CSCI B659 Topics in Artificial Intelligence (3 CR)
Web Mining

Class Projects

1. Mining Pharmacogenomics Information using Topical Web Crawlers
2. Pagerank & Sample Size
3. Can we beat Cinematch (Netflix Recommendation System)?
4. SMILES Index for Retrieving Chemical Information
5. Social network community emergence around new digital media
6. Trend Prediction
7. Evaluating Hypertext Documents for Authenticity

1. Usage Statistics of Robots Exclusion Standard (paper in Proc. IADIS WWW/Internet 2006)
2. Mining for Blog communities
3. Directed News Analysis
4. KidsCrawler
5. Ontology Generation from Specialized Corpora
6. Using Online Social Networks in Topical Sentiment Analysis
7. Using Page History to Rank Search Results
8. Web Mining Developmental Trends in Social Networks
9. Web Topology of the Indiana University Domain (paper in Proc. IV07)
10. Web user profiling and its applicability to system security

Top

1. Multilingual news search
2. Effects of guided summarization on QA using the Web
3. Mining people connections
4. Personalized search by history context
5. Phishing Attacks Using Social Networks (see coverage in IDS, IDS, and Slashdot; paper to appear in CACM)
6. Structural evolution of Web content
7. Clustering of political opinion sites using unsupervised techniques

Top

1. Experiments with PageRank Computation
2. Clustering Weblogs using LSA and Link-based Methods
3. Sherlock News Search Engine
4. Focused Crawlers vs Accelerated Focused Crawlers
5. Domain-Based PageRank Personalization (paper presented at WebKDD 2004!)

Top

• Wikipedia's publicly-available data set provides a wealth of information about the structure and evolution of a dynamic socially-edited forum. Possible research includes automated fact-checking and user authority measures (date of registration, number of posts, average size and longevity of posts, etc). Another avenue of investigation would be correlating IP addresses / usernames with changes to specific hot-topic issues. An automated method to detect suspicious editing / revisions could aid in the identification of biased or self-serving modifications leading to the identification of the offending individuals / organnizations. (Submitted by Mike Conover, based on discussion in class and prior conversations with Virgil Griffith)
• Create a meta-search engine that finds potentially "embarrassing" personal material (photos, videos, text, etc) by mining various sources such as search engines, social network sites, photo and video sites, etc. This could play an educational service by highlighting the dangers of posting private information on the Web. (Based on an idea by Mike Conover)
• Develop a desktop application (eg for the Google Desktop, the Yahoo Desktop, Windows Desktop, or Apple Dashboard Widget) that implements some (simplified/extended) version of the HITS algorithm, discussed in class. This would be a client-based solution for the query-time analysis. (Inspired by Michel Salim)
• The economics of Google Ads are inducing (creating incentives for) a "pollution" of information on the Web. People create fake "original content" with popular query terms to attract traffic and make a profit through advertising. This is being done both manually (by underpaid hired writers) and with automatic text generation scripts. Can we devise techniques to clean the Web from such pollution? (Submitted by Fil)
• Try to come up with a list of top-X sites frequented by spam harvesters. For example I created a gmail account to submit a script to CPAN, which posts the email of authors on their site. That account has quickly become a honeypot with thousands of spam messages. So clearly spammers harvest emails from CPAN. What other sites? What are the worst? You could write a crawler that automatically posts email addresses to sites it encounters (message boards, etc) and then monitors which sites generate the most spam. (Submitted by Fil)
• A machine learning method to classify an arbitrary Web page as blog or not blog, for crawling purposes. (Submitted by Alex Breuer)
• A text mining algorithm to find a huge set of triples (email address, name, address) from crawls of personal websites, blogs, bios, etc. and cross-reference with structured databases such as ip-to-zip converter, phone book, and other online resources. This problem is one of the holy grails in phishing right now, and nobody has addressed how it can be done or protected against. (Submitted by Markus Jakobsson, who could act as external co-supervisor)

Top

1. Why? (Motivate your idea; is it interesting, important, relevant?)
2. What? (Exactly what do you propose?)
3. How? (State your hypothesis and evaluation procedure)
4. When? (You need a realistic timetable and deliverable; is it doable?)

   Top