Compressed File Kernel for SVM

This project provides an implementation of a custom kernel for Support Vector Machines that utilizes normalized compression distance (or NCD). The kernel is designed to work with file paths. This method can be particularly useful in scenarios where file content similarity affects the outcome, such as in document classification or in distinguishing types of media files.

System Requirements

Actual requirements are the following:

numpy
joblib
zlib (usually included with Python)

But usually you'd also like to have something like:

scikit-learn
pandas

Latter two will have everything you need (like numpy and joblib), so I've included them to requirements.

Installation

Clone the repository to your local machine:

git clone https://github.com/b0nce/content_agnostic_svm.git
cd content_agnostic_svm

Install the requirements:

pip install -r requirements.txt

Usage

To use this SVM kernel, you need to have a dataset where each instance corresponds to a file path pointing to the file to be used in the classification. Ensure your dataset includes a 'path' column with file paths and a 'label' column with the target labels.

import pandas as pd
from sklearn.svm import SVC
from ncd_kernel import get_kernel

train_df = pd.read_csv("some_random_train_dataset.csv")
test_df = pd.read_csv("some_random_test_dataset.csv")

assert "path" in train_df and "path" in test_df  # X
assert "label" in train_df and "label" in test_df  # y

# If you want to precompute compressed sizes
# kernel = get_kernel(train_df["path"].values, n_jobs=-1)
kernel = get_kernel(n_jobs=-1)
clf = SVC(kernel=kernel)

clf.fit(train_df["path"].values, train_df["label"].values)

test_accuracy = clf.score(test_df["path"].values, test_df["label"].values)

print(f"{test_accuracy = }")