Group Meeting Minutes

Date: July 21^st, 2005

Place: SB232C

Attendees: Wai Gen, Linh, Dongmei

1/ Dongmei presented experiment results for the uniform distribution of documents, using additional ranking functions.

2/ Dongmei’s presentation on the paper “Content-Based Community Formation in Hybrid Peer-to-Peer Networks”:

• The authors model digital libraries as leaf nodes and regional directory services as ultrapeers in the hybrid peer-to-peer architectures. They used community formation method that allows nodes in the networks gradually change their set of neighbors towards a topology that improves IR efficiency.
• In order to enable users whose content interests and IR patterns are mutually beneficial to collocate in the same or nearby community, the authors proposed a valuation function for estimating the value of establishing a connection between a pair of peers, and a community formation algorithm.
• Valuation function: a query model of a peer is composed based on the peer’s past queries. In addition, a content model of a peer is also built. The K-L divergence method is used to measure the similarity between past queries model of one peer and content model of another peer.
• Formation algorithm: given a set of foreign peers, a peer will evaluate the benefit of establishing a connection with each new peer using the valuation function. If the benefit of connecting to any new peer is greater than the benefit of an existed peer in the neighbor set, that peer will be replaced by the new peer. The set of foreign peers may be retrieved through a host cacher or through PONG messages.
• Other extensions: pruning method and intelligence host cache server are proposed as additional techniques to enhance the IR efficiency.

3/ Linh reported a bug in the simulator that makes the performance of all ranking functions the same as performance of ranking by arrival, for the case of server masking. The solution is to change the value of one parameter.

4/ Linh reported cutting half of the average number of responses per query while maintaining approximately same performance for server masking, by early stop masking when there are some responses. The idea come from query relaxation techniques, that deal with the “failed-queries” (queries that return nothing) in database. The question is how this “early stop masking” technique affects different ranking functions?

5/ Assigments:

Dongmei:

• Create a parameter for the uniform distribution of documents and post the source code on IR machine.
• Explain the experiments’ results (may come out with hypothesises and new design of experiments).
• Big goal: building Limewire and the Journal paper.
• Remember the conference deadline.

Linh:

• Try different masking algorithms (combination of existing ones).
• Try backup masking algorithms
• Explore query relaxation.
• Remember the conference deadline.

6/ Paper for next meeting on Aug 4:  “Sloppy Hashing and Self-Organizing Clusters”

http://iptps03.cs.berkeley.edu/final-papers/coral.pdf