Information Retrieval Techniques for PeertoPeer Networks

1
Information Retrieval Techniques for Peer-to-Peer
Networks

• Demetrios Zeinalipour-Yazti
• Dimitrios Gunopulos
• Vana Kalogeraki
• Computing in Science engineering IEEE 2004

2
Outline

• Introduction
• Peer-to-Peer network Information Retrieval
  techniques
• Experiments
• Conclusion

3
Introduction (cont.)

• Powerful PCs shifts the client-server model
  toward a peer-to-peer(P2P) toplogy.
• Peers collaborate in an ad hoc manner and share
  information in large-scale distributed
  environments without centralized coordination.
• alternative model for the architecture of
  centralized Web crawlers.( JXTA)

4
Introduction (cont.)

• P2P Information-retrieval (IR) environment we
  discuss here assumes the each peer has a database
  of documents that if shares in the network.
• A node send Query messages contain sets of
  keywords to its peers.
• The peer replies message contains pointer to the
  matching documents if the evaluation is
  successful.

5
Introduction

• Traditional IR algorithms do not apply directly
  to P2P system because there is no central
  repository.
• Improving search capabilities is an important
  step in making P2P systems applicable to a wide
  set of applications beyond simple object storage.

6
P2P Network IR Techniques (cont.)

• Breadth-First Search (BFS)
• used in Gnutella
• Node q generates Query message to all its
  neighbors( peer).
• peer p forward the query to all its peer except
  that the sender and then searches its local
  repository.
• QueryHit messages travel along the same paths
  that carried the query messages.

7
BFS (cont.)
8
BFS

• Too many messages
• network utilization
• processing resources
• To solve by associating each query with
  time-to-live (TTL) parameter.

9
Random Breadth-First-Search (RBFS) (cont.)

• peer q forwards search message to only a fraction
  of its peers, selected at random.
• does not require global knowledge.
• the query might not reach some large network
  segments because nodes dont understand that a
  particular link could take the query to that
  large segment.

10
RBFS
11
Intelligent Search Mechanism (ISM) (cont.)

• peer propagates the query message to those peers
  more likely to reply.
• profile mechanism profile of neighboring peers.
• relevance rank uses profiles to select the
  relevant neighbors.
• once profile repository is full, the node employs
  a least recently used (LRU) policy to keep the
  most recent queries.

12
ISM (cont.)
13
ISM (cont.)

• RRpl(Pi, q) Qsim(qj, q)aS(Pi, qj).
• exP1 replied to queries q1 and q2 with
  similarities Qsim(q1, q) 0.5and Qsim(q2, q) 0.1
  to the query q.
• for P2, Qsim(q3, q) 0.4, Qsim(q4, q) 0.3
• a10, 0.510 0.110 gt 0.410 0.310
• a 1, 0.5 0.1 lt 0.4 0.3
• a 0, 1 1 1 1 (gtRES)

14
ISM

• potential disadvantage
• always query the same neighbors
• newly added neighbors are not given the
  opportunity to be explored
• pick a small random subset of peers and add it to
  the set of most relevant peers for each query.

15
gtRES (cont.)

• forwards a query to a subset of peers based on
  some aggregated statistics.
• a peer q forwards a search message to k peers,
  which had returned the most results for the last
  m queries

16
gtRES
17
P2P Network IR Techniques (cont.)

• Random-Walker Searches
• each node randomly forwards a query message,
  called a walker, to one of its peers.
• messages increase linearly
• adaptive probabilistic search (APS) feedback
  from previous search ( rather than randomly
  forward)
• Local Routing indices
• a node knows which peers lead to the desirable
  documents, but it doesnt know the exact path to
  those documents.
• push update( each node sends information to its
  peers)
• (It completes the ISM approach)

18
P2P Network IR Techniques

• randomized gossiping
• bad in scalability
• Centralized Approaches
• Central repository for all peers index of shared
  documents.
• Searching Object Identifiers
• use object identifiers (hashcode on the name of
  file) rather then keywords
• but cant capture the relevance of the documents.
• Distributed IR
• assume that the querying party has some
  statistical knowledge about each databases
  contents ? need global view of the system.

19
PeerWare Infrastructure and Experiments (cont.)

• Evaluation metrics
• To implement only algorithms that requires local
  knowledge BFS, RBFS, gtRES, and ISM
• recall rate the fraction of documents each of
  the search mechanisms retrieves
• technique efficiency of messages needed to
  find the results

20
Experiment( cont.)
TTL 4 ( BFS, ISM) ISM achieved alomost a 90
recall rate while using only 38 of the messages
BFS required. Started out with a low recall rate
gtRES ISM initially choose their neighbors.
21
Experiment

• TTL 5
• ISM uses only 57 of message but discovers almost
  all documents.
• (ISM, RBFS, and gtRES had the same query response
  time (QRT) 250ms)

22
Conclusion

• P2P computing model distributes the incurred
  network load on all interested parties
• Efficient search and retrieval still remains a
  challenge to be explored.
• Another challenge is to create overlay topologies
  in which close-by nodes have semantically
  related documents and interests.
• To test our middleware infrastructure publicly
  available and test it in a larger and more
  realistic environment.