Group Meeting Minutes

Date: July 7^th, 2005

Place: SB232C

Attendees: Wai Gen, Linh, Dongmei

1/ KDD conference in Chicago: Dongmei, Linh and some other IIT students are volunteers:

• Make personal webpage
• Make business card
• Prepare a 30 seconds speech about what you (we) are doing

2/ Oncoming deadline:

• SAC:             Sep 3, Dijon-France
• EDBT:            Sep 23, Munich
• ICDCS:          Nov
• SIGMOD:      Nov, Chicago-USA
• WWW:           Nov
• SIGIR:           Jan, Seattle-USA
• VLDB:           Feb
• MDM:            Oct, Japan
• ECIR:             Oct, Portugal

Linh: work on “query masking” paper for EDBT.

Dongmei: work on the prototype for SIGMOD.

3/ Wai Gen will talk about the intersection of P2P technologies and mobile computing at UIC on July 20.

4/ Wai Gen presented experiments’ results of query masking technique:

• Based on the results, ranking by arrival is as effective as other ranking techniques. Something must be wrong ! (bug in simulator, wrong parameter settings ?)
• Using query masking, tf ranking function is the best. Not using query masking, gsize ranking function is best. Why?
• Can we devise better metric, that combine both performance and cost.

5/ Dongmei’s presentation on the paper “An architecture for information retrieval over semi-collaborating peer-to-peer networks”. I. A. Klampanos and J. M. Jose. In Proceedings of the 2004 ACM Symposium on Applied Computing, volume 2, pages 1078–1083, Nicosia, Cyprus, March 14–17 2004:

• Two important goals of a p2p Information Retrieval system might be efficiently identifying of relevant information sources and efficiently routing queries to those sources. The paper presented a cluster-based architecture for IR over p2p networks, targeting to these two goals.
• The term “semi-collaborating” means nodes in p2p network cooperate in order to perform information retrieval, but they don’t have to share any detailed information, nor do they have to be consistent with respect to the IR systems they use.
• Each node in the network may choose to implement one or more of the following services: Client service, that provides end-user interface; Information Provider service, that shares local documents to other nodes in the network; Hub service, that provides message routing service to other nodes, and Fusion service, that handles the fusion of retrieved results, on behalf of querying node.
• Content-Aware Clustering: two clustering stages are applied: in-peer document clustering and peers clustering.
• In-peer document clustering: each document is represented by its term frequency vector. Initially, each cluster contains a single document. Cosine-similarity formular is used to measure distance between any two vectors. At each step, two clusters closest to each other are merged to form a new cluster. The clustering process stops when there is no single-document cluster left.
• Peer clustering: peers are clustered into Content-Aware Groups (CAGs), based on their internal clusters, the variance of their clusters, and the participation level of their clusters. One peer may belong to many CAGs. A peer is merged with a CAG if the different between its centroi vector and the one of the CAG is smaller than a threshold value, its variance is smaller than the CAG variance, and its participation level is greater than the one of the CAG.
• Query routing:

§ Each hub-enabled peer must maintain the descriptors of all CAGs in the network.

§ Upon receiving a query, a hub-enabled peer will score and rank all the CAGs based on their distance to the query, their variance, and their participation level. The query will be sent to the top n CAGs.

§ Within each CAG, peers get ranked following the same procedure, and the query is routed to the top m peers.

· Combination of results: using Dempster-Shafer (D-S) theory of evidence combination (?). Each peer is assigned an un-trust coefficient based on its ranking score, and each result retrieved from a peer is scored based on its ranking score and the total number of results. The product of the peer’s un-trust coefficient and the result’s ranking score is used as global ranking score for that result.

· Evaluation: experiments were conducted to prove that the technique can actually work efficiently.

6/ Dongmei presented experiments’ results based on the uniform distribution of documents: need to modify the simulator and rerun experiments.

7/ Assigments:

Linh: critique the ICDE paper (query masking paper); finish experiments; continue to work on the query masking technique

Dongmei: modify the simulator and rerun experiment

8/ Paper for next meeting:  Making Peer-to-Peer Keyword Searching Feasible Using Multi-level Partitioning. Shuming Shi et al. Tshinghua Univ. Third International Workshop on Peer to Peer Systems 2004 (IPTPS’04).

http://iptps04.cs.ucsd.edu/papers/shi-keysearch.pdf