4/21/16 1 Learning from Peers on Social Media Platforms Yingda Lu Param Singh Baohong Sun Cornell April 2016 1 Context of the Study • A crowdsourced customer support platform of a telecom company • Company employs this customer-driven platform to help answer customers’ questions 2
33
Embed
Learning from Peers on Social Media Platforms · 2016. 6. 16. · 4/21/16 1 1 Learning from Peers on Social Media Platforms Yingda Lu Param Singh Baohong Sun Cornell April 2016 2
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
4/21/16
1
Learning from Peers on Social Media Platforms
Yingda Lu Param Singh Baohong Sun
Cornell April 2016
1
Context of the Study
• A crowdsourced customer support platform of a telecom company
• Company employs this customer-driven platform to help answer customers’ questions
2
4/21/16
2
The Emergence of Crowdsourced Customer Support Platforms
3
4/21/16 4
4/21/16
3
Industry Background
• Anybody can register as a user • True identities of both knowledge sharer and
seekers are revealed – Act as quality control in the absence of feedback
• Most of the questions are technical questions – “How to insert and retrieve multilingual data using ORACLE
NCLOB?”
• Multiple answers are sequenced according to time stamps
• Top management use the forum to identify experts. Active contributors have higher probabilities of receiving bonus or promotion
• We focus on most basic features 4/21/16 5
4/21/16 6
4/21/16
4
Data Description
Table 2. Data Description
Variables Statistics Number of Users 1558 Number of Periods 30 Total Number of Questions 3748 Total Number of Solutions 2684 Average Number of Questions Asked per User 2.41 Average Number of Solutions per User 1.73 Mean of Tenure (Number of Weeks Registered Before Period 0) 25.08 Average Number of Views per Post 130.59
!
4/21/16 7
Learning Mechanism Enabled by Web 2.0
• Seeking knowledge – Ask question – Everybody can
answer – Knowledge seeker
gains knowledge – Whole community
gains knowledge
• Sharing knowledge
– Answer a question – Knowledge seeker
gains knowledge – Whole community
gains knowledge
Social network enabled very different learning mechanism 1. Learning from peers: learning cannot be achieved without
contribution from peers 2. Externality of learning: any contribution improves
knowledge level of the whole community Naturally, these properties imply decision process that is 1. Inter-dependent 2. Forward-looking 3. Future reward reciprocated by peers “Reciprocal Altruism”: an act of helping others at a cost is
beneficial if there is a chance of being in a reverse situation where the person whom I helped before may perform an altruistic act towards me.
4/21/16 8
4/21/16
5
Core/Periphery Structure
Customer Support Platform Network Structure
9
How to Improve Knowledge Sharing
• Concern of the firm: core/periphery structure might hinders problem solving for consumers
• Firm proposal: improving knowledge sharing by anonymizing participants’ identity
10
4/21/16
6
Research Questions
• What drives the formation of Core/Periphery structure?
• How does the Core/Periphery structure influence user participation and knowledge sharing on this customer support platform?
• How should we improve the design of the platform to improve knowledge sharing
11
Summary of Major Findings
• Core/periphery structure implications – Discourage individuals with low social status to contribute to the
community – Hurt knowledge sharing in online social media
• State variables: – Knowledge of self and of peers – Social reputation score of self and of peers
! !!!!!!!! !! ! !! ,
4/21/16 21
Oblivious Equilibrium and Estimation
• OE provides an appealing behavioral model – Large number of peers – Each individual makes nearly optimal decisions based
on her own state and the long-run market state. – The oblivious value function
– Strategy optimizes an
• Adopt EM algorithm by Arcidiacono and Miller (2011) to control for unobserved heterogeneity
1. !!(!!,!|!!! ,!! ) = ![ !!−!∞!=! !! ! !!,! ]
4/21/16 22
4/21/16
12
Estimation Results
23
Utility Function Parameters Type 1 Type 2 Impact from Knowledge 0.042*** 0.070*** Impact from Social Status 2.927*** 4.040*** Network Position Effect on Social Status 0.069*** 0.097*** Constant for Cost of Asking a Question 2.863*** 2.106*** Constant for Cost of Answering a Question 18.975*** 11.852*** Percentage of Customer in this Type 77.66% 22.34%
• Type 1 user derives more utility from knowledge and social status, and less cost of asking and answering questions
• A user would derive higher social status if she solves questions of other high status individuals
Whose Question Gets Solved
• Knowledge seeker with higher social status of more likely to receive a solution
24
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10.45
0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
Knowledge Seeker Social Status Level
Pro
babi
lity
of R
ecei
ving
a S
olut
ion
4/21/16
13
Whether to Participate
• High social status users more likely to ask questions – Because high status individuals are more likely to
receive solutions to their questions.
25 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
Knowledge Seeker Social Status Level
Pro
babi
lity
of A
skin
g a
Que
stio
n
Dynamics of Core/Periphery Formation
Core users: – Prefer to interact with core users – Ask more questions
26
Peripheral users: – Prefer to interact with core users – Receive no interact with others – Ask few questions
Formation of core/periphery
4/21/16
14
Dynamics of Core/Periphery Formation
Network 1 Network 2 Network 3 Core-periphery Degree 0.0509 0.0641 0.0484 Avg Knowledge for Core 13.119 13.418 12.482 Avg Knowledge for Peri 6.655 6.413 7.818
27
Randomly adding 500 ties
Adding 500 ties among core
Adding 500 ties among periphery
• Core-periphery structure leaves knowledge seekers with low social status positions in disadvantageous situations.
• Once a core appears, it reinforces itself through pattern of future interactions.
Sensitivity Analysis
• Anonymizing knowledge seeker’ identity
28
Benchmark Anonymity
Avg Core Knowledge Level 12.712 13.240
Avg Peripheral Knowledge Level 6.815 7.356
Avg Community Knowledge Level 8.132 8.670
4/21/16
15
Sensitivity Analysis
• Decrease the discount factor of individual contribution level to 0.9
29
Benchmark Faster Decay
Avg Core Knowledge Level 12.712 15.084
Avg Peripheral Knowledge Level 6.815 7.173
Avg Community Knowledge Level 8.132 8.940
Sensitivity Analysis
• Rewarding knowledge seeking behavior by giving reward that equals 10% of the baseline cost of asking questions
30
Benchmark Reward Qs
Avg Core Knowledge Level 12.712 17.173
Avg Peripheral Knowledge Level 6.815 7.978
Avg Community Knowledge Level 8.132 10.032
4/21/16
16
Summary of Major Findings
• Core/periphery structure implications – Discourage individuals with low social status to contribute to the
community – Hurt knowledge sharing in online social media
structure – Platform features may be used to prevent formation of core/
periphery structure
31
Robustness Check
• Using Absolute social status instead
• Changing discount factors
• Removing reciprocal answers
32
4/21/16
17
Yingda Lu Param Vir Singh
Baohong Sun
Cornell April, 2016
Is Core-eriphery Network Good for Knowledge Sharing? -- A Structural Model of Endogenous Network Formation on a Crowdsourced Customer Support Forum
4/21/16 33
Early Adopters Observe Two Challenges
First Challenge: Slow adoption
0
10000
20000
30000
40000
50000
60000
70000
1-A
pr-0
6 1-
May
-06
1-Ju
n-06
1-
Jul-0
6 1-
Aug
-06
1-S
ep-0
6 1-
Oct
-06
1-N
ov-0
6 1-
Dec
-06
1-Ja
n-07
1-
Feb-
07
1-M
ar-0
7 1-
Apr
-07
1-M
ay-0
7 1-
Jun-
07
1-Ju
l-07
1-A
ug-0
7
Figure 2A. The Total Number of Answers and Questionsa
Total Number of Answers Total Number of Questions
0
20
40
60
80
100
120
140
160
Figure 2B. The Number of People Who Participate in the Forum for the First Time
in Each Periodb
Number of People Who Become Active
4/21/16 34
4/21/16
18
Second Challenge: Most activities are generated by a few users Concentrated Network Structure
Figure 1. The Core/Peripheral Network Structurea a. Individuals are represented by spheres. Lines connecting two individuals represent the presence of a knowledge-sharing relationship between them. The arrow heads point towards the individual who answers the question. More active participants are indicated by larger spheres.
4/21/16 35
Research Questions
• 1) How the processes of “learning from peers” and “knowledge spill-over” influence the dynamics of individual knowledge seeking and sharing decisions;
• 2) Which drive the formation of core/periphery structure in social media platforms;
• 3) How the core/periphery structure influences user participation and knowledge sharing on a customer support forum;
• 4) How should we improve the design of social media platform to facilitate knowledge sharing on the social media platforms.
4/21/16 36
4/21/16
19
Related Literature • Marketing literature
– Consumer online decision making • Montgomery et al 2004, Eric et al 2004, Ghose and Yang 2009, Moe and Faer 2004, Sudhir 2004
– Word-of-mouth and social influence • E.g. Manchananda et a l2008, Hartman 2010, Nam et al 2010, Godes and Mayzlin 2004, Chevalier and
Mayzlin 2006, 2009, Weiss et a 2008 – Formation of network
• Narayan and Yang 2007 – Most assume exogenous network and adopt statistical approach
• Organizational behavior literature on knowledge sharing – Knowledge gains through interaction with peers improve productivity
• (i.e. Argote 1999, Levitt and March 1988, Benkard 2000, Argot and Epple 1990, Hatch and Mowery 1998, Bho et al 2007)
– Learning from peers • (i.e. Dar et al. 1995, Larson and Christensen 1993, Grewal et al 2005, Singh et al 2010, Hoh et al 2007,
Ingram and Simons 2002) – Do not directly observe learning
• Learning models – Learning through consumption experience, signals contained in price and advertising
• (i.e. Erdem and Keane 1996, Erdem, Keane and Sun 2008) – Atomistic view of individual experimentation behavior
– K_it: the knowledge level accumulated up to time t – R_it: social reputation score – Sum S_ijt: number of questions answered by i in t – Z_i: observed user characteristics – C(.): cost to be estimated – epsilon: private shock, distributed as type-I extreme
• State variables: – Knowledge of self and of peers – Social reputation score of self and of peers
! !!!!!!!! !! ! !! ,
4/21/16 46
4/21/16
24
Oblivious Equilibrium and Estimation
• OE provides an appealing behavioral model – Large number of peers – Each individual makes nearly optimal decisions based
on her own state and the long-run market state. – The oblivious value function
– Strategy optimizes an
• Adopt EM algorithm by Arcidiacono and Miller (2011) to control for unobserved heterogeneity
1. !!(!!,!|!!! ,!! ) = ![ !!−!∞!=! !! ! !!,! ]
4/21/16 47
– Reputation effect is not trivial – Cost of answering is higher than cost of asking. – Benefit cannot compensate for the cost in the current period. 4/21/16 48
4/21/16
25
Whose Questions to Answer?
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10.5
1
1.5
2
2.5
3
3.5
4
4.5
5
Own Social Status Level
Expe
cted
Num
ber o
f Ans
wers
Rece
ived
per Q
uesti
on
• Questions from high social status users are more likely to be answered • High status users are more likely to share knowledge
• Implication: high status users are likely to answer each other’s questions and thus forms a cohort
4/21/16 49
Individual Optimal Decision Rules -Whether to Participate?
Figure 2A. Probability of Receiving a Solution to a Question Given Knowledge Seekers’ Social Status Level
Figure 2B. Probability of Posting a Question given Social Status Level
• Prob of K seeking increases with reputation
• Prob of K sharing increases with reputation
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10.45
0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
Knowledge Seeker Social Status Level
Prob
abili
ty o
f Rec
eivi
ng a
Sol
utio
n
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
Knowledge Seeker Social Status Level
Prob
abili
ty o
f Ask
ing
a Q
uest
ion
4/21/16 50
4/21/16
26
Explain “Free Riding” Behavior
0
10000
20000
30000
40000
50000
60000
70000
Figure 2A. The Total Number of Answers and Questionsa
Total Number of Answers Total Number of Questions
0
20
40
60
80
100
120
140
160
Figure 2B. The Number of People Who Participate in the Forum for the First Time
in Each Periodb
Number of People Who Become Active
• Findings: • Slow start: low overall knowledge level • Low increasing rate: formation of cohort discourage users with low position from participating
• Managerial implications: • Some users are “forced” not to participate • The adoption does not really take off
4/21/16 51
Formation of Cohort Affects Effectiveness of Knowledge Sharing
Findings: • Only centralized users benefit from the formation of cohort.
Managerial implication: • New users do not get help • Flow of knowledge is not from high K users to low K users • K sharing is hurt when cohort structure is carried over to other question categories
4/21/16 52
4/21/16
27
Better Design to Align with User Behavior • Three Alternative Policies
– Rank based on recency – Hide the identify of knowledge seeker – Reward knowledge seeking
• Simulation takes into account ripple effect in the community – i asks or answers a question – The state variables of everybody in the whole community
change – All peers alter their decisions about asking and answering
questions – …… – The whole process continues
4/21/16 53
Breaking the Cohort and Reward Knowledge
Seeking
Findings: • Individual benefits more than the community • Higher knowledge increment when asking than answering questions
Managerial implications: • It is not about donation, it is about learning. • Proactive learning is more effective than reactive learning
Finding: • Hiding identity of knowledge seeker can improve knowledge sharing by 36%.
Managerial implication: • Encourage competition for reputation, but break the cohort • Alternative design: periodically reset the record, financial incentive to encourage individuals to answer questions from users with low social status
4/21/16 54
4/21/16
28
Managerial Implication
– Higher knowledge increment when asking than answering questions
– Hiding identity of knowledge seeker can improve knowledge sharing by 36%.
• It is not about donation, it is about learning.
• Proactive learning is more effective than reactive learning
• Encourage competition for reputation, but break the cohort
• Alternative design: periodically reset the record, financial incentive to encourage individuals to answer questions from users with low social status
4/21/16 55
Summary of Major Findings
• (1) Knowledge seeking and sharing are strategic decisions driven by knowledge and network position of herself and those of the community: users choose to seek and share knowledge for future rewards reciprocated by her peers.
• (2) Users are more likely to seek and share knowledge when her peers are more knowledgeable.
• (3) While both knowledge and reputation motivate users to share knowledge, a cohort is formed over time that has the privilege to obtain help from each other and in the meanwhile, exclude other users from participating.
• (4) “Free-riding” behavior of inactive contributors could be an equilibrium result: the earlier low
community knowledge level and the later formation of cohort “force” low ranked users from participating.
• (5) Proactive learning by asking is much more effective than reactive learning by observing
• (6) Current design of the open forum is not aligned with dynamic, interrelated and inter-dependent user decision process. An alternative design that breaks the cohorts can improve the knowledge sharing by 36%.
4/21/16 56
4/21/16
29
Conclusion • Formally investigated knowledge seeking and sharing decisions
– A new learning mechanism: learning from peers – Endogenize the formation of social network structure – Rationalize a seemingly altruism behavior: “reciprocal altruism” – Using observed decisions to integrate some economic, social and
psychological behavior – Provide explanations to the two observed challenges
• Some suggestions for adaptors – Recognize the conspicuous nature of platform adoption
• Knowledge Sharing Day (possibly?) – Build a formal reward system to recognize reputation building
• It is a double sided sword – Top management should change the mind set
• It is not a platform for donation • Proactive knowledge seeking behavior should be motivated
– Be aware that silos typically formed offline also form online. • Modified designs that encourage competition for reputation, but break the
cohort
4/21/16 57
Innovative Research
• Dynamic consumer decision making
• Technology enabled dynamic consumer decisions
• Firms act on customer information
4/21/16 58
4/21/16
30
THANK YOU!
4/21/16 59
Web 2.0 and Knowledge Sharing • An increasing focus on social software applications and
services to – Break the silos – Encourage knowledge sharing across department and
locations
• 90% are building the culture – IBM, CISCO, Infosys, Dell, Sun, Oracle – Knowledge sharing, idea generation – Customer service
• Customer help each other • Significant savings of service costs
• Many discussions about how Software/Media revolutionizing the working world of the future
4/21/16 60
4/21/16
31
Marketing Applications of Knowledge Sharing
• Customer service – Adobe and Oracle – OSI saw 22% decrease in time to solve customer
support issue
• Manage business process – 3 months less to complete project
• Production innovation (Crowdsourcing and ideation) – 36% decrease in time to enact key business changes
based on customer feedback
• Prediction market 4/21/16 61
Knowledge Updating is Different from Reputation Updating
• Reputation is directly affected by own decisions.
• Reputation ranking can go up or down • Compete for reputation ranking.
• Thus, – High reputation users may not have high knowledge – Low reputations users may have high knowledge
4/21/16 62
4/21/16
32
Our Study … • Explicitly model the dynamic and interdependent decision process
– rationalize the key driving forces behind knowledge seeking and sharing decisions
– not altruism but future reward reciprocated by peers
• Endogenize formation of a social network – demonstrates the formation of social network as a result of strategic
interaction.
• Recognize the unique learning mechanism enabled by social media platform – Treats knowledge sharing as consequence of dynamic interactions of
individuals. • learning from peers vs learning by doing • interdependent vs. dependent
– Integrate social psychology, information system and marketing
• Managerially, we evaluate current design of web 2.0. 4/21/16 63
Timeline of Decisions
• Everyone observes their own states as well as the states of everyone else in the community.
• Everyone receives their private shocks on the decision of asking question.
• Everyone makes predictions on their peers’ decisions based on equilibrium strategy given their information others’ states in current period. Using this prediction everyone simultaneously makes decisions on whether they are going to ask question.
• Everyone observes the outcomes of asking question decisions—they know who ask questions in current period.
• Everyone receives their private shocks on the decision of answering question.
• Given information on who ask questions in current period and the predictions on others’ decisions of answering questions, everyone simultaneously makes decisions on whether to provide answer for each one of the questions proposed.
• State variables, Accumulated Knowledge and Reputation Status , are both updated.
4/21/16 64
4/21/16
33
Model Comparison
Benchmark model 1: replace relative ranking by absolute ranking Benchmark model 2: remove forward-looking decision making Benchmark model 3: remove knowledge spill-over effect Benchmark model 4: remove the interaction among users