CSE 592 Final Project Synopsis


·        The Audio Genre Identification System (AGIS) -- Eric Fisk and Scott Baird

AGIS is an intelligent classification system that is applied to a sampled audio stream.  Using AGIS it should be possible to analyze samples from an audio stream and identify the strength of characteristics such as "talk", "country music", and "classical music".  The audio samples will be filtered down into a set of numbers, which will be fed to a neural network. Once the neural net is trained, it will hopefully be able to classify further audio samples that are given to it.

·        PathGuru - Using Hidden Markov Models for Web Path Prediction -- Howell Pan and Keith Folsom

Web server access logs contain a wealth of information about the pages of a web site that is being visited.  The data is in the form of a single time-stamped line per page hit, with the most interesting information in the line being the time of the hit, the source IP, and the web page on the server.  Our objective was to use an AI tool to learn the most common paths taken by visitors through a web site.  The AI tool we used is a custom implementation of the Baum-Welch or forward-backward algorithm applied to Hidden Markov Models for each path-starting page we found in the access logs.  We applied this to over a month's worth of access logs from the Pacific Lutheran University web site at http://www.plu.edu.  We've designed a CGI interface that acts as a wrapper for web pages on the PLU site and appends a "footer" block to each page that shows the top 3 most-traveled paths, given
the current page, as previously learned by our AI tool.  We are evaluating the success of our approach by comparing our predictions against the actual paths taken as recorded in a second set of access log records. URL:   http://www.plu.edu/htbin/nph-pathguru.cgi

    

·        Log Analysis for Resource Request Prediction -- Shaula Massena & Diane Moore

The objective of this project is to incorporate AI tools into the analysis of web log data in order to make predictions for a user's future resource requests from the server.  By breaking down web log data into instances of user sessions with all of that user's requests for the most popular resources, we can employ machine learning in determining relationships among the requests.  We are exploring various WEKA machine learning implementations to explore what gives us better results.  We will evaluate the utility of the tools by testing them on a separate set of data from the same web log with user instances and their requests.  If the prediction that X resource will be requested given previous resources is in the set of a test instance user's future requests, we will consider the prediction successful.

 

·        Intelligent Web Query Filter -- Matt Lyons and Jeff Orkin

We plan to make an intelligent searching agent to use as a supplement to existing web site catalogs/search engines (Altavista, Google, etc.)  The basic idea is that the user can start off with a relatively simple query, and the refinement can be assisted by the agent instead of trying to guess the "right" key words.  From the training data provided, the agent will be able to suggest future preferred URLs using a classification algorithm.

 

·        A web page prediction system for predicting user's future document requests based on the past ones -- Abha Bhatia and Sumit Chawla

Our aim is to develop a web-page prediction system for predicting user's future document requests. We plan on organizing documents in a fashion
similar to the http://www.cs.indiana.edu web server - i.e. we will assign a unique identifier to each document and organize documents in hierarchical
categories.  We intend to manually categorize and organize documents on our web server. Once documents have been organized, we plan to use Decision Tree Learning algorithm to do the actual prediction.  We have implemented our own decision tree algorithm.

 

·        CD WebLearner -- Rob Aldinger, Paul McDowell
 
We were attempting to create a system for learning user preferences for CDs.  The site allows the user to store profiles of past intersts and CD
purchases an then tries to guess what other types of CDs they might like. The web site uses NaiveBayes to learn a set of keywords associated with a
CD, and based past choices of interest\disinterest, makes recommendations.

 

·        Adaptive Web Sites - Sites that can adapt to user's interests --  Magesh Narayan, Prakash Channagiri, Robert Fleming

 

 The problem being addressed in this project is personalization - how do we build a website that can adapt to a user's interests and
preferences. Currently sites like CNN use a very rudimentary method wherein the user explicitly specifies his/her interests and tastes. Our goal is to
build a system that automatically learns a user's interests based on his/her historical viewing patterns. We are using machine learning algorithms to
build a website that would initially use the user's history files as a training set. Every webpage visited from then on becomes another addition to
the training set. This technique does not take into account a user's short term interests but is quite effective towards long term interests. Output of
the experiment can be presented in different ways - a web page suggesting other current articles of interest, a web page suggesting other articles
that he has already viewed in the past that might be of interest, or the web site itself can be configured to bring up an article based on what he/she
likes most. Evaluation techniques would involve feeding the system (i.e. website) history folders from different users and observing how closely the
output matches the user's interests.

 

·        Adaptive/Intelligent User Interfaces for Composite Multimedia Systems  -- Greg Taleck, Kenneth Ma, Leo Shum
 
Description: As we witness the convergence of many multimedia delivery systems (i.e. Cable, Internet, Broadcast) content will increase
tremendously.  How can you leverage AI tools to simplify interfaces and make a user's experience more gratifying? (finding acceptable content for
a user?) We explore the use of belief networks, machine and reinforcement learning applied to this domain.