Features and Algorithms Paper by: XIAOGUANG QI and BRIAN D. DAVISON Presentation by: Jason Bender
Outline Introduction to Classification Background Classification Types Classification Methods Applications Features Algorithms Evolution of Websites
What is web page classification? The process of assigning a web page to one or more predefined category labels (ex: news, sports, business…) Classification is generally posed as a supervised learning problem Set of labeled data is used to train a classifier which is applied to label future examples
Background - Classification Types Supervised learning problem broken into sub problems: Subject Classification Functional Classification Sentiment Classification Other types of Classification
Subject Classification Concerned with subject or topic of the web page Judging whether a page is about arts, business, sports, etc… Functional Classification Role that the page is playing Deciding a page to be a personal homepage, course page, admissions page, etc…
Sentiment Classification Focuses on the opinion that is presented in a web page Other types of Classification Such as genre classification and search engine spam classification
Background - Classification Methods Binary vs. Multiclass Single Label vs. Multi Label Soft vs. Hard Flat vs. Hierarchical
Binary vs. Multiclass Classification
Single-Label vs. Multi-Label Classification
Soft vs. Hard Classification
Flat vs. Hierarchical Classification
Applications Why is classification important and how can we use it efficiently?
Constructing, maintaining, or expanding web directories Web directories provide an efficient way to browse for information within a predefined set of categories Example: Open Directory Project Currently constructed by human effort 78,940 editors of ODP
Improving the quality of search results Big problem with search results is search ambiguity
Helping question and answering systems Can use classification systems to help improve the quality of answers Example: Wolfram alpha Other applications Contextual advertising
Features What features can we extract from a web page to use to help classify it?
Features - Introduction Because of features such as the hyperlink …, webpage classification is vastly different from other forms of classification such as plaintext classification. Features organized into two groups: ○ On-page features – directly located on page ○ Neighbor features – found on related pages
On Page Features Textual Contents & Tags Bag-of-words ○ N-gram feature Rather than analyzing individual words, group them into clusters of n-words. -Ex: New York vs. new ….. ….. York Yahoo! Has used a 5-gram feature HTML tags – title, heading, metadata, main text URL
On Page Features Visual Analysis Each page has two representations ○ Text via HTML ○ Visual via the browser Each page can be represented as a visual adjacency multigraph
Features of Neighbors What happens when a page’s features are missing or are unrecognizable?
Features of Neighbors Assumptions If page1 is in the neighborhood of many “sports” pages then there is an increasing probability that page1 is also a “sports” page. Linked pages are more likely to have terms in common
Features of Neighbors Neighbor Selection Focus on pages within 2 steps of target 6 types: parent, child, sibling, spouse, grandparent, and grandchild
Features of Neighbors Labels Anchor Text Surrounding Anchor Text By using the anchor text, surrounding text, and page title of a parent page in combination with text from target page, classification can be improved.
Features of Neighbors Implicit Links Connections between pages that appear in the results of the same query and are both clicked by users
Algorithms What are the algorithmic approaches to webpage classification? Dimension reduction Relational learning Hierarchal classification Information combination
Dimension Reduction Boost classification by emphasizing certain features that are more useful in classification Feature Weighting ○ Reduces the dimensions of feature space ○ Reduces computational complexity ○ Classification more accurate as a result of reduced space
Dimension Reduction Methods Use first fragment K-nearest neighbor algorithm ○ Weighted features ○ Weighted HTML Tags ○ Metrics Expected mutual information Mutual information
Relational Learning Relaxation Labeling
Hierarchical Classification Based on “divide and conquer” Classification problems split into hierarchical set of sub problems. Error Minimization When a lower level category is uncertain of whether page belongs or not, shift assignment one level up.
Information Combination Combine several methods into one Information from different sources are used to train multiple classifiers and the collective work of those classifiers make a final decision.
Conclusion Webpage classification is a type of supervised learning problem aiming to categorize a webpage into a predefined set of categories. In the future, efforts will most likely be focused on effectively combining content and link information to build a more accurate classifier
Evolution of Websites Apple in 1998
Evolution of Websites Apple 2008
Evolution of Websites Nike in 2000
Evolution of Websites Nike in 2008
Evolution of Websites Yahoo in 1996
Evolution of Websites Yahoo in 2008
Evolution of Websites Microsoft in 1998
Evolution of Websites Microsoft in 2008
Evolution of Websites MTV in 1998
Evolution of Websites MTV in 2008
Sources Web Page Classification: Features and Algorithms by Xiaoguang Qi & Brian D. Davison Visual Adjacency Multigraphs – A Novel Approach for a Web Page Classification by Milos Kovacevic, Michelangelo Diligenti, Marco Gori, and Veljko Milutinovic The Evolution of Websites popular-websites.aspx