CSE5230/DMS/2002/11 Data Mining - CSE5230 Web Mining CSE5230 - Data Mining, 2002 Lecture 11.1 Lecture Outline How big is the web? What is “web data”? A taxonomy of web mining tasks Example: targeted advertising Example: personalization References CSE5230 - Data Mining, 2002 Lecture 11.2 How big is the web? It is not easy to determine the size of the web In 1999, one estimate was that there were approximately 350 million web pages, growing at about 1 million pages per day In 2001, Google announced that they were indexing around 3 billion web documents matter which of these is more accurate – it’s very big! We can view the web as the world’s biggest database No The word “database” is used loosely here, because the web has no real formal structure or database schema » This makes the application of data mining to the web potentially very useful, but also difficult CSE5230 - Data Mining, 2002 Lecture 11.3 What is “web data”? Web data can be classified as follows [Dun2002]: The actual content of web pages (text, images, multimedia) Intrapage structure – the HTML or XML mark-up specifying the organization of the page content Interpage structure – the links into and out of web pages Usage data describing how the users of a web site access pages – navigation patterns User profiles – these can include demographic data obtained from a registration process, or perhaps IP addresses. It can also include information found in cookies CSE5230 - Data Mining, 2002 Lecture 11.4 A taxonomy of web mining tasks (1) Web Mining Web Content Mining Web Page Content Mining From Web Structure Mining Search Result Mining Web Usage Mining General Access Pattern Tracking Customized Usage Tracking [Dun2002], following [Zai1999]. CSE5230 - Data Mining, 2002 Lecture 11.5 A taxonomy of web mining tasks (2) Web content mining Examines the contents of web pages (text, graphics) Examines the results of web searches » Mining systems built on top of existing search engines Similar to traditional information retrieval (text categoriation, text filtering, etc.) » Often goes further than simple keyword search – e.g. may cluster similar pages Web structure mining Looks at page structure » e.g. text in <H1> tags may be more important Links between pages » e.g. pages with many incoming links may be more useful CSE5230 - Data Mining, 2002 Lecture 11.6 A taxonomy of web mining tasks (3) Web usage mining Looks at log files of web access General access tracking looks at history of pages visited Customised usage tracking may be focused on particular kinds of usage, or particular users Involves mining of sequential patterns » Can use association rule discovery, or HMMs » These patterns can be clustered to reveal users with similar access behaviour Can be used to » improve web site design » Customize presentation via collaborative filtering CSE5230 - Data Mining, 2002 Lecture 11.7 Example: targeted advertising (1) In marketing, targeting is any technique used to direct marketing or advertising effort to the portion of the population thought to be most valuable to the business, e.g. those Likely to purchase Likely to spend a lot The business wants to avoid spending money on sending advertising to people who will not respond to it In the web context, this can mean displaying an add for a web site on a different web site Can use web usage information to work out what kind of people use a site: target demographics Sell advertising to companies wanting to target that demographic CSE5230 - Data Mining, 2002 Lecture 11.8 Example: targeted advertising (2) For example, the Rugby Heaven web site (http://rugbyheaven.smh.com.au/) is today hosting advertising for: MLC life insurance Fintrack Financial Services Business Review Weekly (BRW) They appear to think that this site is likely to be popular with older people who have money! The URL for the BRW ad. is: http://campaigns.f2.com.au/event.ng/Type=click&FlightID=10928&AdID=24947&TargetID=2389&Se gments=2,13,23,31,35,77,81,88,93,94,153,855,976,993,1145,1301,1989,2320,2389,2394,2396,247 7,2534,2576,2581,2689&Targets=535,2389,40,60,1834&Values=25,31,43,48,50,60,72,81,91,100,1 10,135,150,157,233,239,366,422,605,791,804,805,806,1203,1278,1403,1432,1476,1485,1499&Ra wValues=&Redirect=http:%2F%2Fwww.brw.com.au%2Fsubscription%2Fsubscribe.asp It is clear that some sophisticated targeting is going on CSE5230 - Data Mining, 2002 Lecture 11.9 Example: personalization (1) Personalization spans the areas of web content mining and web usage mining Personalization aims to modify document contents or access patterns to better match the preferences of a particular user Personalization can involve Dynamically creating and serving web pages that are unique to an individual user Determining which pages to retrieve or link to on a userby-user basis CSE5230 - Data Mining, 2002 Lecture 11.10 Example: personalization (2) Unlike targeting, with personalization can be done for the target web page (unlike a targeted advertisement for another site) Simple example: including the name of the user in the page content Personalization techniques include Use of cookies Use of user databases Use of web usage patterns to identify similar users (for use in collaborative filtering) requires a user to log in – this part is not data mining Often CSE5230 - Data Mining, 2002 Lecture 11.11 Example: personalization (3) A classic example of personalization is the recommending to a user of a product very similar to something they have bought before (if the web site is selling something) Content that is similar to something they have used before Personalization techniques can be based on clustering, classification or even prediction With classification, the desires of a user are determined based on the class to which he/she is assigned. Classes may be predetermined by experts. With clustering, clusters of users with similar navigation or purchasing behaviour are found, and the user’s desires are determined on this basis CSE5230 - Data Mining, 2002 Lecture 11.12 Example: personalization (4) Amazon.com makes use of personalization, as we will see in an on-line example They make use of both the user’s past behaviour They also use collaborative filtering – they recommend products bought by users who have similar profiles to the current user Could use clustering, or information filtering techniques CSE5230 - Data Mining, 2002 Lecture 11.13 References [Dun2002] Margaret H. Dunham, Data Mining: Introductory and Advanced Topics, Prentice Hall, Upper Saddle River, NJ, USA, 2002, pp. 195-220. [Zai1999] Osmar R. Zaïane, Resource and Knowledge Discovery from the Internet and Multimedia Repositories, PhD Thesis, Simon Fraser University, Canada, March 1999. CSE5230 - Data Mining, 2002 Lecture 11.14