Makes Weka algorithms available in scikit-learn.
Built on top of the python-weka-wrapper3 library, it uses the python-javabridge library under the hood for communicating with Weka objects in the Java Virtual Machine.
The following is currently available:
- Classifiers (classification/regression)
- Clusters
- Filters
Things to be aware of:
- You need to start the JVM in your Python code before you can use Weka (and stop it again).
- Unlikely to work in multi-threaded/process environments (like flask).
- Jupyter Notebooks do not play nice with python-javabridge, as you might have to restart the kernel in order to be able to restart the JVM (e.g., with additional packages).
- The conversion to Weka data structures involves guesswork, i.e., if targets are to be treated as nominal, you need
to convert the numeric values to strings (e.g., using
to_nominal_labelsand/orto_nominal_attributesfunctions fromsklweka.datasetor theMakeNominaltransformer fromsklweka.preprocessing). - Check the list of known problems before reporting one.
The library has the following requirements:
-
Python 3 (does not work with Python 2)
- python-weka-wrapper (>=0.2.5, required)
-
OpenJDK 8 or later (11 is recommended)
-
install the python-weka-wrapper3 library in a virtual environment, see instructions here:
https://fracpete.github.io/python-weka-wrapper3/install.html
-
install the sklearn-weka-plugin library itself in the same virtual environment
-
latest release from PyPI
./venv/bin/pip install sklearn-weka-plugin -
from local source
./venv/bin/pip install . -
from Github repository
./venv/bin/pip install git+https://github.com/fracpete/sklearn-weka-plugin.git
-
Here is a quick example (of which you need to adjust the paths to the datasets, of course):
import sklweka.jvm as jvm
from sklweka.dataset import load_arff, to_nominal_labels
from sklweka.classifiers import WekaEstimator
from sklweka.clusters import WekaCluster
from sklweka.preprocessing import WekaTransformer
from sklearn.model_selection import cross_val_score
from sklweka.datagenerators import DataGenerator, generate_data
# start JVM with Weka package support
jvm.start(packages=True)
# regression
X, y, meta = load_arff("/some/where/bolts.arff", class_index="last")
lr = WekaEstimator(classname="weka.classifiers.functions.LinearRegression")
scores = cross_val_score(lr, X, y, cv=10, scoring='neg_root_mean_squared_error')
print("Cross-validating LR on bolts (negRMSE)\n", scores)
# classification
X, y, meta = load_arff("/some/where/iris.arff", class_index="last")
y = to_nominal_labels(y)
j48 = WekaEstimator(classname="weka.classifiers.trees.J48", options=["-M", "3"])
j48.fit(X, y)
scores = j48.predict(X)
probas = j48.predict_proba(X)
print("\nJ48 on iris\nactual label -> predicted label, probabilities")
for i in range(len(y)):
print(y[i], "->", scores[i], probas[i])
# clustering
X, y, meta = load_arff("/some/where/iris.arff", class_index="last")
cl = WekaCluster(classname="weka.clusterers.SimpleKMeans", options=["-N", "3"])
clusters = cl.fit_predict(X)
print("\nSimpleKMeans on iris\nclass label -> cluster")
for i in range(len(y)):
print(y[i], "->", clusters[i])
# preprocessing
X, y, meta = load_arff("/some/where/bolts.arff", class_index="last")
tr = WekaTransformer(classname="weka.filters.unsupervised.attribute.Standardize", options=["-unset-class-temporarily"])
X_new, y_new = tr.fit(X, y).transform(X, y)
print("\nStandardize filter")
print("\ntransformed X:\n", X_new)
print("\ntransformed y:\n", y_new)
# generate data
gen = DataGenerator(
classname="weka.datagenerators.classifiers.classification.BayesNet",
options=["-S", "2", "-n", "10", "-C", "10"])
X, y, X_names, y_name = generate_data(gen, att_names=True)
print("X:", X_names)
print(X)
print("y:", y_name)
print(y)
# stop JVM
jvm.stop()See the example repository for more examples:
https://github.com/fracpete/sklearn-weka-plugin-examples
Direct links:
You can find the project documentation here:
