Date post: | 18-Jan-2016 |
Category: |
Documents |
Upload: | abigail-parrish |
View: | 217 times |
Download: | 0 times |
1
Design and Evaluation of P2P Transactive Memory System
Fu-ren LinInstitute of Technology Management
National Tsing-hua UniversityHsin-chu Taiwan 300
2
As We May Think
• by Vannevar Bush • Originally published in the July 1945 issue
of The Atlantic Monthly • “Consider a future device for individual use,
which is a sort of mechanized private file and library. It needs a name, and to coin one at random, ``memex'' will do. A memex is a device in which an individual stores all his books, records, and communications, and which is mechanized so that it may be consulted with exceeding speed and flexibility. It is an enlarged intimate supplement to his memory. “
Vannevar Bush (1890-1974 )
3
Rationale
• A peer as a personalized agent can simulate its master’s knowledge sharing behavior.
• A peer builds its social networks through communicating with other peers in computer networks.
• A peer-to-peer network can be viewed as a distributed knowledge management system by imbedding transactive memory system.
4
Transactive memory system
• Transactive memory theory explains how interdependent people within a knowledge network, each with their own set of skills and expertise, develop cognitive knowledge networks that help them identify the skills and expertise of others in the network .
• Three memory types– Internal memory: what you know– External memory: what others know– Transactive memory: know who knows what
5
Transactive memory system (cont.)
• 4 inter-related processes for an agent to develop transactive memory system:– Expertise recognition
• Identifying who knows what
• KOxij indicates agent i’s perception of agent j’s level of expertise on a particular item, X.
• The actual expertise of agent j on item X would be defined as KIxj.
6
Transactive memory system (cont.)
• Directory updating – learning who knows what in the group
– KIxi = f [KIxi, INFxi, CAIxji] – KOxij = f [Ai, KIxi, COMij, Â((COMik)(KOxkj))]
7
Transactive memory system (cont.)
• Information allocation– assigning memory items to group members
– CAIxij = f [KOxij, COMij, INFxi, Ai-Aj]
8
Transactive memory system (cont.)
• Retrieval coordination– planning how to find items in a way that takes adv
antage of who knows what
– CRIxij = f [TASKxi, KOxij, COMij, Ai-Aj]
9
P2P Transactive Memory System (cont.)
• A P2P network embedded with transactive memory can achieve the dual purposes: – Knowledge network development – Individual autonomy
10
Research objectives
• Developing a P2P knowledge management system with transactive memory to assist peer’s expertise recognition, knowledge network maintenance, information allocation, and retrieval coordination.
• Designing a system model which is obedient to human nature and follows the system development trend of decentralization.
• Observing the evolution of knowledge network in the community.
11
P2P transactive memory system
Directory updating Expertise recognition
Information allocation Retrieval coordination
Information allocation module Retrieval coordination module
Expertise recognition
moduleStereotype
module
Cognitive K.N.maintenance
module
Authoritycomputing
module
Cognitive K.N. exchange module
12
Transactive Memory for Virtual Team Development
A1A2
A3A4
B7
B1B3
B5
B4
B6B4
B2
C1
C2
A4
A3 D3
D2
D4D1
Virtual Team
Virtual Community (A P2P Network)
Sub-group
Knowledge Relation
Transactive Memory System
Expertise
Recognition
Directory
Updating
Information
Allocation
Retrieval
Coordination
13
Roles of peers
• Authority– A peer is an authority when the peer is an expert in a topic
and many other peers refer it when they need the knowledge of this topic.
– The role of an authority plays a knowledge center to distribute knowledge to peers.
– Determining who is an authority is based on the linkages of whole network.
• Hub• Referral
14
Roles of peers: hub
• A hub connects to multiple relative authoritative peers.
• We can see a hub peer as a recommender to indicate who an authority is.
• The hub peer pulls together authorities on a topic and allows us to ignore unrelated peers.
15
Roles of peers: referral
• A referral makes a bridge between requestors and providers.
• When a peer wants to get something from authorities, the criteria for an authority to decide whether to provide requested objects is the peer’s propensity to share which consists of its altruism and social network strength.
• A referral may be a friend or others who have direct or indirect relationship with requestors and providers.
16
Directory updating
• Cognitive knowledge network maintenance module
• Authority computing module• Cognitive network exchange module
– The willingness of sharing knowledge depends on the peer’s altruism and the social network strength. It can be represented by following equation.
– SKNij = function [SOCij, Ali ]
17
Expertise recognition• Expertise recognition module
– The expertise recognition module uses knowledge base to provide the mapping between expertise and knowledge items. When a peer receives advertises of other peers, the information will be processed by this module and map to knowledge items which the recipient maybe knows.
• Stereotype module– The stereotype module plays the role as a knowledge
base to provide necessary mappings, such as profession-expertise and expertise-knowledge, for expertise recognition module. When a peer can’t find the sources of needed knowledge item, the stereotype module provides a substitute to look up the possible alternative sources.
18
Information allocation
• Information allocation decides how to store the new information to an appropriate peer.
• When the related knowledge items are collected and stored by certain peers, it is easy to retrieve later.
• These authorities become the knowledge center and have the responsibility to store and share these knowledge items.
• The information allocation module extracts the key concept from the incoming file, and obtains the authority list from the cognitive knowledge network maintenance module.
• After mapping the correlation, the file will be sent to the authoritative peers to store.
19
Retrieval coordination
• The retrieval coordination module receives the search results from the expertise recognition module and the authority list from cognitive knowledge network maintenances module to retrieve the knowledge items from authorized peers.
20
Knowledge sharing decision model
• DMxij = f [REQxj, SOCij, Ali, THRxj].
Peer j sends a request about knowledge item
x to peer iPeers i and j’s social r
elation strength Peer i’s altruism
threshold
21
Evaluation
• This study has developed a prototyping P2P system to evaluate the performance of a transactive memory system on knowledge sharing and task collaboration.
• In experiments, knowledge networks on a P2P network are updated based on two schemes: exploration and exploitation.
22
Evaluation (cont.)
• Through exploration, a peer is a risk seeker to search potential knowledge owners through its acquainted peers.
• Through exploitation, a peer acquires information of other peers’ expertise via its cognitive knowledge network based on its transactive memory in terms of authority and hub.
• In experiments, different degrees of exploration and exploitation during knowledge sharing and task collaboration may result in different team performance.
23
Experimental settings
• Initialization– 6 stages of interactions
• Knowledge items (categories): between 3 and 7 items initially
• Propensity to share: three levels • Learning curve• Network status measure index:
– egocentric– whole networks.
24
Experimental settings (cont.)
• A peer’s knowledge sharing decision making– Propensity to share
• Altruism
• Strength of social network
– Ability• The ability on certain knowledge is growing according
to a peer’s learning curve.
• Learning curve is the path recording the progress track of a peer along its interactions with others.
25
Learning Curve
k Y(x) = 0.5 Y(x) 0.99≧
0.4 x = 12 x = 24
0.3 x = 16 x = 32
0.2 x = 23 x = 46
0.1 x = 46 x = 92 )( '
1
1)(
xxkexY
26
Network status measure index
Index1: )1(*21
p
s
N
NI , 0 ≦ 1I ≦ 1
Index2: )1(*2
pp
k
NN
NI , 0≦ 2I ≦ 1
sN is the number of links a peer connects,
pN the maximum number of possible connections a peer can have in the network
kN the number of existing links of the whole network.
27
Priority sequence of peer inquiry
• The different degrees of exploration and exploitation are measured in three parameters:
– authority value,
– cumulative inquiry success rate, and
– risk aspect.
28
Variables
• Ai denotes the authority values of peer i which owns a knowledge item in the knowledge network.
• SRi denotes the cumulative inquiry success rate of past transactions with peer i.
• Rij represents the risk aspect of of peer i toward peer j depends on the interaction frequency between peer i and j.
• If two peers have no interactions before, the risk is 1.0; if two peers had interaction, but did not succeed, the risk is set to 0.7.
29
Priority sequence of peer inquiry (cont.)
• Exploration: Y = Sweight * SRit + Rweight *Ri, where we set Sweight = index1, and Rweight = 1 - index1
• Exploitation: Y = Ai * Aweight + SRi * Sweight + Ri * Rweight
• A peer selects one out of three peers suggested by exploration and exploitation.
30
Performance criteria
• Cumulative inquiry success
• Learning outcomes
31
Experimental design
1. Initializing a P2P network
2. Constructing cognitive knowledge networks• Six stages, each stage performs five Q&A
activities
3. Exchange peers’ knowledge networks
4. Comparing the evolutions of two groups via exploration and exploitation, respectively.
32
Experimental results-cumulative success rate
Cumulative Success Rate of Stages
0.240.3 0.32 0.37 0.42
0.48
0.280.41
0.550.66
0.73 0.78
0
0.2
0.4
0.6
0.8
1
1 2 3 4 5 6
Stages
Cum
ulat
ive
Succ
ess Rat
e
exploration schema
exploitation schema
33
Experimental results-learning outcomes
Learing Outcome of Stages
0.013 0.023 0.0420.087
0.19
0.32
0.013 0.0290.1
0.27
0.45
0.6
0
0.10.2
0.3
0.4
0.50.6
0.7
1 2 3 4 5 6
Stages
Lear
ing
Out
com
e
exploration schema
exploitation schema
34
Experimental results-changes of indices 1 and 2
Stage1
Stage2
Stage3
Stage4
Stage5
Stage6
Index1 0.43 0.64 0.76 0.87 0.91 0.94 E1
Index2 0.09 0.13 0.15 0.18 0.18 0.19
Index1 0.42 0.63 0.7 0.71 0.71 0.71 E2
Index2 0.09 0.16 0.21 0.25 0.30 0.33
E1: the group with exploration schemeE2: the group with exploitation scheme
35
Findings
• Index1 and Index2 are designed to measure the status of egocentric and the whole networks.
• E1 increases cumulative success rate through the growing knowledge of the egocentric network.
• If a peer explores its egocentric network completely, it will find all peers in the network and raise the cumulative success rate.
• The increase of E1’s cumulative success rate ascribes to Index1’s increase.
36
Findings (cont.)• E2’s peers observe the whole network to fi
nd the authoritative peers. • E2’s members exchange their cognitive kn
owledge networks with others and increase the understanding of the whole network.
• E2’s Index2 increases faster than E1’s, but E2’s Index1 stops increasing in the later stages. – This is because E2’s peers quickly find all thei
r authoritative peers in the network and stop unnecessary interactions.
37
Findings (cont.)
• In this comparison, we found that the exploitation scheme is superior to the exploration in finding authorities.
• This finding is consistent with the reality of our human society. Through interactions with someone and recognition from other people, we can make an impression quickly and fairly.
38
Evolutions of Peers’ KN By Exploration
Stage 1. Index1=0.44, Index2=0.09 Stage 2. Index1=0.72, Index2=0.14
Stage 3. Index1=0.83, Index2=0.17 Stage 4.Index1=0.94, Index2=0.19
Stage 5. Index1=0.94, Index2=0.19 Stage 6. Index1=1.0, Index2=0.2
39
Evolutions of Peers’ KN By ExploitationStage 1. Index1=0.22, Index2=0.07 Stage 2. Index1=0.61, Index2=0.19
Stage 3. Index1=0.78, Index2=0.28 Stage 4. Index1=0.78, Index2=0.38
Stage 5. Index1=0.78, Index2=0.4 Stage 6. Index1=0.78, Index2=0.47
40
Conclusion• A transactive memory system has been designed and prototype
d to assist peer’s expertise recognition, knowledge network maintenance, information allocation, and retrieval coordination.
• Evaluate cumulative inquiry success rate and learning outcomes.• The evolutions of cognitive knowledge network show that trans
active memory can help peer improve the development of their cognition knowledge networks.
• The use of transactive memory to design P2P knowledge system is – not only obedient to human nature and follows the system d
evelopment trend of decentralization, – but also enhances the mechanisms of privacy, autonomy, an
d self-organization which a centralized architecture is hard to achieve.
41
Ongoing Works
• Developing P2P application systems to conduct field experiments.