Project Guidelines (please read carefully):

• 2-3 people per project.
• Deadlines (no late submissions accepted):
  • 1-2 page project proposal (include workplan) due Feb 20,
  • 1st project submission on March 20,
  • Final submission around May 1.
• Choosing your project:
  • Projects can be any of the suggested projects below.
  • You are welcome to choose  projects outside this list; BUT the project proposal must be approved.
  • Projects must have direct relevance to the material covered in this course. The project work that you want to be evaluated on MUST NOT overlap with similar work/project for another course.
  • The 1-2 page proposal to be submitted on Feb 20 must have a clear goal and  workplan.
    • Acceptable projects will consist of:
    • New algorithm/method/technique for handling an old/new problem,
    • Known algorithm applied to a new domain,
    • Comprehensive comparison of existing algorithms/techniques on a particular application/data set.
    • Implementation of some web-mining technique.
    • NOTE: Simply a literature survey of existing methods is NOT acceptable.

• After you select a project, you should create a project web page and then send the TA the URL.

Project Suggestions:

• 1. WEB-USAGE MINING:
  Best if you have access to logs of a frequently visited web-site. I can provide logs of www.sulekha.com under NDA. Determine different types of visitors based on their "footprints". Find correlations between different parts of the website based on similar visitations. You can cluster visitors based on their URL traversal patterns and then model/characterize each cluster. Usage studies also  suggest ways of rearranging/adapting the website, as well as dynamically customizing it for each visitor. Web-log analysis can also be used to improve the performance of a website by helping caching policies, load balancing, etc.

  Additional References:

  Yan T.W., Jacobsen M., Garcia-Molina H., Dayal U., "From User Access Patterns to Dynamic Hypertext Linking", Fifth International World Wide Web Conference, Paris, 1996. http://www5conf.inria.fr/fich_html/papers/P8/Overview.html

  Shahabi, C., Zarkesh A., Adibi J., Shah V., "Knowledge Discovery From Users Web-Page Navigation", Proc. Seventh IEEE Intl Workshop on Research Issues in Data Engineering(RIDE) 1997. http://dimlab.usc.edu/publications.asp

  Schechter S., Krishnan M., Smith M., "Using path profiles to predict HTTP requests The Seventh International World Wide Web Conference", Brisbane, Australia http://www.eecs.harvard.edu/~stuart/

  Nasraoui O., Frigui H., Joshi A., Krishnapuram R., "Mining Web Access Logs Using Relational Competitive Fuzzy Clustering", Proc. Eight International Fuzzy Systems Association World Congress - IFSA 99, August 1999. http://www.cs.umbc.edu/~ajoshi/web-mine/publications.html

  Workshops:
  KDD'99 Workshop on Web Usage Analysis http://www.acm.org/sigkdd/proceedings/webkdd99/toconline.htm

  KDD'00 Workshop on Web Mining for E-Commerce http://www.wiwi.hu-berlin.de/~myra/WEBKDD2000/schedule.html

  User Modeling'99:
  http://www.cs.usask.ca/UM99/papers.shtml

  Also see the forthcoming Workshop on Web Mining for additional topics of interest.
  http://www.lans.ece.utexas.edu/workshop_index.htm

  Software:
  http://www.hypernews.org/HyperNews/get/www/log-analyzers.html packages on log analysis.

  WUM: Web Utilization Miner
  http://wum.wiwi.hu-berlin.de/ (please contact Dr. Ghosh if you want to use this package).
   

• 2. PERSONALIZATION:
  The August issue of Comm. ACM has 26 articles on personalization.

  Possible projects include:

  A. Collaborative Filtering:
  How can the importance of different items (e.g. genre vs. director vs. length vs. actor for movies) be customized for different users? Can use EachMovie data set. How can collaborative filtering approach  be more scalable to handle larger numbers of people and items? How can content and/or usage (web logs) be used to augment collaborative filtering?

  See Collaborative Filtering Mailing List Archive By Thread:
  http://info.berkeley.edu/resources/mailing-lists/collab/
  Also the Syskill Webert data from UCI for rating web pages.

  B. Infomediaries:
  These agents can modify information streams at the client end, server end, or somewhere in-between. They can be used to customize, filter, annotate, aggregate or transcode in a personalized way. See http://wwwcssrv.almaden.ibm.com/wbi/intermediaries.html and http://www.almaden.ibm.com/cs/wbi/papers/chi97/wbipaper.html Also a WBI plugin is available at: http://www.almaden.ibm.com/cs/wbi/ .
   

  C. Personalized Web Searching/Browsing/Filtering
  Develop agents that work on behalf of the individual user, a personal e-valet, if you will. They can identify and gather interesting news articles, extract info about specific products, filter emails etc etc.

  Additional References:

  Clexandros Moukas and Pattie Maes Amalthaea: "An Evolving Multi-Agent Information Filtering and Discovery System for the WWW", http://www.media.mit.edu/~moux/papers/jaamas98.ps

  Liren Chen and Katya Sycara, "Webmate : A personal agent for browsing and searching", In Proc of 2nd International Conference on Autonomous Agents, 1998.
  http://www.cs.cmu.edu/~softagents/webmate/aa98-webmate.ps

  J. Allan, "Incremental relevance feedback for information filtering". In Proc. ACM SIGIR Conf., Zurich, Switzerland, August 1996.
  http://www-ciir.cs.umass.edu/~allan/Papers/sigir96.ps
   

• 3. XML:
Considered the NO.1 Information Technology in 2000, and a core technology for implementing B2B exchanges, next generation data repositories as well as search and querying engines that can exploit XML tags for better accuracy.

See www.xml.org and www.xml.com for starters. Also the articles: Top Tech Trends for 2001
http://www.infoworld.com/articles/op/xml/01/01/08/010108opvizard.xml and XML enlivens e-business transactions: http://www.infoworld.com/supplements/2000toy/010129tcxml.html.

Possible projects include:
- develop an architecture for searching/querying a collection of XML documents
- determine ways of finding similarity between two XML documents, that take into account both the words and the tags.
- ways to (semi)-automatically converting an XML document into another one with a different DTD: useful for information exchange, say between two companies with different conventions for marking up documents.
- extracting meta-data from web pages, and using them to augment a web content mining application.
- develop XML tools to integrate both content and usage data and domain knowledge such as ontologies or word vocabularies, to model users in an e-commerce sites.

We have purchased TREC 4 and TREC 5 CDs for the class. These have many large document collections in SGML, a superset of XML, and thus can possibly be adapted for your XML project.
 
 

• 4. Document similarity for plagiarism detection:
  Design and evaluate new algorithm(s) for scientific literature (e.g. use the  CiteSeer database) or other databases.

  Andrei Broder, Steve Glassman, Mark Manasse, Geoffrey Zweig, Syntactic Clustering of the Web, Sixth International World Wide Web Conference, 391--404, 1997. SRC Technical Note, 1997-015.
  http://gatekeeper.dec.com/pub/DEC/SRC/technical-notes/SRC-1997-015-html

  N. Shivakumar, H. Garcia-Molina. SCAM: A Copy Detection Mechanism for Digital Documents. Proceedings of the 2nd International Conference on Theory and Practice of Digital Libraries, Austin, Texas, 1995. Available from http://www-db.stanford.edu/~shiva/Pubs/scam.ps

  Misc Stuff:
   

  • Einat Amitay's WEB IR & IE resources. http://www.ics.mq.edu.au/~einat/web_ir/
  • GISTNIC. This is a site where one can access a large number of databases related to statistics from India. Need to register; slow, but fascinating.
    http://gist.ap.nic.in/

 
• 5. Information Retrieval (TREC/Web):
  • TREC queries. (major skills:graph theory, some programming).
The TREC (http://trec.nist.gov/) and SMART (ftp://ftp.cs.cornell.edu/pub/smart/) data sets contain large document collections, various queries and the corresponding relevance judgements (which are a hand-assembled list that indexes the documents relevant to each query). The goal is to have a query retrieval system that does a "perfect" job on all queries (100% precision and 100% recall). Of course this is not possible, so the realistic goal is to have high precision and recall.

The TREC and SMART queries are of different flavors. Some queries are short while others have a more verbose description. Many queries in TREC are such that the documents that are judged to be relevant to these queries do not contain any of the words in the query.

The first phase of this project would be to explore how "far" relevant documents are from the words contained in a query. For example, are all relevant documents two "hops" from the query, that is, do all relevant documents either (i) contain words in the query or (ii) contain words that are contained in documents that contain words in the query. Such questions can be answered by looking at a graph between the documents and words.

The second part of the project would be to explore algorithms (graph algorithms?) that would exploit the information discovered in the above phase to get better precision and recall.

•  Cluster web search results. (major skills:programming).
  Some search engines (see Northern Light) cluster their search results into various categories. Typically they use words either in the text or in the HTML tags to cluster their results. This project would cluster web search results (say those returned by Google) by augment words with links in a principled manner.

  See "Clustering Hypertext with Applications to Web Searching" (http://ww.lans.ece.utexas.edu/course/ee380l/2001sp/readinglist/toric.ps.gz) for one approach.

• Email classification --- Investigate Linear Discriminat Analysis and computationally superior variants for text classification (mathematical,programming).
  Various text classification algorithms exist, for example, the simple Naive Bayes to more advanced probabilistic algorithms (see http://www.cs.cmu.edu/~webkb/). This project would explore the feasibility of using Fisher's linear discriminats for classifying small document collections, and then explore computationally efficient variants.

  Classifying email automatically into different folders is a challenging and worthwhile problem. See http://www.ai.mit.edu/~ychang/learningmail.pdf  for a comparison of Naive Bayes and Fisher's Linear Discriminant. Source code is available too! Also see last year's CS395T project: http://www.cs.utexas.edu/users/zhux2/datamining/

  The following paper uses a so-called Fisher index:
  http://www.cs.berkeley.edu/~soumen/doc/www99focus/html/

• Estimation of number of clusters. (in clustering document collections) based on Gap statistic, and other measures. (major skills: mathematical,statistical,some programming). See http://www-stat.stanford.edu/~tibs/ftp/gap.ps
• Given a web page, find related web pages. Compare results with   Google, Netscape (skills:programming). I feel this problem is not adequately solved in Google for less "authoritative" pages. See:
  http://www.research.digital.com/SRC/publications/copyright/elsevier99.html

 
• 6. Bioinformatics:
  •  Clustering DNA Microarray Data (mathematical, programming).
    For experiments(support vector machines, decision trees, linear discriminants) on classifying gene expression data, see
    http://www.cs.columbia.edu/~bgrundy/papers/genex.pdf
    http://www-stat.stanford.edu/~hastie/Papers/shave.ps http://www-stat.stanford.edu/~tibs/ftp/sjcgs.ps
  
   
• 7. Face Recognition: (major skills: mathematical, MATLAB programming)
  This project would explore different methods for the face recognition, i.e., given a database of face images and a query face, what face is most similar to the query face? Some of the methods used for this problem are similar to the methods used in text classification. Some successful methods are based on eigenfaces (using SVD) and Fisherfaces (using Fisher's linear discriminant).

  Last year's class project (http://www.cs.utexas.edu/users/subbiah1/mining/) did an excellent job in building prototype software in MATLAB for this problem. The scope of this year's project would be to use this software and extend it to use computationally different (and more efficient) methods.

  Resources

  Eigenfaces and Face Recognition at the MIT Media Lab.
  Papers and Faces Database by Larry Sirovich.
  Eigenfaces vs. Fisherfaces: Recognition Using Class Specific Linear Projection by Peter Belhumeur and Jo Hespanha and David Kriegman, July 1997.
  Class Project that contains a report and some code too.
  Face Database at AT&T Labs, Cambridge.
  Papers about fast computation of eigenvectors of images (1 and 2).
  Additional resources and tools for web mining provided by Noppadon.