| Title: | Drift Adaptable Models |
| Version: | 1.2.707 |
| Description: | In streaming data analysis, it is crucial to detect significant shifts in the data distribution or the accuracy of predictive models over time, a phenomenon known as concept drift. The package aims to identify when concept drift occurs and provide methodologies for adapting models in non-stationary environments. It offers a range of state-of-the-art techniques for detecting concept drift and maintaining model performance. Additionally, the package provides tools for adapting models in response to these changes, ensuring continuous and accurate predictions in dynamic contexts. Methods for concept drift detection are described in Tavares (2022) <doi:10.1007/s12530-021-09415-z>. |
| License: | MIT + file LICENSE |
| URL: | https://cefet-rj-dal.github.io/heimdall/, https://github.com/cefet-rj-dal/heimdall |
| Encoding: | UTF-8 |
| RoxygenNote: | 7.3.2 |
| Imports: | stats, caret, daltoolbox, ggplot2, reticulate, pROC, car |
| Config/reticulate: | list( packages = list( list(package = "scipy"), list(package = "torch"), list(package = "pandas"), list(package = "numpy"), list(package = "matplotlib"), list(package = "scikit-learn") ) ) |
| NeedsCompilation: | no |
| Packaged: | 2025-05-13 05:24:37 UTC; gpca |
| Author: | Lucas Tavares [aut],
Leonardo Carvalho [aut],
Rodrigo Machado [aut],
Diego Carvalho [ctb],
Esther Pacitti [ctb],
Fabio Porto [ctb],
Eduardo Ogasawara 
[aut, ths, cre],
CEFET/RJ [cph] |
| Maintainer: | Eduardo Ogasawara <eogasawara@ieee.org> |
| Repository: | CRAN |
| Date/Publication: | 2025-05-13 05:40:02 UTC |
ADWIN method
Description
Adaptive Windowing method for concept drift detection doi:10.1137/1.9781611972771.42.
Usage
dfr_adwin(target_feat = NULL, delta = 2e-05)
Arguments
target_feat |
Feature to be monitored. |
delta |
The significance parameter for the ADWIN algorithm. |
Value
dfr_adwin object
Examples
#Use the same example of dfr_cumsum changing the constructor to:
#model <- dfr_adwin(target_feat='serie')
Autoencoder-Based Drift Detection method
Description
Autoencoder-Based method for concept drift detection doi:0.1109/ICDMW58026.2022.00109.
Usage
dfr_aedd(
encoding_size,
ae_class = autoenc_encode_decode,
batch_size = 32,
num_epochs = 1000,
learning_rate = 0.001,
window_size = 100,
monitoring_step = 1700,
criteria = "mann_whitney",
alpha = 0.01,
reporting = FALSE
)
Arguments
encoding_size |
Encoding Size |
ae_class |
Autoencoder Class |
batch_size |
Batch Size for batch learning |
num_epochs |
Number of Epochs for training |
learning_rate |
Learning Rate |
window_size |
Size of the most recent data to be used |
monitoring_step |
The number of rows that the drifter waits to be is updated |
criteria |
The method to be used to check if there is a drift. May be mann_whitney (default), kolmogorov_smirnov, levene |
alpha |
The significance threshold for the statistical test used in criteria |
reporting |
If TRUE, some data are returned as norm_x_oh, drift_input, hist_proj, and recent_proj. |
Value
dfr_aedd object
Examples
#See an example of using `dfr_aedd` at this
#https://github.com/cefet-rj-dal/heimdall/blob/main/multivariate/dfr_aedd.md
Cumulative Sum for Concept Drift Detection (CUMSUM) method
Description
The cumulative sum (CUSUM) is a sequential analysis technique used for change detection.
Usage
dfr_cusum(lambda = 100)
Arguments
lambda |
Necessary level for warning zone (2 standard deviation) |
Value
dfr_cusum object
Examples
library(daltoolbox)
library(heimdall)
# This example uses an error-based drift detector with a synthetic a
# model residual where 1 is an error and 0 is a correct prediction.
data(st_drift_examples)
data <- st_drift_examples$univariate
data$event <- NULL
data$prediction <- st_drift_examples$univariate$serie > 4
model <- dfr_cusum()
detection <- NULL
output <- list(obj=model, drift=FALSE)
for (i in 1:length(data$prediction)){
output <- update_state(output$obj, data$prediction[i])
if (output$drift){
type <- 'drift'
output$obj <- reset_state(output$obj)
}else{
type <- ''
}
detection <- rbind(detection, data.frame(idx=i, event=output$drift, type=type))
}
detection[detection$type == 'drift',]
Adapted Drift Detection Method (DDM) method
Description
DDM is a concept change detection method based on the PAC learning model premise, that the learner’s error rate will decrease as the number of analysed samples increase, as long as the data distribution is stationary. doi:10.1007/978-3-540-28645-5_29.
Usage
dfr_ddm(min_instances = 30, warning_level = 2, out_control_level = 3)
Arguments
min_instances |
The minimum number of instances before detecting change |
warning_level |
Necessary level for warning zone (2 standard deviation) |
out_control_level |
Necessary level for a positive drift detection |
Value
dfr_ddm object
Examples
library(daltoolbox)
library(heimdall)
# This example uses an error-based drift detector with a synthetic a
# model residual where 1 is an error and 0 is a correct prediction.
data(st_drift_examples)
data <- st_drift_examples$univariate
data$event <- NULL
data$prediction <- st_drift_examples$univariate$serie > 4
model <- dfr_ddm()
detection <- NULL
output <- list(obj=model, drift=FALSE)
for (i in 1:length(data$prediction)){
output <- update_state(output$obj, data$prediction[i])
if (output$drift){
type <- 'drift'
output$obj <- reset_state(output$obj)
}else{
type <- ''
}
detection <- rbind(detection, data.frame(idx=i, event=output$drift, type=type))
}
detection[detection$type == 'drift',]
Adapted EWMA for Concept Drift Detection (ECDD) method
Description
ECDD is a concept change detection method that uses an exponentially weighted moving average (EWMA) chart to monitor the misclassification rate of an streaming classifier.
Usage
dfr_ecdd(lambda = 0.2, min_run_instances = 30, average_run_length = 100)
Arguments
lambda |
The minimum number of instances before detecting change |
min_run_instances |
Necessary level for warning zone (2 standard deviation) |
average_run_length |
Necessary level for a positive drift detection |
Value
dfr_ecdd object
Examples
library(daltoolbox)
library(heimdall)
# This example uses a dist-based drift detector with a synthetic dataset.
data(st_drift_examples)
data <- st_drift_examples$univariate
data$event <- NULL
model <- dfr_ecdd()
detection <- NULL
output <- list(obj=model, drift=FALSE)
for (i in 1:length(data$serie)){
output <- update_state(output$obj, data$serie[i])
if (output$drift){
type <- 'drift'
output$obj <- reset_state(output$obj)
}else{
type <- ''
}
detection <- rbind(detection, data.frame(idx=i, event=output$drift, type=type))
}
detection[detection$type == 'drift',]
Adapted Early Drift Detection Method (EDDM) method
Description
EDDM (Early Drift Detection Method) aims to improve the detection rate of gradual concept drift in DDM, while keeping a good performance against abrupt concept drift. doi:2747577a61c70bc3874380130615e15aff76339e
Usage
dfr_eddm(
min_instances = 30,
min_num_errors = 30,
warning_level = 0.95,
out_control_level = 0.9
)
Arguments
min_instances |
The minimum number of instances before detecting change |
min_num_errors |
The minimum number of errors before detecting change |
warning_level |
Necessary level for warning zone |
out_control_level |
Necessary level for a positive drift detection |
Value
dfr_eddm object
Examples
library(daltoolbox)
library(heimdall)
# This example uses an error-based drift detector with a synthetic a
# model residual where 1 is an error and 0 is a correct prediction.
data(st_drift_examples)
data <- st_drift_examples$univariate
data$event <- NULL
data$prediction <- st_drift_examples$univariate$serie > 4
model <- dfr_eddm()
detection <- NULL
output <- list(obj=model, drift=FALSE)
for (i in 1:length(data$prediction)){
output <- update_state(output$obj, data$prediction[i])
if (output$drift){
type <- 'drift'
output$obj <- reset_state(output$obj)
}else{
type <- ''
}
detection <- rbind(detection, data.frame(idx=i, event=output$drift, type=type))
}
detection[detection$type == 'drift',]
Adapted Hoeffding Drift Detection Method (HDDM) method
Description
is a drift detection method based on the Hoeffding’s inequality. HDDM_A uses the average as estimator. doi:10.1109/TKDE.2014.2345382.
Usage
dfr_hddm(
drift_confidence = 0.001,
warning_confidence = 0.005,
two_side_option = TRUE
)
Arguments
drift_confidence |
Confidence to the drift |
warning_confidence |
Confidence to the warning |
two_side_option |
Option to monitor error increments and decrements (two-sided) or only increments (one-sided) |
Value
dfr_hddm object
Examples
library(daltoolbox)
library(heimdall)
# This example uses an error-based drift detector with a synthetic a
# model residual where 1 is an error and 0 is a correct prediction.
data(st_drift_examples)
data <- st_drift_examples$univariate
data$event <- NULL
data$prediction <- st_drift_examples$univariate$serie > 4
model <- dfr_hddm()
detection <- NULL
output <- list(obj=model, drift=FALSE)
for (i in 1:length(data$prediction)){
output <- update_state(output$obj, data$prediction[i])
if (output$drift){
type <- 'drift'
output$obj <- reset_state(output$obj)
}else{
type <- ''
}
detection <- rbind(detection, data.frame(idx=i, event=output$drift, type=type))
}
detection[detection$type == 'drift',]
Inactive dummy detector
Description
Implements Inactive Dummy Detector
Usage
dfr_inactive()
Value
Drifter object
Examples
# See ?hcd_ddm for an example of DDM drift detector
KL Distance method
Description
Kullback Leibler Windowing method for concept drift detection.
Usage
dfr_kldist(target_feat = NULL, window_size = 100, p_th = 0.05, data = NULL)
Arguments
target_feat |
Feature to be monitored. |
window_size |
Size of the sliding window (must be > 2*stat_size) |
p_th |
Probability theshold for the test statistic of the Kullback Leibler distance. |
data |
Already collected data to avoid cold start. |
Value
dfr_kldist object
Examples
library(daltoolbox)
library(heimdall)
# This example uses a dist-based drift detector with a synthetic dataset.
data(st_drift_examples)
data <- st_drift_examples$univariate
data$event <- NULL
model <- dfr_kldist(target_feat='serie')
detection <- NULL
output <- list(obj=model, drift=FALSE)
for (i in 1:length(data$serie)){
output <- update_state(output$obj, data$serie[i])
if (output$drift){
type <- 'drift'
output$obj <- reset_state(output$obj)
}else{
type <- ''
}
detection <- rbind(detection, data.frame(idx=i, event=output$drift, type=type))
}
detection[detection$type == 'drift',]
KSWIN method
Description
Kolmogorov-Smirnov Windowing method for concept drift detection doi:10.1016/j.neucom.2019.11.111.
Usage
dfr_kswin(
target_feat = NULL,
window_size = 1500,
stat_size = 500,
alpha = 1e-07,
data = NULL
)
Arguments
target_feat |
Feature to be monitored. |
window_size |
Size of the sliding window (must be > 2*stat_size) |
stat_size |
Size of the statistic window |
alpha |
Probability for the test statistic of the Kolmogorov-Smirnov-Test The alpha parameter is very sensitive, therefore should be set below 0.01. |
data |
Already collected data to avoid cold start. |
Value
dfr_kswin object
Examples
library(daltoolbox)
library(heimdall)
# This example uses a dist-based drift detector with a synthetic dataset.
data(st_drift_examples)
data <- st_drift_examples$univariate
data$event <- NULL
model <- dfr_kswin(target_feat='serie')
detection <- NULL
output <- list(obj=model, drift=FALSE)
for (i in 1:length(data$serie)){
output <- update_state(output$obj, data$serie[i])
if (output$drift){
type <- 'drift'
output$obj <- reset_state(output$obj)
}else{
type <- ''
}
detection <- rbind(detection, data.frame(idx=i, event=output$drift, type=type))
}
detection[detection$type == 'drift',]
Mean Comparison Distance method
Description
Mean Comparison statistical method for concept drift detection.
Usage
dfr_mcdd(target_feat = NULL, alpha = 1e-08, window_size = 1500)
Arguments
target_feat |
Feature to be monitored |
alpha |
Probability theshold for all test statistics |
window_size |
Size of the sliding window |
Value
dfr_mcdd object
Examples
library(daltoolbox)
library(heimdall)
# This example uses a dist-based drift detector with a synthetic dataset.
data(st_drift_examples)
data <- st_drift_examples$univariate
data$event <- NULL
model <- dfr_mcdd(target_feat='depart_visibility')
detection <- NULL
output <- list(obj=model, drift=FALSE)
for (i in 1:length(data$serie)){
output <- update_state(output$obj, data$serie[i])
if (output$drift){
type <- 'drift'
output$obj <- reset_state(output$obj)
}else{
type <- ''
}
detection <- rbind(detection, data.frame(idx=i, event=output$drift, type=type))
}
detection[detection$type == 'drift',]
Multi Criteria Drifter sub-class
Description
Implements Multi Criteria drift detectors
Usage
dfr_multi_criteria(drifter_list, combination = "or", fuzzy_window = 10)
Arguments
drifter_list |
List of drifters to combine. |
combination |
How the drifters will be combined. Possible values: 'fuzzy', 'or', 'and'. |
fuzzy_window |
Sets the fuzzy window size. Only if combination = 'fuzzy'. |
Value
Drifter object
Adapted Page Hinkley method
Description
Change-point detection method works by computing the observed values and their mean up to the current moment doi:10.2307/2333009.
Usage
dfr_page_hinkley(
target_feat = NULL,
min_instances = 30,
delta = 0.005,
threshold = 50,
alpha = 1 - 1e-04
)
Arguments
target_feat |
Feature to be monitored. |
min_instances |
The minimum number of instances before detecting change |
delta |
The delta factor for the Page Hinkley test |
threshold |
The change detection threshold (lambda) |
alpha |
The forgetting factor, used to weight the observed value and the mean |
Value
dfr_page_hinkley object
Examples
library(daltoolbox)
library(heimdall)
# This example assumes a model residual where 1 is an error and 0 is a correct prediction.
data(st_drift_examples)
data <- st_drift_examples$univariate
data$event <- NULL
data$prediction <- st_drift_examples$univariate$serie > 4
model <- dfr_page_hinkley(target_feat='serie')
detection <- c()
output <- list(obj=model, drift=FALSE)
for (i in 1:length(data$serie)){
output <- update_state(output$obj, data$serie[i])
if (output$drift){
type <- 'drift'
output$obj <- reset_state(output$obj)
}else{
type <- ''
}
detection <- rbind(detection, list(idx=i, event=output$drift, type=type))
}
detection <- as.data.frame(detection)
detection[detection$type == 'drift',]
Passive dummy detector
Description
Implements Passive Dummy Detector
Usage
dfr_passive()
Value
Drifter object
Examples
# See ?hcd_ddm for an example of DDM drift detector
Distribution Based Drifter sub-class
Description
Implements Distribution Based drift detectors
Usage
dist_based(target_feat)
Arguments
target_feat |
Feature to be monitored. |
Value
Drifter object
Drifter
Description
Ancestor class for drift detection
Usage
drifter()
Value
Drifter object
Examples
# See ?dd_ddm for an example of DDM drift detector
Error Based Drifter sub-class
Description
Implements Error Based drift detectors
Usage
error_based()
Value
Drifter object
Examples
# See ?hcd_ddm for an example of DDM drift detector
Process Batch
Description
Process Batch
Usage
## S3 method for class 'drifter'
fit(obj, data, prediction, ...)
Arguments
obj |
Drifter object |
data |
data batch in data frame format |
prediction |
prediction batch as vector format |
... |
opitional arguments |
Value
updated Drifter object
Metric
Description
Ancestor class for metric calculation
Usage
metric()
Value
Metric object
Examples
# See ?metric for an example of DDM drift detector
Accuracy Calculator
Description
Class for accuracy calculation
Usage
mt_accuracy()
Value
Metric object
Examples
# See ?mt_accuracy for an example of Accuracy Calculator
FScore Calculator
Description
Class for FScore calculation
Usage
mt_fscore(f = 1)
Arguments
f |
The F parameter for the F-Score metric |
Value
Metric object
Examples
# See ?mt_fscore for an example of FScore Calculator
Precision Calculator
Description
Class for precision calculation
Usage
mt_precision()
Value
Metric object
Examples
# See ?mt_precision for an example of Precision Calculator
Recall Calculator
Description
Class for recall calculation
Usage
mt_recall()
Value
Metric object
Examples
# See ?mt_recall for an example of Recall Calculator
ROC AUC Calculator
Description
Class for QOC AUC calculation
Usage
mt_rocauc()
Value
Metric object
Examples
# See ?mt_rocauc for an example of ROC AUC Calculator
Multivariate Distribution Based Drifter sub-class
Description
Implements Multivariate Distribution Based drift detectors
Usage
mv_dist_based()
Value
Drifter object
Norm
Description
Ancestor class for normalization techniques
Usage
norm(norm_class)
Arguments
norm_class |
Normalizer class |
Value
Norm object
Examples
# See ?norm for an example of DDM drift detector
Memory Normalizer
Description
Normalizer that has own memory
Usage
nrm_memory(norm_class = minmax())
Arguments
norm_class |
Normalizer class |
Value
Norm object
Examples
# See ?nrm_mimax for an example of Memory Normalizer
Reset State
Description
Reset Drifter State
Usage
reset_state(obj)
Arguments
obj |
Drifter object |
Value
updated Drifter object
Examples
# See ?hcd_ddm for an example of DDM drift detector
Synthetic time series for concept drift detection
Description
A list of multivariate time series for drift detection
example1: a bivariate dataset with one multivariate concept drift example
#'
Usage
data(st_drift_examples)
Format
A list of time series.
Source
References
Examples
data(st_drift_examples)
dataset <- st_drift_examples$example1
Stealthy
Description
Ancestor class for drift adaptive models
Usage
stealthy(
model,
drift_method,
monitored_features = NULL,
norm_class = daltoolbox::zscore(),
warmup_size = 100,
th = 0.5,
target_uni_drifter = FALSE,
incremental_memory = TRUE,
verbose = FALSE,
reporting = FALSE
)
Arguments
model |
The algorithm object to be used for predictions |
drift_method |
The algorithm object to detect drifts |
monitored_features |
List of features that will be monitored by the drifter |
norm_class |
Class used to perform normalization |
warmup_size |
Number of rows used to warmup the drifter. No drift will be detected during this phase |
th |
The threshold to be used with classification algorithms |
target_uni_drifter |
Passes the prediction target to the drifts as the target feat when the drifter is univariate and dist_based. |
incremental_memory |
If true, the model will retrain with all available data whenever the fit is called. If false, it only retrains when a drift is detected. |
verbose |
if TRUE shows drift messages |
reporting |
If TRUE, some data are returned as norm_x_oh, drift_input, hist_proj, and recent_proj. |
Value
Stealthy object
Examples
# See ?dd_ddm for an example of DDM drift detector
Update State
Description
Update Drifter State
Usage
update_state(obj, value)
Arguments
obj |
Drifter object |
value |
a value that represents a processed batch |
Value
updated Drifter object
Examples
# See ?hcd_ddm for an example of DDM drift detector