Eliciting Truthful Information

Verfiable information
Unverifiable information
Decentralized machine learning

Game Theory for Data Science:

Eliciting High-Quality Information

Boi Faltings1 Goran Radanovic2

1 AI Laboratory, EPFL

[email protected]
2 EconCS, Harvard

[email protected]

December 11th, 2017

c Boi Faltings, Goran Radanovic Game Theory for Data Science 1/92
 Eliciting Truthful Information
Verfiable information
Quality control
Unverifiable information
Decentralized machine learning

Background Material

Boi Faltings and Goran Radanovic:

Game Theory for Data Science:
Eliciting Truthful Information,
Morgan & Claypool Publishers, 2017.
15% discount with code: authorcoll

c Boi Faltings, Goran Radanovic Game Theory for Data Science 2/92
 Eliciting Truthful Information
Verfiable information
Quality control
Unverifiable information
Decentralized machine learning

Big data

Organizations and individuals base decisions on data rather than

principles:
financial markets
medicine
choosing a restaurant/hotel/spouse
law enforcement
...
Often, data must be obtained from others.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 3/92
 Eliciting Truthful Information
Verfiable information
Quality control
Unverifiable information
Decentralized machine learning

Product reviews

Reviews and ratings great to avoid poor products.

Having reviews is essential for selling the product.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 4/92
 Eliciting Truthful Information
Verfiable information
Quality control
Unverifiable information
Decentralized machine learning

Why do we need to worry about quality?

Amazon ratings distribution for music CD Mr. A-Z, reported by Hu et al. (2006), and empirical observation.

Laziness: most people do not write reviews.

Self-selection: most reviews are written for ulterior motives,
e.g. reviews paid for by hotel, push your own opinion, etc.
Malicious participants: paint a fake picture of reality.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 5/92
 Eliciting Truthful Information
Verfiable information
Quality control
Unverifiable information
Decentralized machine learning

Forecasting polls

Will Scotland become independent?

Internet can be used to collect forecasts of important events.

Important for many high-stakes decisions.
Need to encourage knowledgeable participants and accurate
estimates.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 6/92
 Eliciting Truthful Information
Verfiable information
Quality control
Unverifiable information
Decentralized machine learning

Crowdwork

Human computation: tasks solved by workers recruited

through the internet (e.g. Amazon Mechanical Turk).
Peer grading: students grade each others’ homework.
Huge benefits for knowledge acquisition, online courses, etc.
Need to make workers invest effort to obtain quality results.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 7/92
 Eliciting Truthful Information
Verfiable information
Quality control
Unverifiable information
Decentralized machine learning

Quality control options

Filtering: eliminate outliers, inconsistent data.

Reputation: eliminate bad providers.
Incentives: encourage participation and effort of good data
providers.
All 3 can be used together - we focus on incentives and reputation.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 8/92
 Eliciting Truthful Information
Verfiable information
Quality control
Unverifiable information
Decentralized machine learning

The promise of incentives

Filtering, reputation need to throw away data ⇒ wasteful.

Incentives can also increase the amount of good data.
Incentives can be inaccurate as long as participants believe
that they are right on average.
However, participants may misunderstand or not care.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 9/92
 Eliciting Truthful Information
Verfiable information
Quality control
Unverifiable information
Decentralized machine learning

Setting

Limit to variables with discrete values x1 , .., xk .

c Boi Faltings, Goran Radanovic Game Theory for Data Science 10/92
 Eliciting Truthful Information
Verfiable information
Quality control
Unverifiable information
Decentralized machine learning

Choosing a strategy

Agent has to choose strategy:

heuristic:
report a constant value.
report a random number.
report ...
honest/truthful: perform accurate measurement and report
truthfully.
Rational agent: chooses strategy with highest payoff.
Mechanism: influence choice through payment rule.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 11/92
 Eliciting Truthful Information
Verfiable information
Quality control
Unverifiable information
Decentralized machine learning

Principle underlying truthful mechanisms

Reward reports according to consistency with a reference:

verifiable information: ground truth g will become known and
can be used as a reference.
unverifiable information: ground truth will never be known.
Reference is constructed from peer reports.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 12/92
 Eliciting Truthful Information
Verfiable information
Quality control
Unverifiable information
Decentralized machine learning

Roadmap

Verfiable information
Unverifiable, objective information
Parametric mechanisms for unverifiable information
Non-parametric mechanisms for unverifiable information
Distributed machine learning

c Boi Faltings, Goran Radanovic Game Theory for Data Science 13/92
 Eliciting Truthful Information
Proper scoring rules
Verfiable information
Prediction markets
Unverifiable information
Reputation
Decentralized machine learning

Eliciting verifiable, objective information

Forecasting, estimation, cumulative phenomena: truth can be

verified later.
⇒ payment can use verification.
eliciting a value: reward if report is accurate prediction.
eliciting a probability distribution: scoring rules.
eliciting a consensus probability distribution: prediction
markets.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 14/92
 Eliciting Truthful Information
Proper scoring rules
Verfiable information
Prediction markets
Unverifiable information
Reputation
Decentralized machine learning

Agent beliefs

Agent i has beliefs about what others observe:

1 prior probability distribution Pri (x).
We abbreviate Pri (x) = pi (x) or just p(x).
Often common to all agents (e.g. current review scores).
Maximum likelihood: agent endorses x p = argmaxx p(x).
2 measures signal si and forms a posterior distribution Pri (x|si ).
We abbreviate Pri (x|si ) = qi (x) or q(x).
Update prior → posterior often different for each agent.
Agent endorses x q = argmaxx q(x).
Beliefs motivate agent actions: crucial for incentives.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 15/92
 Eliciting Truthful Information
Proper scoring rules
Verfiable information
Prediction markets
Unverifiable information
Reputation
Decentralized machine learning

Eliciting a value

Truth Matching mechanism:

t1 agent makes observation and forms posterior belief q.
t2 agent reports a value v to the center.
t3 ground truth g observed; center pays reward only if v = g .
Expected reward E [pay ] = q(v )
⇒ maximized by choosing v = x q = argmaxx q(x):
Rational agent reports its best estimate truthfully.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 16/92
 Eliciting Truthful Information
Proper scoring rules
Verfiable information
Prediction markets
Unverifiable information
Reputation
Decentralized machine learning

Costly measurements

Measurement has a cost m.

If q ' p, agent may decide to skip the measurement!
m should not exceed

q(x q ) − p(x p )
| {z } | {z }
Epost [pay ] Eprior [pay ]

m
Scale payment by α ≥ q(x q )−p(x p ) to ensure this condition!
α depends on measurement technology - could also be
inferred from agent behavior.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 17/92
 Eliciting Truthful Information
Proper scoring rules
Verfiable information
Prediction markets
Unverifiable information
Reputation
Decentralized machine learning

Discouraging random reports

Even without measurement, agent still gets a reward by

reporting its prior most likely value x p .
⇒ system will be polluted with many uninformed reports!
Subtract Eprior [pay ] = p(x p ).
Designer needs to estimate p(x p ), q(x q ); can use some
background constraints, e.g. p(x p ) ≥ 1/N.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 18/92
 Eliciting Truthful Information
Proper scoring rules
Verfiable information
Prediction markets
Unverifiable information
Reputation
Decentralized machine learning

Components of payment schemes

Final payment rule:

  
p 1 if x = g
pay (x, g ) = α −p(x ) +
0 otherwise
= α [1x=g − p(x p )]

Components:
incentive for truthful report (1 if ground truth is matched)
offset to make expected reward of random reports = 0
scale to compensate cost of measurement
Focus on incentives for truthful reports.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 19/92
 Eliciting Truthful Information
Proper scoring rules
Verfiable information
Prediction markets
Unverifiable information
Reputation
Decentralized machine learning

Reporting probability distributions

Report is not a value, but probability distribution A.
Proper scoring rule = payment function pay (A, g ) such that:
X X
(∀q 0 6= q) q(x) · pay (q, x) > q(x) · pay (q 0 , x)
x x

Examples:
quadratic scoring rule:
X
pay (A, g ) = 2 · A(g ) − A(x)2
x∈X
Gneiting T., and Raftery A.
logarithmic scoring rule:
Strictly proper scoring rules,
pay (A, g ) = C + ln A(g ) prediction, and estimation.
JASA 2007.
c Boi Faltings, Goran Radanovic Game Theory for Data Science 20/92
 Eliciting Truthful Information
Proper scoring rules
Verfiable information
Prediction markets
Unverifiable information
Reputation
Decentralized machine learning

Eliciting forecasts with proper scoring rules

”What will be the weather next Sunday: Rain, Cloudy or Sunny?”
Agent prior = historial averages (for example):

Rain Cloud Sun

p=
0.2 0.3 0.5

Agent studies data ⇒ posterior belief:

Rain Cloud Sun

q=
0.8 0.15 0.05

Agent reports the entire posterior belief.

On Sunday, using the logarithmic scoring rule he gets paid
C + ln 0.8 = C − 0.22 if it rains, and C + ln 0.05 = C − 3 if it
is sunny.
c Boi Faltings, Goran Radanovic Game Theory for Data Science 21/92
 Eliciting Truthful Information
Proper scoring rules
Verfiable information
Prediction markets
Unverifiable information
Reputation
Decentralized machine learning

Why is this truthful?

Expected reward using log scoring rule:
X X
E [pay (A, g )] = q(x) · pay (A, x) = q(x) · [C + ln(A(x))]
x x

and the difference between truthful/non-truthful reporting:

E [(pay (q, g )] −E [pay (A, g )]
X
= q(x) · [C + ln q(x)) − (C + ln A(x))]
x
X q(x)
= q(x) · ln
x
A(x)
= DKL (q||A)
By Gibbs’ inequality, DKL (q||A) ≥ 0, so reporting an A 6= q can
only get a lower payoff!
c Boi Faltings, Goran Radanovic Game Theory for Data Science 22/92
 Eliciting Truthful Information
Proper scoring rules
Verfiable information
Prediction markets
Unverifiable information
Reputation
Decentralized machine learning

Prediction markets
Model on a financial market.

Market = trade securities σ(xi )

for predictions xi that pay $1 if
g = xi and $0 otherwise.

Every security has a market

price π(xi ).
Competitive equilibrium = Hanson R. Logarithmic market scoring
π(xi ) is a consensus probability rules for modular combinatorial information
estimate for Pr(g = xi ). aggregation. JPM 2007.
Chen Y., and David P. A utility framework
Bigger investment ⇔ bigger for bounded-loss market makers. UAI 2007.
influence, but also risk.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 23/92
 Eliciting Truthful Information
Proper scoring rules
Verfiable information
Prediction markets
Unverifiable information
Reputation
Decentralized machine learning

Automated market makers

Consider security R that pays 1 if it rains at any time next

Saturday, and 0 if it doesn’t rain.
To ensure that there is always a way to buy and sell this
security, we introduce an automated market maker agent.
Market maker maintains a net balance n of securities R that
have been bought or sold to it (negative means sold short).
Q: how to choose a price function π(n) for buying or selling
the next share, so that it shows the estimated probability of
the event?
A: Agents should be paid for how much their actions improved
or deteriated the predicted probability, according to a proper
scoring rule.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 24/92
 Eliciting Truthful Information
Proper scoring rules
Verfiable information
Prediction markets
Unverifiable information
Reputation
Decentralized machine learning

Example

Security R: Rain(Saturday) pays 1 if and only if it rains at any

time on Saturday, 0 otherwise. Initially π(0) = p(R) = 0.5.
t1 Agent 1 uses prior probability (historical average) p(R) = 0.2,
short sells n1 securities R until price becomes π = 0.2.
t2 Agent 2 studies data and concludes: q(r ) = 0.8
⇒ buys n2 securities R as long as π < 0.8.
t3 next, Agent 3 studies the data and concludes q(r ) = 0.5
⇒ short sells n3 securities R as long as π > 0.5.
te on Saturday, it rains and agents should get the following
rewards (using log2 scoring rule Sr (p, x) = b log2 p(x):
1 Sr (0.2, 1) − Sr (0.5, 1) = b(log2 0.2 − log2 0.5) = −1.322b
2 Sr (0.8, 1) − Sr (0.2, 1) = b(log2 0.8 − log2 0.2) = 2b
3 Sr (0.5, 1) − Sr (0.8, 1) = b(log2 0.5 − log2 0.8) = −0.678b

c Boi Faltings, Goran Radanovic Game Theory for Data Science 25/92
 Eliciting Truthful Information
Proper scoring rules
Verfiable information
Prediction markets
Unverifiable information
Reputation
Decentralized machine learning

Market makers with a logarithmic scoring rule

Assume participant believes that true probability of outcome

xi is π ∗ (xi ) > π(xi ).
⇒ buys m securities and makes the price increase to some
π(n + m) = π 0 > π(n).
⇒ he should make a profit of Sr (π 0 , 1) − Sr (π, 1) if the outcome
is indeed xi :
Z n+m
m− π(µ)dµ = Sr (π(n + m), 1) − Sr (π(n), 1)
n

dSr (π(n)) dSr dπ

(1 − π(n)) = =
dn dπ dn
2n/b
LMSR: Sr (π) = b log2 π =⇒ π(n) = 2n/b +1
c Boi Faltings, Goran Radanovic Game Theory for Data Science 26/92
 Eliciting Truthful Information
Proper scoring rules
Verfiable information
Prediction markets
Unverifiable information
Reputation
Decentralized machine learning

Example
In the example, assuming b = 1: Securities bought/sold:
2n
t1 : π(n) = = 0.2 ⇒ n(t1 ) = −2 : n1 = −2
2n + 1
2n
t2 : π(n) = n = 0.8 ⇒ n(t2 ) = 2 : n2 = 4
2 +1
2n
t3 : π(n) = n = 0.5 ⇒ n(t3 ) = 0 : n3 = −2
2 +1
1 2 shares sold that drive the price from 0.5 to 0.2:
price/share = (log2 0.2 − log2 0.5)/2 = −0.339
2 4 shares bought that drive the price from 0.2 to 0.8:
price/share = (log2 0.8 − log2 0.2)/4 = 0.5
3 2 shares sold that drive the price from 0.8 to 0.5:
price/share = (log2 0.5 − log2 0.8)/2 = −0.661
c Boi Faltings, Goran Radanovic Game Theory for Data Science 27/92
 Eliciting Truthful Information
Proper scoring rules
Verfiable information
Prediction markets
Unverifiable information
Reputation
Decentralized machine learning

Example..

At the end:
Agent 1 has to pay for 2 borrowed shares, net gain
= 2 · 0.339 − 2 = −1.322.
Agent 2 gets 4 for this shares, net gain = 4 − 4 · 0.5 = 2.
Agent 3 has to pay for 2 borrowed shares, net gain
= 2 · 0.661 − 2 = −0.678
Correspondence with scoring rules makes it optimal for agents to
act according to their true beliefs about the posterior probability q!

c Boi Faltings, Goran Radanovic Game Theory for Data Science 28/92
 Eliciting Truthful Information
Proper scoring rules
Verfiable information
Prediction markets
Unverifiable information
Reputation
Decentralized machine learning

”Irrational” agents

Some agents do not respond to incentives

faulty agents, who do not consider the incentive or who are
unable to provide correct data.
malicious agents, who want to insert fake data for ulterior
motives, for example to hide pollution.
Approach: limit their negative influence on the learned model
through reputation.

Resnick P. and Sami R. The influence limiter: provably manipulation-resistant

recommender systems. RecSys 2007.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 29/92
 Eliciting Truthful Information
Proper scoring rules
Verfiable information
Prediction markets
Unverifiable information
Reputation
Decentralized machine learning

Information fusion

Estimate 1 Estimate 2 Estimate 3

Evaluate
Update reports

Center
Ref. agent
Report

Agent 1 Agent 2 Agent 3

c Boi Faltings, Goran Radanovic Game Theory for Data Science 30/92
 Eliciting Truthful Information
Proper scoring rules
Verfiable information
Prediction markets
Unverifiable information
Reputation
Decentralized machine learning

Information fusion in community sensing

Evaluate  contribu/ons  
Trusted  sensor  
  Pollu/on     Pollu/on     Pollu/on    
Map  P1   Map  P2   Map  Pn  

Update   Update   Update  

Pollu%on  model  M  

Report  1   Report  2   Report  n  

Sensor  1   Sensor  2   Sensor  n  

c Boi Faltings, Goran Radanovic Game Theory for Data Science 31/92
 Eliciting Truthful Information
Proper scoring rules
Verfiable information
Prediction markets
Unverifiable information
Reputation
Decentralized machine learning

Reputation principle

Agents interact with the system over time t.

Assign a reputation score that determines if an agent is
misbehaving.
The reputation of an agent is based on the agent’s influence.
The report of an agent changes prediction for location of
reference measurement from p(x) to q(x).
Evaluate the quality by proper scoring rule Sr on reference
measurement gt :

scoret = Sr (q, gt ) − Sr (p, gt ) ∈ [−1, +1]

Use scoret to update the reputation rept .

c Boi Faltings, Goran Radanovic Game Theory for Data Science 32/92
 Eliciting Truthful Information
Proper scoring rules
Verfiable information
Prediction markets
Unverifiable information
Reputation
Decentralized machine learning

Simple reputation system

Thresholding: submitting data requires a minimal reputation.
Most common representation — β reputation system:
αt
rept =
αt + βt
P
where αt =Pα0 + s∈{scoresτ >0} |s| and
βt = β0 + s∈{scoresτ <0} |s|.

However, allows manipulation:

Provide good data that does not change the model.
⇒ build up reputation.
Use reputation to insert bad data that does change the model.

Josang A., and Ismail R. The beta reputation system. EC 2002.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 33/92
 Eliciting Truthful Information
Proper scoring rules
Verfiable information
Prediction markets
Unverifiable information
Reputation
Decentralized machine learning

Stochastic influence limiter

rept
Stochastic information fusion: with probability rept +1 accept
report.
Exponential reputation update:
1
rept+1 = rept · (1 + · scoret )
2
Main properties:
1 Negative influence is upper bounded by 2 · init.rep.
2 Information loss is upper bounded by a constant.
Empirical performance often much better.

Radanovic G., and Faltings B. Limiting the influence of low quality

information in community sensing. AAMAS 2016.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 34/92
 Eliciting Truthful Information
Proper scoring rules
Verfiable information
Prediction markets
Unverifiable information
Reputation
Decentralized machine learning

Empirical evaluation (Pollution Sensing)

Reputation systems:
CSIL - stochastic influence limiter
BETA - beta reputation system
Pollution model of Strassbourg (France):

c Boi Faltings, Goran Radanovic Game Theory for Data Science 35/92
 Eliciting Truthful Information
Proper scoring rules
Verfiable information
Prediction markets
Unverifiable information
Reputation
Decentralized machine learning

Performance results

c Boi Faltings, Goran Radanovic Game Theory for Data Science 36/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Ground truth is never known

In many cases, ground truth is

never known:
product reviews
community sensing
predictions about
hypothetical questions

Peer consistency: evaluate consistency with peer reports.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 37/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Peer consistency mechanisms

Reporting information becomes a game among agents: reward

depends on actions of agent and peer agent.
Optimal strategy = equilibrium of the game.
Truthtelling becomes an equilibrium: if peers are truthful,
truthtelling is the best response.
Equilibria also depend on agents’ beliefs about others.
Agents need to have similar beliefs so that the same
mechanism works for all of them!

c Boi Faltings, Goran Radanovic Game Theory for Data Science 38/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Output agreement

Term coined by von Ahn:

Ask 2 people to solve a task.
Pay a constant reward if two people give the same answer.
Q: When does this incentivize truthfulness/maximum effort?
A: In objective tasks: agents believe that honest peers are most
likely to obtain the same answer.

Truthful reporting is an equilibrium.

von Ahn, L. and Dabbish, L. Designing games with a purpose.

Communications of the ACM 2008.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 39/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

ESP game

Guess keywords to label

image.
Matching guess of an
(unknown) partner gives
points.
Taboo words to exclude
trivial choices.
von Ahn, L. and Dabbish, L. Labeling
images with a computer game. HFCS
2004.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 40/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Objective vs. Subjective

objective data: all agents observe a noisy version of the same

realization of the phenomenon. Example: temperature at
point x and time t. Center wants to know ground truth.
subjective data: each agent observes a different realization of
the phenomenon. Example: service in a restaurant, quality of
a product. Center wants to know distribution.
Goal: predict observations of a new agent.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 41/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Subjective observations

Reporting on the quality of service of Blue Star Airlines, with

very high reputation.
My plane is late and baggage lost. Should I report poor
service?
A: no, because most people enjoy good service, so my report
is less likely to match the peer!
Not categorical: agents do not believe that the same bad
service is most likely for others.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 42/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Types of peer consistency mechanisms for subjective tasks

There is no truthful peer consistency mechanism without

assumptions about agent beliefs!
Known mechanisms make assumptions about prior and posterior
beliefs about the distribution of peer agent reports:
Homogeneous agent population with identical and known
prior and posterior beliefs, example: peer prediction
Common and known prior beliefs, but belief updates can be
heterogeneous as long as they satisfy a self-predicting
condition, for example: Peer truth serum (PTS)
Utilize the multi-task structure of crowdsourcing to
accomodate subjective beliefs and provide stronger incentives,
for example: PTS for crowdsourcing

c Boi Faltings, Goran Radanovic Game Theory for Data Science 43/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Peer Truth Serum

Assume center maintains and publishes distribution R of all

prior reports (initially uniform). When agent reports ui = xj ,
center updates:

R̂t+1 = (1 − δ)Rt + δx j

Reward by impact of xj on accuracy of model, evaluated using

proper scoring rule SR for predicting a random peer report xp :

pay (Rt , xj ) = SR(R̂, xp ) − SR(R, xp )

= SR((1 − δ)R + δx j , xp ) − SR(R, xp )

c Boi Faltings, Goran Radanovic Game Theory for Data Science 44/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Derivative of reward function

Assume we use log scoring rule:

SR(R, xp ) = ln r (xp )

with derivative:

∂SR(R, xp ) 1/r (x) x = xp 1x=xp
= =
∂r (x) 0 x=6 xp r (x)

Partial derivatives of R̂ with respect to parameter δ are as

follows:

dr̂ (x) 1 − r (x) x = xj
= = 1x=xj − r (x)
dδ −r (x) x = xk 6= xj

c Boi Faltings, Goran Radanovic Game Theory for Data Science 45/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Approximation by Taylor expansion

Approximate model improvement on a random peer report xp

by the first term of the Taylor expansion:
X ∂SR(R̂, xp ) dr̂ (z)
SR(R̂, xp ) − SR(R, xp ) ≈ δ
z
∂r̂ (z) dδ
X 1z=xp



= δ 1z=xj − r (z)
z
r (z)
!
X 1z=xj 1z=xp X r (z)
= δ − 1z=xp
z
r (z) z
r (z)
1xj =xp
 
= δ −1
r (xj )

c Boi Faltings, Goran Radanovic Game Theory for Data Science 46/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Reward for quadratic scoring rule

Using the quadratic scoring rule:

X
SR(R̂, xp ) = r̂ (xp ) − 0.5 r̂ 2 (x)
x

with derivative:

∂SR(R, xp ) 1 − r (x) x = xp
= = 1x=xp − r (x)
∂r (x) −r (x) x=6 xp

we obtain:
X


pay (q̂, xp ) = δ 1z=xp − r (z) 1z=xj − r (xj )
z


= δ 1xj =xp − r (xj )

c Boi Faltings, Goran Radanovic Game Theory for Data Science 47/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Payment

Agent i reproduces the identical calculation ⇒ incentive for

optimally improving the center’s estimate.
Assume Bayesian update; δ for agent i is unknown...
...but reward is proportional to δ: δ is just a scaling factor!
⇒ choose payment proportional to the improvement:
1xj =xp
pay (xj , xp ) = −1
r (xj )
Faltings, B., Jurca, R., and Radanovic, G. Peer Truth Serum: Incentives for
Crowdsourcing Measurements and Opinions. CoRR abs/1704.05269 2017.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 48/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Incentive Compatibility

Incentive compatibility condition for xl 6= xj :

EP(x|xj ) [pay (xj , x)] = q(xj ) · pay (xj , xj ) = q(xj )/r (xj )
> EP(x|xj ) [pay (xl , x)] = q(xl ) · pay (xl , xj ) = q(xl )/r (xj )

Assume (for now) that agent adopts R as its prior P.

⇒ translates to self-predicting condition:

p(xj |xj ) p(xl |xj )

> , l 6= j
p(xj ) p(xl )

Satisfied for Bayesian belief updates.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 49/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Belief updates
Assume agent receives a signal si :
s i = (Pr (obs|x1 ), Pr (obs|x2 ), ..., Pr (obs|xk ))
Bayesian update:
ui (x|obs) = αpi (x)Pr (obs|x)
P
where α = 1/Pr (obs) set so that u = 1.
Objective update: agent trusts its measurement
qi (x) = ui (x)
Subjective update: observation is one of many data points
qi (x) = (1 − δ)pi (x) + δui (x)
with δ = 1/n if p is formed by n − 1 other observations.
Sensing error: agent might treat objective data as subjective.
c Boi Faltings, Goran Radanovic Game Theory for Data Science 50/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Self-predicting belief updates

Maximum likelihood: Agent endorses xi = argmaxx Pr (obs|x)

Bayesian update for subjective data:

qi (x) = pi (x)(1 − δ + δαPr (obs|x))

⇒ update is self-predicting:

qi (xi )/pi (xi ) > qi (xj )/pi (xj ), xj 6= xi

since xi = argmaxx Pr (obs|x) = argmaxx δαPr (obs|x)

c Boi Faltings, Goran Radanovic Game Theory for Data Science 51/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Helpful reporting
What if R 6= P (for example, on initializing the mechanism)?
Consider that P is more informed, i.e. closer to true distribution
P ∗ than R (in the interval between P ∗ and R).
⇒ Agents partition values into:
under-reported: r (x) < p(x) ⇔ r (x) < p ∗ (x)
over-reported: r (x) ≥ p(x) ⇔ r (x) ≥ p ∗ (x)
Non-truthful strategy: report x instead of y :
May be profitable if x under-reported or y over-reported.
Never profitable if x over-reported and y under-reported.
Helpful strategy: never report over-reported x for under-reported y .

Jurca R., and Faltings B. 2011. Incentives for answering hypothetical

questions. SCUGC 2011.
c Boi Faltings, Goran Radanovic Game Theory for Data Science 52/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Asymptotic accuracy

Assume center maintains R as an aggregate over reports

received over time (for example histogram).
Asymptotically accurate: R converges to true distribution P ∗ .
Any mechanism that induces helpful reporting is
asymptotically accurate.
Peer truth serum admits equilibria in helpful strategies.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 53/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Properties of the Peer Truth Serum

Optimal: when loss function is logarithmic scoring rule,

incentive for agent reports is to minimize loss function for
center.
Unique: any payment function that incentivizes truthful
reporting with only the self-predicting condition must have the
form f = 1/p(xj ) + g (−xj ) where g (−xj ) is a function
independent of the report xj .
Maximal: weakening conditions leads to impossibility results.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 54/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Peer prediction method

Rather than reward the most likely Prob. distribution q

z
value... Peer
q(x)=0.25
x y

Peer prediction method: q(y)=0.75

Payment
rule

Each value for answer xi is Report x Estimate q* Report z

associated with an assumed Agent Center

Pay ln(q*(z))
Peer

posterior distribution Example: shadowing

q̂(x) = P̂r(x|xi ). p q* q

q̂ is skewed so that xi is more

likely than in prior. Miller N., et al. Eliciting in-
Use a proper scoring rule to formative feedback: the peer
score this posterior against a prediction method. Manage-
random peer report. ment Science 2005.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 55/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Other equilibria...

All agents report x with smallest r (x).

⇒ equilibrium with highest possible payoff.
Will lead to uniformative, uniform distribution.
Can be detected: distribution of reports varies a lot over time.
⇒ penalize agents for such behavior.
More elegant solution: do not publish distribution R, but
derive it from multiple answers: PTSC.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 56/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Knowing Agent Beliefs

Mechanism design requires knowledge of agent beliefs (prior, prior

+ posterior).
Elicit beliefs through additional reports, example: Bayesian
Truth Serum.
Learn distributions from data; agents believe that mechanism
has correct observation ⇒ beliefs are identical to measured
distribution, example: Peer Truth Serum for Crowdsourcing.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 57/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Bayesian Truth Serum

obtain agent beliefs through an additional prediction report:

estimate of probability distribution of values in other agents’
reports.
prediction report indicates agents’ beliefs.

Prelec D. 2004. A bayesian truth serum for subjective data. Science, 34(5695):
462466.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 58/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Decomposable BTS mechanisms

Keep the decomposable structure of the score. Example:

1xi =xj 1X
τdecomp (xi , Fi , xj , Fj ) = + Fi (xj ) − Fi (z)2
Fj (xi ) 2 z
| {z } | {z }
information score prediction score

Requires additional constraint on agents’ beliefs:

y = argmaxz Pr (xi = y |xj = z)

Witkowski J., and Parkes D. 2012. A robust Bayesian truth serum for small
populations. AAAI 2012.
Radanovic G., and Faltings B. 2013. A robust Bayesian truth serum for
non-binary Signals. AAAI 2013.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 59/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Divergence-based BTS mechanism

Drawback of the original BTS:

requires a large number of agents
robust (decomposable) versions require additional constraints
Alternative approach: penalize agents for inconsistencies
Information score: penalize agents who have the same
information reports, while significantly different predictions.
Prediction score: score an agent’s posterior against a peer
report with a proper scoring rule.

Radanovic G., and Faltings B. Incentives for truthful information elicitation of

continuous signals. AAAI 2014.
Kong Y., and Schoenebeck G. Equilibrium selection in information elicitation
without verification via information monotonicity. Working paper 2016.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 60/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Multi-task crowdsourcing
2
1. Evaluate task
2. Update belief
Pr(sp,q|sw)
3. Report answer
4. Reward worker
sw Worker w
3
1
4
sp
Worker p
tasks tw, tp, tq Center
sq

Worker q

c Boi Faltings, Goran Radanovic Game Theory for Data Science 61/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Peer truth serum for crowdsourcing (PTSC)

Idea: collect R from agents’ reports, but keep it private.

R = histogram of reports from a set of many similar tasks,
e.g. multiple agents evaluate different airlines.
Peer report is chosen from reports on the same task.
Agent should believe that:
P ' R (in the limit of infinitely many tasks).
for its own task, q(x)/r (x) is maximized for its own
observation xi .
Radanovic G., et al. Incentives for effort in crowdsourcing using the peer truth
serum. ACM TIST 2016.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 62/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Algorithm (PTSC)
1 Collect answers to a set of
similar tasks T from
crowdworkers.
2 For worker w , calculate
Rw (x) = Pnum(x) , where
y num(y )
reports by worker w are
excluded.
3 For each task tw carried out by
worker w , select a peer worker p
that has solved the same task.
If they gave the same answer x,
reward w with
α · (1/Rw (x) − 1), otherwise
charge α.
c Boi Faltings, Goran Radanovic Game Theory for Data Science 63/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Example (PTSC)

Task Answers g
t1 b,a,a,c a
t2 b,b,b,a b
Probability of different answers
t3 a,a,b,a a
across all tasks:
t4 a,d ,a,a a
t5 c ,c ,a,b c Answer a b c d
t6 d ,a,d ,d d Count 20 12 4 4
t7 a,a,c ,a a R 0.50 0.30 0.1 0.1
t8 b,b,a,b b
t9 a,a,a,a a
t10 b,b,a,b b

c Boi Faltings, Goran Radanovic Game Theory for Data Science 64/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Example (PTSC)

Consider an agent ai who solves t7 and has xi = a.

Suppose p(x) ← R(x) and q(x) ← freq(x|a): self-predicting
condition satisfied!
Expected payoffs:
honest, report a:
E [pay (a)] = 0.75
0.5 − 1 =
1
2
strategic, report c:
0.1
E [pay (a)] = 0.1 −1=0
random, report according to r :
E [pay ([0.5, 0.3, 0.1, 0.1])] =
0.5 · 0.75 0.1 0.1
0.5 + 0.3 · 0.3 + 0.1 · 0.1 + 0.1 ·
0.05
0.1 −1=0

c Boi Faltings, Goran Radanovic Game Theory for Data Science 65/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Example (PTSC)
Probability of different answers across tasks with the same answer:
Correct Observed answer
answer a b c d
a Count(a) 15 2 2 1
freq(·|a) 0.75 0.1 0.1 0.05
b Count(b) 3 9 0 0
freq(·|b) 0.25 0.75 0 0
c Count(c) 1 1 2 0
freq(·|c) 0.25 0.25 0.5 0
d Count(d) 1 0 0 3
freq(·|d) 0.25 0 0 0.75
Count 20 12 4 4
R 0.5 0.3 0.1 0.1
⇒ for each task, reporting correct answer has highest prob. of
matching peer and payoff!
c Boi Faltings, Goran Radanovic Game Theory for Data Science 66/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Properties (PTSC)

Truthful equilibrium when agents’ beliefs satisfy self-predicting

condition.
Expected payoff = 0 for heuristic reporting, e.g., random
answers according to R
Truthful equilibrium has the highest payoff.
Agents do not need to have common prior distribution.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 67/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Large number of peers - Log PTS

If each agent has a large number of peers, the self-predicting

condition is not needed.
Logarithmic Peer Truth Serum:

freqlocal (xagent )
τ (xagent , ...) = ln
Rw (xagent )

where freqpeers (xagent ) is a (normalized) frequency of reports

equal to xagent among the peers who solve the same task.
Radanovic G. and Faltings B. Incentive Schemes for Participatory Sensing.
AAMAS 2015.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 68/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Empirical evaluation

Pollution model of Strassbourg (France):

Based on actual measurements, discretized to 4 values.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 69/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Dependence on number of Tasks

Few tasks ⇒ distribution RW is noisy.

⇒ truthful incentive may be weakened.
Min. number of tasks depends on confidence of worker:
 
q(x)
minx −1
p(x)

c Boi Faltings, Goran Radanovic Game Theory for Data Science 70/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Accurate reports pay off

Collusive and inaccurate reporting strategies are worse than

accurate reporting.
Random reports carry no payoff.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 71/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Incentives per sensor

Sensors have different payoffs (depending on how much

pollution varies).
For each and every sensor, reporting accurate and truthful
data is better than other strategies.
c Boi Faltings, Goran Radanovic Game Theory for Data Science 72/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Robustness to small populations

50 50
45 45
40 40
Avg. payoff (median)

Avg. payoff (median)

35 35
30 30
25 honest 25 honest
inaccurate inaccurate
20 20
collude collude
15 15
10 10
5 5
0 0
100/13 80/11 60/9 40/7 100/13 80/11 60/9 40/7
Num. sensors/Num. peers Num. sensors/Num. peers

Payoffs for different strategy profiles: PTSC (left) vs. Log PTS (right).

Both PTSC and Log PTS encourage truthful reporting

Log PTS is more sensitive to the decrease in the number of
sensors/peers

c Boi Faltings, Goran Radanovic Game Theory for Data Science 73/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Heterogenous Agent Beliefs

Agents answer multiple tasks and use the same strategy

everywhere.
Agents and center know and agree on sign of correlation
among each answer pair for different agents/same task.
Nothing else is known about agent beliefs.
Distinguishing correlated values is not important.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 74/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Correlated Agreement
Rewards are given through comparison of report x with a randomly
chosen peer’s answer y .
Idea 1: base payment on correlation matrix ∆ of signals:
∆(x, y ) = Pr (x, y ) − Pr (x)Pr (y ).
Define score for agent report x, peer report y as:

1 if ∆(x, y ) > 0
S(x, y ) =
0 otherwise

Idea 2: compare scores x, y for same task t1 with score for

randomly chosen different tasks using reports v of agent for t2
and w of peer agent for t3 :

Pay (x, y ) = S(x, y ) − S(v , w )

c Boi Faltings, Goran Radanovic Game Theory for Data Science 75/92
 Eliciting Truthful Information
Output agreement
Verfiable information
Parametric mechanisms
Unverifiable information
Non-parametric mechanisms
Decentralized machine learning

Correlated Agreement (2)

Expected payment for truthful reporting is the sum of all

positive entries in ∆:
X X
E [pay ] = ∆(xi , xj )S(xi , xj ) = ∆(xi , xj )
i,j i,j,∆(xi ,xj )>0

Non-truthful strategies would sum other elements: can only

achieve smaller sum.
Truthful strategies result in highest-paying equilibrium!

Dasgupta, A. and Ghosh, A.. Crowdsourced judgement elicitation with

endogenous proficiency. WWW 2013
Shnayder, V., Agarwal, A., Frongillo, R. and Parkes, D. Informed Truthfulness
in Multi-Task Peer Prediction. EC 2016
c Boi Faltings, Goran Radanovic Game Theory for Data Science 76/92
 Eliciting Truthful Information Group dynamics
Verfiable information Self-selection
Unverifiable information Scaling positive incentives
Decentralized machine learning Agent selection

Managing the Information Agents

Group dynamics: learning in repeated applications.

Self selection.
Low-quality signals.
Agent selection.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 77/92
 Eliciting Truthful Information Group dynamics
Verfiable information Self-selection
Unverifiable information Scaling positive incentives
Decentralized machine learning Agent selection

Group Dynamics

Peer-based mechanisms assume that argents coordinate

through a signal observed from the phenomenon.
In repeated elicitations with the same peers, agents may learn
other heuristic strategies (e.g. always report the same value).
Can be studied using replicator dynamics.
Output agreement/peer prediction vulnerable, but CA and
PTSC are not.
Gao, A., Mao, A., Chen, Y/ and Adams,R. Trick or Treat: Putting Peer
Prediction to the Test, EC 2014
Shnayder, V., Frongillo, R. and Parkes, D. Measuring performance of peer
prediction mechanisms using replicator dynamics, IJCAI-16

c Boi Faltings, Goran Radanovic Game Theory for Data Science 78/92
 Eliciting Truthful Information Group dynamics
Verfiable information Self-selection
Unverifiable information Scaling positive incentives
Decentralized machine learning Agent selection

Self-selection

Center can create incentives for measurements of uncertain

signals.
However, center does not know what it doesn’t know!
Self-selection: agents decide themselves what to measure and
contribute.
Requires that mechanism gives an incentive to measure the
interesting (uncertain) signals, and to provide accurate values.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 79/92
 Eliciting Truthful Information Group dynamics
Verfiable information Self-selection
Unverifiable information Scaling positive incentives
Decentralized machine learning Agent selection

2 Scenarios for Comparison

Novelty: posterior indicates a different value from the prior:

P1 = (0.1, 0.8, 0.1), Q1 = P1 vs.

P2 = (0.1, 0.8, 0.1), Q2 = (0.8, 0.1, 0.1)

Precision: lower precision (3 values) vs. higher precision (5

values):

P3 = (0.3, 0.4, 0.3), Q3 = (0.1, 0.8, 0.1)vs.

P4 = (0.1, 0.2, 0.4, 0.2, 0.1) Q4 = (0.05, 0.1, 0.7, 0.10.05)

c Boi Faltings, Goran Radanovic Game Theory for Data Science 80/92
 Eliciting Truthful Information Group dynamics
Verfiable information Self-selection
Unverifiable information Scaling positive incentives
Decentralized machine learning Agent selection

Expected Payments
Mechanism Expected Payment Novelty Precision
Truth Matching (value) maxx q(x) − maxx p(x) 0 vs. 0 0.4 vs. 0.4
Truth Matching (log rule) H(P) − H(Q) 0 vs. 0 0.648 vs. 0.728
Truth Matching (quadratic rule) λ(Q) − λ(P) 0 vs. 0 0.32 vs. 0.28
Output Agreement maxx q(x) − maxx p(x) 0 vs. 0 0.4 vs. 0.4
Peer Prediction (log rule) H(P) − H(Q) 0 vs. 0 0.648 vs. 0.728
Peer Prediction (quadratic rule) λ(Q) − λ(P) 0 vs. 0 0.32 vs. 0.28
Peer Truth Serum maxx γ(x) 0 vs. 7 1 vs. 1.33
Correlated Agreement maxx [q(x) − p(x)] 0 vs. 0.7 0.4 vs. 0.4
PTS for Crowdsourcing maxx γ(x) 0 vs. 7 1 vs. 1.33
Logarithmic PTS DKL (Q||P) 0 vs. 2.1 0.483 vs. 0.492
Bayesian Truth Serum H(P) − H(Q) 0 vs. 0 0.648 vs. 0.728
Divergence-based BTS (log) H(P) − H(Q) 0 vs. 0 0.195 vs. 0.221
Divergence-based BTS (quadratic) λ(Q) − λ(P) 0 vs. 0 0.32 vs. 0.28
P
H(P) = P − x p(x) log p(x) (Shannon Entropy),
λ(P) = x p(x)2 (Simpson’s diversity index),
γ(x) = q(x)/p(x) − 1 (Confidence).

c Boi Faltings, Goran Radanovic Game Theory for Data Science 81/92
 Eliciting Truthful Information Group dynamics
Verfiable information Self-selection
Unverifiable information Scaling positive incentives
Decentralized machine learning Agent selection

Novelty scenario

Value does not change (P1 /Q1 ): all schemes have expected
reward = 0.
Value changes (P2 /Q2 ): only PTS, CA, PTSC and LPTS
provide an expected incentive!
⇒ with other mechanisms, agents would not want to measure
novel data!
Center has to provide extra incentives ⇒ center has to know
what it doesn’t know.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 82/92
 Eliciting Truthful Information Group dynamics
Verfiable information Self-selection
Unverifiable information Scaling positive incentives
Decentralized machine learning Agent selection

Precision scenario

Mechanisms with constant rewards (truth matching, output

agreement, correlated agreement) are incentive-neutral.
Mechanisms based on quadratic scoring rule discourage
precision!
Mechanisms based on log. scoring rule (including PTS) give
incentive for higher precision.
Incentives for precision are also important to discourage reporting
low-quality signals.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 83/92
 Eliciting Truthful Information Group dynamics
Verfiable information Self-selection
Unverifiable information Scaling positive incentives
Decentralized machine learning Agent selection

Exploiting self-selection

Self-selection is an important idea: only information agents

know what the center does not know.
However, only some of the schemes provide the right
incentives.
Further research required to focus on this design criterion as
well.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 84/92
 Eliciting Truthful Information Group dynamics
Verfiable information Self-selection
Unverifiable information Scaling positive incentives
Decentralized machine learning Agent selection

Avoiding low-quality signals

Agents could collude to report something else than the true

signal.
Example: hash task description into the answer space ⇒
answers depend on the task, but not in the right way.
Requires coordination among information agents; feasible only
in some scenarios.
Best counter: spot-check with trusted reports and penalize
disagreement (as in influence limiter).

c Boi Faltings, Goran Radanovic Game Theory for Data Science 85/92
 Eliciting Truthful Information Group dynamics
Verfiable information Self-selection
Unverifiable information Scaling positive incentives
Decentralized machine learning Agent selection

Scaling incentives...

Traditional approach — scale so that:

1 Expected payments for uninformed reporting = 0.
2 Expected payments for accurate reporting > 0.
3 Possible to learn a scaling parameter.*

Two drawbacks:
1 Requires negative payments.
2 Susceptible to large noise.

*Liu, Y. and Chen Y. Learning to incentivize : eliciting effort via output

agreement. IJCAI 2016

c Boi Faltings, Goran Radanovic Game Theory for Data Science 86/92
 Eliciting Truthful Information Group dynamics
Verfiable information Self-selection
Unverifiable information Scaling positive incentives
Decentralized machine learning Agent selection

Traditional approach — susceptibility to noise

Crowdsourcing Community sensing

payment
time

noisy peer answers noisy measurements

How to boost the difference between the payments? — apply the
reputation based approach!
Radanovic, G. and Faltings, B. Learning to scale payments in crowdsourcing
with PropeRBoost. HCOMP 2016

c Boi Faltings, Goran Radanovic Game Theory for Data Science 87/92
 Eliciting Truthful Information Group dynamics
Verfiable information Self-selection
Unverifiable information Scaling positive incentives
Decentralized machine learning Agent selection

Learning to scale payments

Agents interact with a mechanisms over time.

Use reputations to track the quality of reported information.
The reputation is based on the quality score calculated by a
peer consistency approach.
Peer can be an output of a ”truth estimator” — θ̂F .
Quality score:

πt (x) = 1θ̂F =x − Pr (θ̂F = x)

scoret (x) = (1 − α) · πt (x) − α

Pr (θ̂F = x) can be estimated, α determines the minimal

acceptable quality.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 88/92
 Eliciting Truthful Information Group dynamics
Verfiable information Self-selection
Unverifiable information Scaling positive incentives
Decentralized machine learning Agent selection

ProperBoost — performance results

Crowdsourcing Community sensing
payment

payment
time time

scale
scale

time

c Boi Faltings, Goran Radanovic Game Theory for Data Science 89/92
 Eliciting Truthful Information Group dynamics
Verfiable information Self-selection
Unverifiable information Scaling positive incentives
Decentralized machine learning Agent selection

Agent and peer selection

For distributed agents, important to define:

1 Possible peers of each agent:

Peer can be an output of a ”truth estimator”

Applying machine learning to obtain an unbiased estimator

2 Agent selection under limited budget:

Each selected agent must have a ”good” peer
Leads to a constrained subset selection problem

Liu, Y. and Chen Y. Machine-learning aided peer prediction. EC 2017

Radanovic, G., Singla, A., Krause, A., and Faltings, B. Information
Gathering with Peers: Submodular Optimization with Peer-Prediction
Constraints. AAAI 2018

c Boi Faltings, Goran Radanovic Game Theory for Data Science 90/92
 Eliciting Truthful Information Group dynamics
Verfiable information Self-selection
Unverifiable information Scaling positive incentives
Decentralized machine learning Agent selection

Conclusions

Game theory allows to make payments for data depend on

accuracy:
1 Dominant strategies for verifiable information.
2 Strongly truthful equilibria for unverifiable information.
Non-parametric mechanisms allow heterogeneous agent
beliefs, under some conditions.
Managing agents to avoid collusion can be important.

c Boi Faltings, Goran Radanovic Game Theory for Data Science 91/92
 Eliciting Truthful Information Group dynamics
Verfiable information Self-selection
Unverifiable information Scaling positive incentives
Decentralized machine learning Agent selection

To read more

Boi Faltings and Goran Radanovic:

Game Theory for Data Science:
Eliciting Truthful Information,
Morgan & Claypool Publishers, 2017.
15% discount with code: authorcoll

c Boi Faltings, Goran Radanovic Game Theory for Data Science 92/92