from skmultiflow.data import FileStream stream = FileStream("./src/skmultiflow/data/datasets/covtype.csv") stream.prepare_for_use() stream.n_classes # Output: 7 stream.target_values # Output: [1, 2, 3, 4, 5, 6, 7]
Thanks @jacobmontiel . I run exactly the same codes with you. But the error still exists. However, when I redownload the data, the error is fixed. So, maybe the data source is not saved properly at first. Thanks for your reply.
Glad to hear that it is working. It is strange that it just went away, in any case we will keep it in mind in case somebody else gets the same error.
Hi @jacobmontiel I want to know if all methods for Concept Drift Detection included in skmultiflow.drift_detection only support 1-D data stream. For example, when using a 2-d (size=[2000,5]) data_stream in the following codes, an error will arise.
import numpy as np
from skmultiflow.drift_detection import PageHinkley
ph = PageHinkley()
data_stream = np.random.randint(2, size=[2000,5])
for i in range(999, 2000):
data_stream[i] = np.random.randint(4, high=8,size=5)
for i in range(2000):
print('Change has been detected in data: ' + str(data_stream[i]) + ' - of index: ' + str(i))
from skmultiflow.drift_detection import PageHinkley data_stream = np.concatenate((np.random.randint(2, size=1000), np.random.randint(4, size=1000))) ph = PageHinkley() for i, val in enumerate(data_stream): ph.add_element(val) if ph.detected_change(): print('Change has been detected in data: ' + str(data_stream[i]) + ' - of index: ' + str(i)) ph.reset()
scikit-multiflow. As you mention the sklearn implementation works in batches of data, if you wan to update the densities you have to define data update strategy. This is very similar to how the
KNNClassifieris implemented. You will see there that the data is stored in a sliding window. Regarding drift detection, ADWIN as all other drift detectors take as input 1-dimensional data. You can check the
KNNADWINClassifierwhich uses ADWIN to monitor the classification performance of the basic KNN model. If ADWIN detects a change in classification performance, then the model is reset.
Hi @dossy , here is one
from skmultiflow.data import DataStream import numpy as np n_features = 10 n_samples = 50 X = np.random.random(size=(n_samples, n_feature y = np.random.randint(2, size=n_samples) stream = DataStream(data=X, y=y) # stream.prepare_for_use() # if using the stable version (0.4.1) stream.n_remaining_samples()
Last line return
np.ndarrayas long as you define the index of the target column (last column by default).
pandas.DataFrameare also supported, following the same indications.
scikit-multiflowis only a small part of the puzzle and there’s a lot of stuff you have to develop yourself around it?
I know this is a n00b question, but if I’m working with strings, I have to vectorize them first? I can’t just pass in a
pandas.DataFramecontaining strings - only real/int values?
All questions are welcomed. Currently, we only support numerical data. Your data must be pre-processed. As you mention, scikit-multiflow is focused on the learning part. The idea is that you can take it and integrate it as part of your workflow.
scikit-multiflowprocesses data one sample at a time. We provide the
DataStreamclasses for the case when you have data in a file or in memory. Both are extensions of the
Streamclass. If you want to read from log file you could process each log as it arrives (convert to numerical values) and the pass it to a model. The operation to receive and process the log entry can be wrapped in an extension of the
Streamclass. The most relevant method is
Hi, I am building a HoeffdingTree classifier on a heavily imbalanced data stream (only ~1 in 1000 data points are of the positive class). Using the
EvaluatePrequential evaluator I am able to plot the precision and recall, however, the recall is extremely low as the model learns to predict the negative class almost always (only 50 positive predictions in my stream of 10 million data points).
Tree classifiers often give me class probabilities rather than discrete class outputs, and the actual recall (and precision) is of course threshold-dependent. Is there a way to control the threshold for which I am evaluating the recall?