Number of submissions per task

Introduction to Labor
Marketplaces: Taskcn
Uichin Lee
KAIST KSE
KSE801: Human Computation and Crowdsourcing
Category
Rewards: >=1000CYN
Rewards: 500-1000CYN
Task Categories
Design
•Logo / VI Design
•Graphic design
•Software interface design
•Design of Buildings
•Brochure
•Three-dimensional modeling
•Product identification / packaging
Web Site
•Web Design / Production
•Site planning
•Web application development
•Flash animation
•Construction site as a whole
•Search Engine Optimization
Writing
•Named / slogan
•Technical / Application Writing
•Event Planning
•Business plans / tenders
•Literature Writing / Creative
•Translation and other writing
Programming
•Applications
•Scripts / Tools
•Database Development
•Mobile / embedded development
•System Management
Multimedia
•PPT presentation / courseware
•Video capture / editing
•Photography / photo post
•Audio / audio processing
•Multimedia data collection
Task Classification
Reward
Remaining
Time
Signed up /
Submitted
20% commission (regular user) vs. 18% commission (gold user)
Winner takes all (single bid)
1001 viewed; 38 signed up; 27 submitted
Below: viewing submissions (all 27 submissions)
Only “gold users”
can hide their
submissions
Taskcn Papers
• Crowdsourcing and Knowledge Sharing: Strategic
User Behavior on Taskcn, Jiang Yang, Lada A.
Adamic, Mark S. Ackerman, EC 2008
• Competing to Share Expertise: the Taskcn
Knowledge Sharing Community, Jiang Yang, Lada
A. Adamic, Mark S. Ackerman, ICWSM, 2008
• The Networks of Markets: Online Services for
Workers and Job Providers (Taskcn), MSR-TR 2010
Skill and Workload vs. Reward
• Human-coded variables: skill and workload
– Skill: minimum skill required to complete a task
– Workload: average time to finish a task
– Ordinal scale is used for rating
• Two raters evaluated 157 randomly selected tasks in the
design category
– Raters do not know the reward of a task
Reward
Minimum skill
0.493
Workload
-0.443
Minimum skill
-0.629
Spearman’s rank correlation coefficient
Taskcn: Stats
• Data set: 4 year data set: 17,000 tasks +
1.7m submissions
RMB/CNY(Chinese Yuan): 5CYN ~ $1
Open Tasks over Time
(Weekly Bins)
Offered Reward per Task
(dev: programming)
(CNY)
Worker Characterization
• 3 groups (based on how many submissions were made)
– At least 10 attempts (submissions), at least 50 attempts, and All
100 CNY
Average revenue per submission (CNY)
Worker Characterization
• # of submissions per task follows power law distribution
Number of submissions per task (CCDF)
Worker Characterization
• Joint distribution of workers’ submissions
across different categories
Market Segmentation
• Individual worker behavior:
– A typical worker tends to focus submissions to a
specific range of rewards.
– Specifically, a typical worker tends to submit most
frequently solutions to tasks from a narrow range of
rewards and attempts higher-reward tasks with
diminishing frequency with the value of the reward.
• Collective worker behavior:
– When workers are viewed as a community, however,
higher rewards tend to attract larger number of
submissions
Individual Behavior
• Histogram of submissions by top 10 workers (left fig)
• Experienced workers have a narrow reward range
Fraction of submissions
unique mode (= occurs most frequently)
1000, 500, 300, 200, ….
Reward (CYN)
Collective Behavior
• Number of submissions per task (across all workers) increases
as the associated reward increases
– Due to large number of workers who only made few attempts and
never came back to Taskcn (heavy tail: # submissions per worker)
Winning as Incentive to Continue
• High number of registered users never attempted any task (89%)
– June 2006 – May 2007 (EC 2008)
– 66,182 registered users
• Appears that people want to avoid the futility of their efforts (like
lotteries?)
• Winning experience is an important incentive:
– First attempt: 2307 won vs. 169,456 others failed
– The winner group has attempted more trials than the loser group
– Cox proportional hazard analysis: 19% lower probability of stopping
after each subsequent attempt
User’s Prestige Network
• Community expertise network (CEN): people’s
expertise can be measured by structural prestige
Centrality Metrics
• Calculate “centrality” metrics of a worker
– Degree centrality: sum of weights of out-edges of
worker u
– Eigenvector centrality: steady-state visit probability of
user u when a random walker traverses the normalized
graph
– Closeness centrality: the inverse of the average length
of a shortest path that originates from worker u and
terminates at worker v (for every worker v in the graph)
– Betweenness centrality: the sum of the fraction of
shortest paths between every pair of workers that pass
through worker u
User’s Prestige Network
• Same two users
compete twice:
– same winner 77%
of the time
(compared to 1/2
chance)
• Same two users
compete 3x:
Node size: proportional to Eigenvector centrality (PageRank)
Blue: a user who has won at least once
28 winners out of total 800 users
– same winner all 3
times in 56% of
the cases
(compared to 1/4
chance)
User’s Prestige Network
• Indegree/outdegree distribution (design)
win
lose
Task’s Prestige Network
• If winners of other tasks lose in this task, this task is
more prestigious...
User A won task X, but lost task Y
Task Y is more prestigious than task X
(directed edge from X to Y)
Motif Profiles
• Motif analysis provides a finer grained, local view into the
networks of users/tasks
• Below table shows the frequencies of dyadic and triadic
motifs
Average Expertise of All Users of a Task
• For a given task, test association between centrality of the
task (Task’s Prestige Net) vs. avg. indegree and PageRank of
the users who submitted to the task (User’s Prestige Net)
– Task outdegree (lost) vs. PageRank (low)
Importance of Experience
• The reward of selected tasks by typical
workers exhibits a diminishing increase with
the number of submissions (of course)
• The expected revenue per submission by
typical workers tends to increase until it
settles around a constant with the number of
submissions.
Importance of Experience
• Users learn to choose tasks with less competitive tasks
• Skilled users survive and continue to participate the work (?)
Importance of Experience
Average Revenue (CNY)
Average Rewards (CNY)
Estimated Prob. of Winning
• Winning probability increases as # submissions increases (also
avg. rewards increases, yet it tapers off); thus avg. revenue
also increases and it tapers off
Summary
• Amount of reward does not correlate with
– # submissions
– Expertise level
• Yet, it does correlate with the number of views
• Can infer expertise from expertise networks
• Successful users
– Choose less popular tasks
– Focus on specific reward range (best suited for one’s
expertise?)
– Increase revenue with # of attempts (but it tapers off)