Searching Social Networks

1
Searching Social Networks

• Bin Yu, Munindar P. Singh
• Presentation by
• Bilge Baskeles
• 2004721025

2
Outline

• Introduction
• Referral Systems
• Experimental Results
• A Prototype System
• Related Work
• Conclusion

3
Introduction

• Finding relevant information which is not indexed
  or cataloged depends on finding right people to
  ask our questions in our social network who has
  the desired information and expertise.
• Listing social relationships in a repository is
  not feasible or possible because of privacy etc.
• Distributed search through referrals is more
  preferable.

4
Introduction

• A referral system is a multiagent system whose
  member agents are capable of giving, following
  and evaluating referrals.
• The agents cooperate by giving and taking
  referrals to help its users exchange information.
• Every agent learns the users preferences and
  interests, keeps a view of its users
  acquaintances and maintains incoming queries by
  issuing referrals or answering them.

5
Introduction

• Two major approaches exist for referral systems
• MINDS
• emphasizes learning heuristics for refrerral
  generation
• ReferralWeb
• Focuses on how to bootstrap the referral system

6
Introduction

• In the paper, the dynamics of social networks and
  the effects of the dynamics on information flow
  is emphasized.
• It is considered how to efficiently search social
  networks with the help of agents with local
  knowledge and how to control search by adaptively
  choosing referrals.

7
Introduction

• Multiagent systems involving specialized agents
  apply in two main scenarios.
• Knowledge Management efficient and effective
  knowledge management system in which agents
  maintain its users information and can search its
  social network.
• Trust and Reputation Management Referrals
  enable agents to share information so that
  untrustworthy parties can be eliminated. Methods
  exist such as combining evidence from witnesses.

8
Referral Systems

• Each agent helps its user maintain his social
  network.
• A query from the user is seen by the agent. The
  agent sends the query to the agents of users
  (contacts) who are eligible.
• When an agent recieves a query, it decides
  whether it suits its user and sends it to the
  user if suitable, otherwise responds with
  referrals to others or discard the query.

9
Referral Systems

• A query specifies what information is being
  sought. A response includes an answer or a
  referral.
• An agent answers only if it is reasonably
  confident of its expertise matching the query.
• A referral depends on the query and on the
  agents models of others and given by agent if
  the agent has confidence of relevance of the
  referred.

10
Referral Systems

• Each agent keeps models of its acquaintances
  closest of which are called neighbors.
• An acquaintance is added by recieving an unknown
  referral as a response to a query submitted to
  acquaintances or by recieveing a query from an
  unknown agent.
• Limited number of neighbors for each agent.
  Promotions occur among acquaintances.

11
Referral Systems

• When a referral is recieved, it is integrated
  into agents models. Based on models, an agent
  decides to follow a referral.
• When the agent recieves an answer, it is used for
  evaluating the expertise of the answering agent
  and agents who referred this agent (chain
  mechanism).

12
Referral Systems

• Each agent has two kinds of models
• A profile for its user
• An acquaintance model for each acquaintance
• VSM (Vector Space Model) is used to capture these
  models.
• The vectors in VSM includes term vectors
  (different areas of expertise) indicating a
  weight for each term.
• The similarity between two term vectors (query
  and expertise) is calculated using cosine of the
  angle between them. (Problem Different scales
  in the same direction are treated alike )

13
Referral Systems
Example Qlt0.1 0.9gt, E1 lt0.5 0.5gt, E2 lt1
1gt In VSM, equal similarities. In defined
approach E2 is better.
14
Referral Systems
When expertise is matched against query, it must
be better enough.
15
Referral Systems
The sociability of an agent is its ability to
give good referrals. Every agent is evaluated by
expertise and sociability.
16
Referral Systems
? is the absolute relevance threshold which is
used to tune the number of retrieved results and
given referrals. ? w and ?0.3
17
Referral Systems
A referral chain of length l is defined by ltAr,
A1, ..., Algt with the originating agent Ar. The
chain constructs a referral graph which is
directed and its root is the originating agent.
The depth of a referral is the shortest path
from the root.
18
Referral Systems
An acyclic graph includes no redundant referrals.
In the example ltA4, A1gt is redundant. black
root white not
queried gray queried
19
Referral Systems

• In weighted referral graphs, the idea is that the
  agent with the greater weight is a better bet.

20
Referral Systems
Calculate w6 0.40.5 0.150.5 0.275 and w5
0.15 0.6 0.09 Recalculate weights of agents
at each referral addition
21
Referral Systems
If A4 referred to A2 then A2 would be a cut-point
since A2 is already referred to A6. In this
situation, relax is applied on A2 to propagate
weight changes to the descendants of A2.
22
Referral Systems

• Referral graph construction
• algorithm considers the
• length of referral chains
• when expanding a leaf
• agent and prefers leaf
• agents with shorter
• referrals if weights are
• equal.

23
Referral Systems

• When an answer is recieved ? operator is used to
  update the sociability and expertise of
  acquaintances by assigning rewards and penalties.
  This operator enables the values of sociability
  and expertise to build up slowly but fall
  quickly.

24
Referral Systems

• Given a referral graph G(Q), suppose Aj returns
  an answer T. The requesting agent Ar updates the
  expertise and sociability of its models as
  follows, where a is the rating given by Ars
  user, ß is the learning rate, -1 a 1 and 0
  ß 1.
• Expertise Ar will update the expertise vector
  for its own user as Er (1-ß)Er ßQ and the
  expertise vector for Aj as Ej (1-ß)Ej aßT. Aj
  updates the expertise vector for Pj as Ej
  (1-ß)Ej ßT.

25
Referral Systems

• Sociability If l is the depth
• of Aj in the referral graph,
• the credits (rewards and
• penalties) to Ajs ancestors
• according to their distance
• is propagated using the
• recursive algorithm.
• propagateCredits(Aj, l-1, a)
• example A6 returns answer of
• quality a, A2 and A4
• gets a, A3 and A1 gets a/2

26
Experimental Results

• A social network of 4933 AI scientists based on
  bibliographic data corpus from proceedings of
  AAAI (1980-2000) and IJCAI (1981-2001)
  conferences.
• For each paper author, title and keyword is
  extracted. Network is constructed as follows
• Each paper is classified into one of the 19
  topics in taxonomy for categorization. Not used
  for rating.
• An author is considered a neighbor of the other
  if only they have coauthored at least one paper.
  Random links are added for relationship.
• TFIDF (term frequency inverse document frequency)
  is used for initializing expertise vectors. Each
  element of ei of expertise vector E e1, e2,
  ..., en is calculated by multiplying TF and IDF.
  Two cases
• For profile in Aj, ek tfj idfk where tfj
  of papers authored by Pj in topic k and idfk
  log(N/nk) where N6635 (total of papers)
• For model of Aj in Ai, ek tfjidfk where tfj
  of papers Pi and Pj in topic k.
• If ekgt8 then the author is considered an expert
  in topic k. a is set to 1 if an expert is found.
  wi and ?i for each agent Ai is set to 0.1. The
  sociability of all agents in acquaintance models
  is set to 0.5. The learning rate ß 0.1.
• Two types of queries home (author has some
  papers) and foreign (author has no paper).

27
Experimental Results

• Effect of branching factor
• F (braching factor) how many neighbors an agent
  should refer to while processing a query
• Depth of referral graph is fixed at 6.
• F3 and F4 were needed to find all suitable
  experts for home and foreign queries.
• The referrals can support a focused search
  without spamming friends and colleagues.

28
Experimental Results

• Depth of referral graphs
• For home queries depth should be 5
• For foreign queries depth should be 6
• Chance for finding an acceptable answer,
• For home queries from 14 to 57
• For foreign queries from 0.6 to 19
• Avg. of experts goes down beyond a certain
  point

29
Experimental Results

• Accuracy of Referral Chains
• Sociability credits the ability to give good
  referrals.
• Referring process considers expertise and
  sociability.
• of neighbors are constant but models of
  acquaintances are updated so that they can be
  promoted to neighbors.
• The graph shows average of experts after
  sending 10 home or foreign queries.
• Learning can be effective especially for home
  queries.

30
Experimental Results

• Minimizing Referral Graphs by using weights.
• When Ar recieves referrals, it queries agent with
  heighest weight.
• Referring stops when an expert is found.
• Graph shows results when (branching factoe) F4.
• The requesting agent can efficiently find short
  paths to the desired experts, even though
  referrals are generated based on local knowledge.

31
A Prototype System

• Multiagent Referral System (MARS)
• is based on definitions mentioned above and
  includes an interface or text queries.
• implemented in Java
• uses IBM ABLE for reasoning
• registration server implemented over Sybase DBMS
• agents use email as transport mechanism
• evaluated among a small group of users of NCSU
  because f email server limitations
• Challenge bootstrapping
• MARS uses a server where new users register
  themselves along with their topics of expertise
  and interests. An agent contacts the registration
  server if it cannot find a suitable contact by
  itself.

32
Related Work

• Referral Systems
• MINDS a distributed information retrieval
  system in which agents share both knowledge and
  tasks in order to cooperate in retrieving
  documents for users.
• Huhns et. al.(1987) present heuristics for
  learning and updating the relevance of documents.
  
• Customizes document retrieval for each user.
• Kautz et. al.(1996) simulated expertise location
  in a large company.
• Shows how the length and accuracy of referral
  chains are affected by the number of users, the
  accuracy and responsiveness of each user.
• Kautz et. al. developed ReferralWeb in which
  co-occurrence of names in close proximity on Web
  pages is used to suggest direct person to person
  relationships.
• ContactFinder reads messages posted on bulletin
  boards and extracts topic areas based on
  heuristics.
• Vivacqua et. al. developed a user-interface
  agent, Expert Finder, which assist a novice user
  in finding experts by matching the profiles of
  novice and the expert.
• MITREs XperNet focuses on identification and
  tracking of expert communities using statistical
  clustering and network analysis.

33
Related Work

• Peer-to-Peer Networks
• A P2P node broadcasts a search request to its
  peers, who propagate the request to their peers,
  and so on.
• In Gnutella, distributed search algorithms
  broadcast a request to all peers in a brute force
  manner.
• Yang et. al. Study performance and tradeoff of
  three search techniques iterative deepening,
  directed BFS, local indices.
• The small-world phenomenon found by
  WattsStrogatz 1998 found that small-world
  networks are neither fully regular nor fully
  random. Such networks are highly clustered with
  just a few random short paths.

34
Related Work

• Matchmaking Systems
• Middle agents help multiagent systems provide
  effective, robust and scalable mechanisms.
• Middle agents cooperate with one another to
  locate agents with desired services. These
  architectures are ill-suited because of the
  presupposition of a fixed configuration.
• Shehory proposed a peer-to-peer location
  mechanism for open multiagent systems in which
  each agent caches a list of agents it knows.
• Matchmaking systems such as SHADE and Yenta group
  or cluster with similar interests. The basic idea
  behind matchmaking systems is bootstrapping each
  agent and finding at least one other agent with
  which to communicate and forming clusters of
  like-minded agents.

35
Conclusion

• Referral networks are prmising because they
  capture two essential aspects of social networks
  how they are applied and evolved.
• Another class of application is using a referral
  system as an ingredient of a practical multiagent
  system. This would make the system more resistant
  to failure.
• Further research may include incorporation of
  other mechanisms into the current MARS system.

36

• Thank You!
• Questions?