manhattan, maximum, minkowski, podani, soergel, wave, whittaker.
It is possible to use a custom distance. The name of the custom distance function is passed as character to the type parameter.
In order to use the parallelDist custom distance option, the custom function must be defined as
explained in "Details: User-defined distance functions" part of parallelDist documentation.
See Examples section below.
The additional parameters for the selected distance (see TSdist, parallelDist) can be passed
as parameters when creating the object, which will be saved in more slot. See Examples section below.
The output is a list containing this information (object@out):
vectors The projection vectors obtained after applying CSP.
eig The eigenvalues obtained after applying CSP.
proy The variance values of the projected signals obtained after applying CSP.
And if training=TRUE the following values are also saved:
acc The mean accuracy value obtained for training data applying cross validation.
used_folds List of the folds used in the cross validation.
folds_acc Accuracy values for each of the folds of the cross validation.
model The trained LDA classifier.
selected_q The number of vectors used when training.
X1list of matrices for data class 1.
X2list of matrices for data class 2.
qinteger value indicating the number of vectors used in the projection, by default q=15.
labelsvector of two strings indicating labels names, by default names of variables X1 and X2.
typestring which sets the type of distance to be considered, by default type='EUCL'. See details section.
wweight for the distances mixture D_mixture = w*D_euclidean + (1-w)*D_type, by default w=0.5.
mixturelogical value indicating whether to use distances mixture or not (EUCL + other), by default mixture=FALSE.
traininglogical value indicating whether to perform the training or not.
foldinteger value, by default fold=10. It controls the number of partitions when training.
If fold==1 a train/test split is performed, with p=0.2 for test indices.
seednumeric value, by default seed=NULL. Set a seed to ensure reproducible results.
eig.tolnumeric value, by default eig.tol=1e-06, tolerance to convert distance matrix to be definite positive.
verboselogical
morelist, additional parameters to be passed to the distance methods. See details section.
outlist containing the output.
dbcsp, print, summary, train, selectQ, predict, plot, boxplot
# To create an instance of a class dbcsp given data from 2 classes x <- AR.data$come[1:20] y <- AR.data$five[1:20] mydbcsp <- new("dbcsp", X1 = x, X2 = y) # CUSTOM DISTANCE x <- AR.data$come[1:10] y <- AR.data$five[1:10] fn <- function(x, y, eps=1) mean(1 - cos(x - y))*eps mydbcsp <- new("dbcsp", X1 = x, X2 = y, type="fn", eps=0.9)# To create an instance of a class dbcsp given data from 2 classes x <- AR.data$come[1:20] y <- AR.data$five[1:20] mydbcsp <- new("dbcsp", X1 = x, X2 = y) # CUSTOM DISTANCE x <- AR.data$come[1:10] y <- AR.data$five[1:10] fn <- function(x, y, eps=1) mean(1 - cos(x - y))*eps mydbcsp <- new("dbcsp", X1 = x, X2 = y, type="fn", eps=0.9)
This function plots an instance before and/or after its DB-CSP projection.
## S3 method for class 'dbcsp' plot(x, class = 1, index = 1, vectors = 1:(x@q*2), pairs=TRUE, before = TRUE, after = TRUE, legend = FALSE, getsignals = FALSE, ...)## S3 method for class 'dbcsp' plot(x, class = 1, index = 1, vectors = 1:(x@q*2), pairs=TRUE, before = TRUE, after = TRUE, legend = FALSE, getsignals = FALSE, ...)
x |
object of class |
class |
integer, which of both classes to access (1 or 2), by default |
index |
an integer, representing which instance of the class to plot, by default |
vectors |
an integer or vector of integers, representing the vectors to plot after the projection, by default all the vectors used in the projection are plotted |
pairs |
logical, if TRUE the pairs of the indicated vectors are also shown, by default |
before |
logical, if TRUE the original signals are plotted, by default |
after |
logical, if TRUE the signals after projection are plotted, by default |
legend |
logical, if true the legend of the transformed signals is shown, by default |
getsignals |
logical, if TRUE the projected signals for the selected class, instance and vectors are returned, by default |
... |
optional arguments inherited from the |
It plots an instance before and/or after being projected with the DB-CSP filter.
Vectors values must lie between 1 and 2*q, being q the number of dimensions used to perform the DB-CSP
algorithm when creating the dbcsp object. The following should be
taken into account when plotting:
The first q values (1,...,q) are indicated as a1...aq, and are plotted with solid lines.
The last q values (q+1,...,2*q) are indicated as b1...bq, and are plotted with dashed lines.
If pairs=TRUE, it is recommended that vectors<q for better understanding,
since their pairs are plotted as well. In case that vectors>q, it should be noted
that the values are displayed from b1 to bq, where b1 and bq represent q+1 vector and 2*q vector,
respectively. The paired vectors (a1-b1, a2-b2, ...) are plotted with the same color, but different line type.
For example if q=15 and plot(object, vectors=16, pairs=FALSE), b1 (16-q=1) vector is
shown.
The number of rows and columns of the layout (mfrow, mfcol) can not be modified, as the function select them
according to before and after parameters.
Displays a plot of the selected instance before and/or after the DB-CSP filter projection.
The vectors shown after the projection are differentiated by the q first and q last vectors,
since the former maximize the variance of one class and minimize the variance of the other, while the latter do the opposite.
If getsignals=TRUE, a matrix with the projected signals shown in the plot is returned.
dbcsp, print, summary, train, selectQ, predict, boxplot
# Read data from 2 classes x <- AR.data$come y <- AR.data$five mydbcsp <- new("dbcsp", X1 = x, X2 = y) plot(mydbcsp) plot(mydbcsp,class=2,index=30,vectors=1:5,before=FALSE, legend=TRUE) pSignals <- plot(mydbcsp,class=2,index=30,vectors=1:5,before=FALSE, legend=TRUE,getsignals=TRUE)# Read data from 2 classes x <- AR.data$come y <- AR.data$five mydbcsp <- new("dbcsp", X1 = x, X2 = y) plot(mydbcsp) plot(mydbcsp,class=2,index=30,vectors=1:5,before=FALSE, legend=TRUE) pSignals <- plot(mydbcsp,class=2,index=30,vectors=1:5,before=FALSE, legend=TRUE,getsignals=TRUE)
This function returns the labels predicted for the input instances. If true_targets are passed as parameter,
the accuracy obtained is printed too.
## S3 method for class 'dbcsp' predict(object, X_test, true_targets=NULL, ...)## S3 method for class 'dbcsp' predict(object, X_test, true_targets=NULL, ...)
object |
object of class |
X_test |
list of matrices for test data. |
true_targets |
vector of true labels of the instances. Note that they must match the names of the labels used when training the model. |
... |
not currently used. |
It gives the predictions for the test data using the model saved in the object, which has been previously trained with
the train.dbcsp function. If the true_targets are indicated, the confusion matrix and obtained accuracy value are
returned too.
The values returned by the LDA predict function, a list with these components:
class The MAP classification (a factor)
posterior Posterior probabilities for the classes
x The scores of test cases on up to dimen discriminant variables
If the true_targets are indicated, two more items are added to the output list:
confusion_matrix The confusion matrix obtained with predicted labels and true labels.
acc The accuracy value obtained for the test instances.
dbcsp, print, summary, train, selectQ, plot, boxplot
# Read data from 2 classes x <- AR.data$come[1:20] y <- AR.data$five[1:20] mydbcsp <- new("dbcsp", X1 = x, X2 = y) mydbcsp <- train(mydbcsp,fold=3) test_data <- c(AR.data$come[20:24], AR.data$five[20:24]) test_labels <- c(rep('x',5),rep('y',5)) predictions <- predict(mydbcsp,test_data,test_labels) # Predicted classes print(predictions$class) # Confusion matrix print(predictions$confusion_matrix) # Accuracy print(predictions$acc)# Read data from 2 classes x <- AR.data$come[1:20] y <- AR.data$five[1:20] mydbcsp <- new("dbcsp", X1 = x, X2 = y) mydbcsp <- train(mydbcsp,fold=3) test_data <- c(AR.data$come[20:24], AR.data$five[20:24]) test_labels <- c(rep('x',5),rep('y',5)) predictions <- predict(mydbcsp,test_data,test_labels) # Predicted classes print(predictions$class) # Confusion matrix print(predictions$confusion_matrix) # Accuracy print(predictions$acc)
This function prints information about dbcsp class.
## S3 method for class 'dbcsp' print(x, ...)## S3 method for class 'dbcsp' print(x, ...)
x |
object of class |
... |
not currently used. |
It provides information about the object and the class.
No return value, called for side effects.
dbcsp, summary, train, selectQ, predict, plot, boxplot
# Read data from 2 classes x <- AR.data$come[1:30] y <- AR.data$five[1:30] mydbcsp <- new("dbcsp", X1 = x, X2 = y) print(mydbcsp)# Read data from 2 classes x <- AR.data$come[1:30] y <- AR.data$five[1:30] mydbcsp <- new("dbcsp", X1 = x, X2 = y) print(mydbcsp)
This function applies DB-CSP and classification with different dimensions to see which gets the best outcomes.
selectQ( object, Q = c(1, 2, 3, 5, 10, 15), train_size = 0.75, CV = FALSE, folds = 10, seed = NULL ) ## S4 method for signature 'dbcsp' selectQ( object, Q = c(1, 2, 3, 5, 10, 15), train_size = 0.75, CV = FALSE, folds = 10, seed = NULL )selectQ( object, Q = c(1, 2, 3, 5, 10, 15), train_size = 0.75, CV = FALSE, folds = 10, seed = NULL ) ## S4 method for signature 'dbcsp' selectQ( object, Q = c(1, 2, 3, 5, 10, 15), train_size = 0.75, CV = FALSE, folds = 10, seed = NULL )
object |
object of class |
Q |
list of integers which represents the dimensions to use, by default |
train_size |
float between 0.0 and 1.0 representing the proportion of the dataset to include in the train split, by default |
CV |
logical indicating if a cross validation must be performed or not (if TRUE, train_size is not used), by default |
folds |
integer, number of folds to use if CV is performed. |
seed |
numeric value, by default |
A data.frame including the dimensions and their corresponding accuracy values.
If CV=TRUE, for each dimension, the standard deviation of the accuracy values of the folds is also included in the data frame.
dbcsp, print, summary, train, predict, plot, boxplot
# Read data from 2 classes x <- AR.data$come y <- AR.data$five mydbcsp <- new("dbcsp", X1 = x, X2 = y) result <- selectQ(mydbcsp) print(result)# Read data from 2 classes x <- AR.data$come y <- AR.data$five mydbcsp <- new("dbcsp", X1 = x, X2 = y) result <- selectQ(mydbcsp) print(result)
This function provides a summary of the dbcsp object and information about the performed process.
## S3 method for class 'dbcsp' summary(object, ...)## S3 method for class 'dbcsp' summary(object, ...)
object |
object of class |
... |
not currently used. |
It prints the following information:
Length and shape of the list of matrices of each class.
The number of vectors (dimensions) used in the CSP projection.
Distance used when performing the Common Spatial Patterns algorithm.
If the training process has already been performed, the obtained training accuracy value.
No return value, called for side effects.
dbcsp, print, train, selectQ, predict, plot, boxplot
# Read data from 2 classes x <- AR.data$come[1:30] y <- AR.data$five[1:30] mydbcsp <- new("dbcsp", X1 = x, X2 = y) summary(mydbcsp)# Read data from 2 classes x <- AR.data$come[1:30] y <- AR.data$five[1:30] mydbcsp <- new("dbcsp", X1 = x, X2 = y) summary(mydbcsp)
This function applies DB-CSP to the instances and perform the training of a Linear Discriminant Analysis (LDA) classifier using the object data.
## S3 method for class 'dbcsp' train(x, selected_q=x@q, fold=x@fold, seed=x@seed, verbose=TRUE,...)## S3 method for class 'dbcsp' train(x, selected_q=x@q, fold=x@fold, seed=x@seed, verbose=TRUE,...)
x |
object of class |
selected_q |
integer value indicating the number of vectors to use when training the model, by default |
fold |
integer value, by default |
seed |
numeric value, by default |
verbose |
logical |
... |
not currently used. |
The dbcsp object with the training results saved as list in x@out:
vectors The projection vectors obtained after applying CSP.
eig The eigenvalues obtained after applying CSP.
proy The variance values of the projected signals obtained after applying CSP.
acc The mean accuracy value obtained for training data applying cross validation.
used_folds List of the folds used in the cross validation.
folds_acc Accuracy values for each of the folds of the cross validation.
model The trained LDA classifier.
selected_q The number of vectors used when training.
dbcsp, print, summary, selectQ, predict, plot, boxplot
# Read data from 2 classes x <- AR.data$come[1:20] y <- AR.data$five[1:20] mydbcsp <- new("dbcsp", X1 = x, X2 = y) mydbcsp <- train(mydbcsp,fold=3) print(mydbcsp@out$acc)# Read data from 2 classes x <- AR.data$come[1:20] y <- AR.data$five[1:20] mydbcsp <- new("dbcsp", X1 = x, X2 = y) mydbcsp <- train(mydbcsp,fold=3) print(mydbcsp@out$acc)