An investigation of implicit features in compression-based learning for comparing webpages

We investigate compression-based learning for image classification tasks. These algorithms are claimed to approximate the Kolmogorov complexity of the difference between two object descriptions, but in practice are a measure over an induced feature space. We investigate if these algorithms can be improved via feature selection. Our experiments cover a corpus of legitimate websites and Phishing websites impersonating them; the task is to classify a webpage as either legitimate or a Phish. We perform feature selection in the feature space induced by a well-known compression algorithm (specifically, the entries of the compression dictionary). We then apply four well-known classification algorithms to the reduced feature sets, and conduct a Receiver Operating Characteristic analysis on them. We find that a subset of the features is sufficient for a near-perfect classification of these webpages.