Meeting Minutes

Date: 06/23/05

Place: SB 232C

From 10:30 am to 1:00 pm

Attenders:         Dr. Wai Gen Yee,

                        Dongmei Jia,

                        Linh Thai Nguyen.

1/ The “query masking” paper (Improving Search Performance in P2P File Sharing Systems by Query Masking) is submitted to ICDE 2006. The main idea of query masking is to use sub-queries to increase the probability of finding desired data objects. In this work, we have tried different combinations of metadata distributing techniques, ranking functions and client-centric masking techniques, to see whether there is a good combination. Experimental results show that query masking technique can actually improve query performance in P2P systems. In the future, we will investigate how server-centric masking techniques can affect query performance.

2/ Dr. Yee’s discover on experiments’ results:

• The simulator is slow. In order to run extensive experiments on big dataset, we need some way to optimize the code.
• The dataset used is the small dataset (part 4). The query length and the degree of masking are both varied.
• Effect of query length to ranking performance: For short queries, there are more results. However, there are more noise in the results, and short queries contain little information, hence it is difficult to rank the results correctly.
• Effect of query length and degree of masking to ranking performance: when query length is 6 and the number of unmasked terms is 1, ranking based on term frequency is the best; when query length is 1 and number of unmasked terms is 1, ranking based on group size is the best. This pattern is quite strange (why?).
• For part 4 data, only 1 query out of 10 works.
• If we increase the number of peers to 5000 (instead of 1000), ranking performance increases significantly (but the simulator runs very slow).

3/ Missing parts:

• We need intelligent arguments that can formally explain why some ranking functions, in some specific situation, outperform other ranking functions.
• Which ranking function works best for uniform distribution of data objects.
• Which is better: client-centric query masking or server-centric query masking. In addition, which query masking technique to be used.

4/ Dongmei’s presentation: SETS: Search Enhanced by Topic Segmentation.

• One site can only belong to one topic segment.
• The system needs one centralized site to serve as administrative site. This site receives summary vectors from other sites in the system and classifies these vectors using k-Mean algorithm, then it distributes the centroid vector of each topic segments to every sites in the system.
• The global and local query routing algorithms are not discussed in detailed.
• The proposed techniques and structures are not novel (represent documents by keyword vectors, classify using k-Mean...) but this work is complete.

5/ Dr. Yee’s recommendations:

• Every reviewed paper need to be summaried.
• Students should try to work independently, write stuff for other to read and comment.
• Be proactive in research.

6/ Assignments for next week:

• Dongmei: continue to work on the prototype.
• Linh: comparison between client-centric masking and server-centric masking; find the best ranking function; find the best masking function.

7/ Miscellaneous:

• one bug of the simulator found and fixed: the parameter related to the computation of inverse ranking score is wrongly set to 1.