AIRG Topics - Spring 2002

Our group meets on Thursdays at 11:00 a.m., FL 246.

Dates and topics for this semester are as follows:

Jan 17

AIRG/DKBRG Organizational Meeting (Coordinators: DCB & EAR)

Jan 24

Aparna Varde,

"The SWECCA Approach"

Jan 31

Adina Florea

Postponed: Faculty Candidate Colloquium

Feb 7

Zhuo Chen,

"Two-level A-Design"

Feb 14

Advising Appointment Day: No meeting

Wed Feb 20, 1:00-2:00, FL 320

Janet Burge

"Software Engineering Using design RATionale (SEURAT)"

PhD Proposal Defense

Feb 21

Dave Brown

"Features"

Feb 28

No meeting: Faculty Candidate Colloquium

Mar 7

Term break: No meeting

Mar 14

No meeting: Faculty Candidate Colloquium

Mar 21

Geraldine Rosario

"What is an Adaptive System?"

Mar 28

Lily Chen

title tba

Apr 4

Mukesh Mulchandani

"Updating XML Views of Relational Data"

Apr 11

Jing Yang

title tba

PhD Qualifying Examination Directed Research Presentation

Apr 18

Wendy Kogel

"Mixture of Experts for Neural Network Recommender Systems"

Master's Thesis Presentation

Advisors: Prof. Carolina Ruiz and Prof. Sergio Alvarez (Boston College)

Abstract:

In this thesis we propose and investigate a new architecture, mixture of experts, for artificial neural networks (ANNs). This architecture greatly reduces the number of connections in the neural network and consequently the amount of time needed to train the net, by exploiting the natural divisions in the input data. We evaluate this new architecture by comparing it against the traditional fully connected ANN architecture in the domain of recommender systems. Our results show that the mixture of experts architecture drastically decreases the amount of training time, which is significant because of the real-time constraint of recommender systems, while preserving the accuracy of the recommendations of the fully connected configuration.

Apr 25

Maged El Sayed

"Incremental prediction for users' access patterns"

PhD Qualifying Examination Directed Research Presentation

Mining frequent sequences is an important component in predicting users' access behavior in web applications, and hence pre-fetching appropriate pages. The process of mining for frequent patterns typically involves handling huge amounts of input sequence data. Many solutions to this problem used Apriori-like candidate generation. In this talk I will be introducing a design for a system that addresses the problem of discovering and storing frequent patterns. This system takes into consideration minimizing required access to the input database, which is needed to discover frequent patterns, and the need to represent the frequent patterns in a compressed format. The system can adapt to changes in users behavior over time, and allows the user to change the system parameters (e.g. min support) incrementally without requiring full recomputation.