Topic Classification of Spam Host based on URLs
Young-joo Chung Masashi Toyoda, Masaru Kitsuregawa · 2010
E-mail: {chung, toyoda, kitsure}@tkl.iis.u-tokyo.ac.jp Abstract In this paper, we determined the main topic of spam hosts based on their uniform resource locator(URL)s. Topic classification of web spam can help personalize spam filters for the web browser, collect topic-specific spam samples with a focused crawler, and understand web spamming activity as a social phenomenon in the cyber space. To classify a URL into different spam topics, we first defined spam topic categories by investigating URLs of spam hosts in our Japanese web archives. Next, we constructed a training set using URLs that were manually categorized into spam topics, and built classifiers using a machine learning approach. In addition, based on the assumption that a small spam link structure consists of pages about a single topic, we used URLs from those structures as additional training data. We categorized URLs of spam hosts from our large scaled Japanese web archive into several topics using two classifiers built by different training sets, and compared the classification results. We improved classification performance from the baseline approach by about 10%. Keyword Web spam, Topic classification, URL, Machine learning