CVM (Work in progress)

A PySpark package that that implements Kernel SVM classification based on the Cascading SVM idea presented in [1].

Note that currently the implementation only runs on single machines (though on multiple cores), as we are not distributing the model object for prediction.

Install

To install the cvm package, simply run

pip install .

To install the cvm package in development mode, run

pip install -e .

Examples

To obtain the data, run the ./get_data.sh script.

MNIST Digit classification

In this example, we classify the handwritten digit.

From the top level directory, run

[SPARKDIR]/bin/spark-submit --driver-memory 2g bin/mnist.py

Forest Covertype classification

In this example, we try to classify whether the forest covertype is of type 2 or not, such that the problem becomes a binary classification problem.

From the top level directory, run

[SPARKDIR]/bin/spark-submit --driver-memory 2g bin/covertype.py

Authors

• Carlos Riquelme
• Lan Huong
• Sven Schmit

References

[1] Graf, Hans P., Eric Cosatto, Leon Bottou, Igor Dourdanovic, and Vladimir Vapnik. "Parallel support vector machines: The cascade svm." In Advances in neural information processing systems, pp. 521-528. 2004.