<?xml version="1.0" standalone="yes"?>
<Paper uid="N06-2035">
  <Title>Weblog Classification for Fast Splog Filtering: A URL Language Model Segmentation Approach</Title>
  <Section position="2" start_page="0" end_page="137" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> The blogosphere, which is a subset of the web and is comprised of personal electronic journals (weblogs) currently encompasses 27.2 million pages and doubles in size every 5.5 months (Technorati, 2006). The information contained in the blogosphere has been proven valuable for applications such as marketing intelligence, trend discovery, and opinion tracking (Hurst, 2005). Unfortunately in the last year the blogosphere has been heavily polluted with spam weblogs (called splogs) which are weblogs used for different purposes, including promoting affiliated websites (Wikipedia, 2006). Splogs can skew the results of applications meant to quantitatively analyze the blogosphere. Sophisticated content-based methods or methods based on link analysis (Gy&amp;quot;ongyi et al., 2004), while providing effective splog filtering, require extra web crawling and can be slow. While a combination of approaches is necessary to provide adequate splog filtering, similar to (Kan &amp; Thi, 2005), we propose, as a preliminary step in the overall splog filtering, a fast, lightweight and accurate method merely based on the analysis of the URL of the weblog without considering its content.</Paragraph>
    <Paragraph position="1"> For quantitative and qualitative analysis of the content of the blogosphere, it is acceptable to eliminate a small fraction of good data from analysis as long as the remainder of the data is splog-free.</Paragraph>
    <Paragraph position="2"> This elimination should be kept to a minimum to preserve counts needed for reliable analysis. When using an ensemble of methods for comprehensive splog filtering it is acceptable for pre-filtering approaches to lower recall in order to improve precision allowing more expensive techniques to be applied on a smaller set of weblogs. The proposed method reaches 93.3% of precision in classifying a weblog in terms of spam or good if 49.1% of the data are left aside (labeled as unknown). If all data needs to be classified our method achieves 78% accuracy which is comparable to the average accuracy of humans (76%) on the same classification task.</Paragraph>
    <Paragraph position="3"> Sploggers, in creating splogs, aim to increase the traffic to specific websites. To do so, they frequently communicate a concept (e.g., a service or a product) through a short, sometimes non-grammatical phrase embedded in the URL of the weblog (e.g., http://adult-video-mpegs.blogspot.com) . We want to build a statistical classifier which leverages the language used in these descriptive URLs in order to classify weblogs as spam or good. We built an initial language model-based classifier on the tokens of the URLs after tokenizing on punctuation (., -,  , /, ?, =, etc.). We ran the system and got an accuracy of 72.2% which is close to the accuracy of humans--76% (the baseline is 50% as the training data is balanced). When we did error analysis on the misclassified examples we observed that many of the mistakes were on URLs that contain words glued together as one token (e.g., dailyfreeipod). Had the words in these tokens been segmented the initial system would have classified the URL correctly. We, thus, turned our attention to additional segmenting of the URLs beyond just punctuation and using this intra-token segmentation in the classification.</Paragraph>
    <Paragraph position="4"> Training a segmenter on standard available text collections (e.g., PTB or BNC) did not seem the way to procede because the lexical items used and the sequence in which they appear differ from the usage in the URLs. Given that we are interested in unsupervised lightweight approaches for URL segmentation, one possibility is to use the URLs themselves after segmenting on punctuation and to try to learn the segmenting (the majority of URLs are naturally segmented using punctuation as we shall see later).</Paragraph>
    <Paragraph position="5"> We trained a segmenter on the tokens in the URLs, unfortunately this method did not provide sufficient improvement over the system which uses tokenization on punctuation. We hypothesized that the content of the splog pages corresponding to the splog URLs could be used as a corpus to learn the segmentation. We crawled 20K weblogs corresponding to the 20K URLs labeled as spam and good in the training set, converted them to text, tokenized and used the token sequences as training data for the segmenter. This led to a statistically significant improvement of 5.8% of the accuracy of the splog filter.</Paragraph>
  </Section>
class="xml-element"></Paper>