Weichao Mao, Zhenzhe Zheng, Fan Wu

Slides:



Advertisements
Similar presentations
Pattern Recognition and Machine Learning
Advertisements

1+eps-Approximate Sparse Recovery Eric Price MIT David Woodruff IBM Almaden.
Bayes rule, priors and maximum a posteriori
Yang Cai Oct 01, An overview of today’s class Myerson’s Auction Recap Challenge of Multi-Dimensional Settings Unit-Demand Pricing.
Chapter 6: Prior-free Mechanisms Roee and Ofir (Also from “Envy Freedom and Prior-free Mechanism Design” by Devanur, Hartline, Yan) 1.
On Optimal Single-Item Auctions George Pierrakos UC Berkeley based on joint works with: Constantinos Daskalakis, Ilias Diakonikolas, Christos Papadimitriou,
Prior-free auctions of digital goods Elias Koutsoupias University of Oxford.
CPS Bayesian games and their use in auctions Vincent Conitzer
Incentivize Crowd Labeling under Budget Constraint
Mechanism Design, Machine Learning, and Pricing Problems Maria-Florina Balcan.
An Approximate Truthful Mechanism for Combinatorial Auctions An Internet Mathematics paper by Aaron Archer, Christos Papadimitriou, Kunal Talwar and Éva.
Hybrid Keyword Auctions Kamesh Munagala Duke University Joint work with Ashish Goel, Stanford University.
A Prior-Free Revenue Maximizing Auction for Secondary Spectrum Access Ajay Gopinathan and Zongpeng Li IEEE INFOCOM 2011, Shanghai, China.
1 Regret-based Incremental Partial Revelation Mechanism Design Nathanaël Hyafil, Craig Boutilier AAAI 2006 Department of Computer Science University of.
Exploiting Sparse Markov and Covariance Structure in Multiresolution Models Presenter: Zhe Chen ECE / CMR Tennessee Technological University October 22,
Ch11 Curve Fitting Dr. Deshi Ye
6.853: Topics in Algorithmic Game Theory Fall 2011 Matt Weinberg Lecture 24.
Revenue Maximization in Probabilistic Single-Item Auctions by means of Signaling Joint work with: Yuval Emek (ETH) Iftah Gamzu (Microsoft Israel) Moshe.
EE462 MLCV Lecture Introduction of Graphical Models Markov Random Fields Segmentation Tae-Kyun Kim 1.
1. problem set 12 from Binmore’s Fun and Games. p.564 Ex. 41 p.565 Ex. 42.
Welfare and Profit Maximization with Production Costs A. Blum, A. Gupta, Y. Mansour, A. Sharma.
Dynamic Spectrum Management: Optimization, game and equilibrium Tom Luo (Yinyu Ye) December 18, WINE 2008.
Sequences of Take-It-or-Leave-it Offers: Near-Optimal Auctions Without Full Valuation Revelation Tuomas Sandholm and Andrew Gilpin Carnegie Mellon University.
Mechanisms for a Spatially Distributed Market Moshe Babaioff, Noam Nisan and Elan Pavlov School of Computer Science and Engineering Hebrew University of.
Machine Learning CUNY Graduate Center Lecture 3: Linear Regression.
Competitive Analysis of Incentive Compatible On-Line Auctions Ron Lavi and Noam Nisan SISL/IST, Cal-Tech Hebrew University.
Mechanisms for Making Crowds Truthful Andrew Mao, Sergiy Nesterko.
Multi-Unit Auctions with Budget Limits Shahar Dobzinski, Ron Lavi, and Noam Nisan.
Sequences of Take-It-or-Leave-it Offers: Near-Optimal Auctions Without Full Valuation Revelation Tuomas Sandholm and Andrew Gilpin Carnegie Mellon University.
Combinatorial Auctions By: Shai Roitman
1 ECE-517 Reinforcement Learning in Artificial Intelligence Lecture 7: Finite Horizon MDPs, Dynamic Programming Dr. Itamar Arel College of Engineering.
ECE 8443 – Pattern Recognition ECE 8423 – Adaptive Signal Processing Objectives: Deterministic vs. Random Maximum A Posteriori Maximum Likelihood Minimum.
By: Amir Ronen, Department of CS Stanford University Presented By: Oren Mizrahi Matan Protter Issues on border of economics & computation, 2002.
Yang Cai Oct 08, An overview of today’s class Basic LP Formulation for Multiple Bidders Succinct LP: Reduced Form of an Auction The Structure of.
Aemen Lodhi (Georgia Tech) Amogh Dhamdhere (CAIDA)
Regret Minimizing Equilibria of Games with Strict Type Uncertainty Stony Brook Conference on Game Theory Nathanaël Hyafil and Craig Boutilier Department.
Spectrum Trading in Cognitive Radio Networks: A Contract-Theoretic Modeling Approach Lin Gao, Xinbing Wang, Youyun Xu, Qian Zhang Shanghai Jiao Tong University,
Yang Cai Oct 06, An overview of today’s class Unit-Demand Pricing (cont’d) Multi-bidder Multi-item Setting Basic LP formulation.
A Passive Approach to Sensor Network Localization Rahul Biswas and Sebastian Thrun International Conference on Intelligent Robots and Systems 2004 Presented.
1 1 Slide © 2007 Thomson South-Western. All Rights Reserved Chapter 8 Interval Estimation Population Mean:  Known Population Mean:  Known Population.
Section 5.5 The Real Zeros of a Polynomial Function.
6.853: Topics in Algorithmic Game Theory Fall 2011 Constantinos Daskalakis Lecture 22.
Optimal Relay Placement for Indoor Sensor Networks Cuiyao Xue †, Yanmin Zhu †, Lei Ni †, Minglu Li †, Bo Li ‡ † Shanghai Jiao Tong University ‡ HK University.
Data Modeling Patrice Koehl Department of Biological Sciences
Carnegie Mellon University
Secretary Markets with Local Information
Near-Optimal Spectrum Allocation for Cognitive Radios: A Frequency-Time Auction Perspective Xinyu Wang Department of Electronic Engineering Shanghai.
To Whom the Revenue Goes: A Network Economic Analysis of the Price War in the Wireless Telecommunication Industry Vaggelis G. Douros, Petri Mähönen   Institute.
Probability Theory and Parameter Estimation I
LECTURE 09: BAYESIAN ESTIMATION (Cont.)
Privacy and Fault-Tolerance in Distributed Optimization Nitin Vaidya University of Illinois at Urbana-Champaign.
Course: Autonomous Machine Learning
Second degree price discrimination
Game Theory in Wireless and Communication Networks: Theory, Models, and Applications Lecture 6 Auction Theory Zhu Han, Dusit Niyato, Walid Saad, Tamer.
Qianyi Huang, Yixin Tao, and Fan Wu
Elasticity Considerations for Optimal Pricing of Networks
Exponential & Logarithmic Equations
Xinbing Wang*, Qian Zhang**
Pricing Model In Cloud Computing
10701 / Machine Learning Today: - Cross validation,
Economics and Computation Week #13 Revenue of single Item auctions
Robust Mechanism Design with Correlated Distributions
Cost-Volume-Profit Analysis and Planning
Parametric Methods Berlin Chen, 2005 References:
Exponential & Logarithmic Equations
Exponential & Logarithmic Equations
Chapter 8 Section 6 Solving Exponential & Logarithmic Equations
Applied Statistics and Probability for Engineers
Auction Theory תכנון מכרזים ומכירות פומביות
CPS Bayesian games and their use in auctions
Presentation transcript:

Pricing for Revenue Maximization in IoT Data Markets: An Information Design Perspective Weichao Mao, Zhenzhe Zheng, Fan Wu Shanghai Jiao Tong University, China May 2, 2019

Outline Motivation Preliminaries Pricing Mechanisms Evaluation Results Simple Setting: Single Buyer Type General Setting: Multiple Buyer Types Practical Considerations Evaluation Results Summary

Outline Motivation Preliminaries Pricing Mechanisms Evaluation Results Simple Setting: Single Buyer Type General Setting: Multiple Buyer Types Practical Considerations Evaluation Results Summary

Data Markets For targeted advertising: For city services: Gnip, Inc.: support.gnip.com HERE Technologies: www.here.com IOTA: www.iota.org DataBroker DAO: databrokerdao.com

Challenges: Time Sensitiveness Historical data vs. Future data

Challenges: Data Piracy Raw data vs. Data services Raw Data Raw Data

Challenges: Data Valuation Data Valuation independent of Data Volume

Design Objective Propose a market model to capture the unique economic properties of IoT data. Present pricing mechanisms that maximize seller revenue in the corresponding market model.

Outline Motivation Preliminaries Pricing Mechanisms Evaluation Results Simple Setting: Single Buyer Type General Setting: Multiple Buyer Types Practical Considerations Evaluation Results Summary

State & Action Nature state set Ω={ 𝜔 1 = , 𝜔 2 = } Action set Ω={ 𝜔 1 = , 𝜔 2 = } Action set 𝐴={ 𝑎 1 = , 𝑎 2 = }

Utility Matrix Buyer utility matrix: action 𝑎 𝑢 𝜔 1 , 𝑎 1 =𝟏 𝑢 𝜔 1 , 𝑎 2 =0 𝑢 𝜔 2 , 𝑎 1 =0 𝑢 𝜔 2 , 𝑎 2 =𝟏 state 𝜔 action 𝑎

Type & Prior Utility Buyer prior estimation, or type 𝜽= 𝜃 1 =0.7, 𝜃 2 =0.3 𝑃𝑟𝑜𝑏 𝜔= =0.7, 𝑃𝑟𝑜𝑏 𝜔= =0.3 Buyer prior utility 𝑢 1 = max 𝑎 𝐸 𝜔 𝑢 𝜔,𝑎 . 𝑢 1 =Prob ×𝑢 +Prob ×𝑢( ) =0.7×1+0.3×0= 𝟎.𝟕

Menu Seller publishes a menu of pricing schemes 𝑀= 𝐼, 𝑡 𝐼 Experiment 𝐼 and the corresponding price 𝑡 𝐼 In experiment ②, if 𝜔= 𝜔 1 : Send signal 𝑠 1 ( ) with 𝑝 𝑠 1 𝜔 1 =0.8, and send 𝑠 2 ( ) with 𝑝 𝑠 2 𝜔 1 =0.2

Bayesian Updating Buyer updates his belief and gets posterior estimation: 𝑝 𝜔 𝑖 𝑠 𝑗 = 𝑝 𝑠 𝑗 𝜔 𝑖 ⋅𝑝( 𝜔 𝑖 ) 𝑝( 𝑠 𝑗 ) Prob 𝜔= | 𝒔 𝟏 =0.86, Prob 𝜔= | 𝒔 𝟏 =0.14 Suppose buyer receives in experiment ②:

Buyer Valuation Buyer expected posterior utility: 𝑢 2 = 𝑗 𝑝( 𝑠 𝑗 ) max 𝑎 𝑖=1 𝑛 𝑝 𝜔 𝑖 𝑠 𝑗 ⋅𝑢( 𝜔 𝑖 ,𝑎) =𝟎.𝟕𝟕 Utility increment, or buyer valuation: 𝑣 𝜽, 𝐼 2 = 𝑢 2 − 𝑢 1 =𝟎.𝟎𝟕 Set price 𝑡 𝐼 =𝟎.𝟎𝟓, and get revenue 0.05 Individual rationality (I.R.): 𝑣=𝑢 2 − 𝑢 1 =0.07>0.05= 𝑡 𝐼

Two Important Experiments Full-information experiment: 𝐼 = 1 0 0 1 100% accurate data, leading to highest utility increment No-information experiment: 𝐼 = 0.5 0.5 0.5 0.5 Random data, leading to zero utility increment Alternative: let 𝑆 be a singleton In our example: 𝑣 𝜽, 𝐼 =0.3, leads to highest revenue.

Data Trading Process

Outline Motivation Preliminaries Pricing Mechanisms Evaluation Results Simple Setting: Single Buyer Type General Setting: Multiple Buyer Types Practical Considerations Evaluation Results Summary

The MSimple Mechanism Only one buyer type 𝜽 exists, and 𝜽 is known to the seller Theorem 1: For the single buyer type case, the optimal menu contains one pricing scheme, which is a full-information experiment with a price equal to the buyer’s valuation.

Outline Motivation Preliminaries Pricing Mechanisms Evaluation Results Simple Setting: Single Buyer Type General Setting: Multiple Buyer Types Practical Considerations Evaluation Results Summary

General Setting: Multiple Buyer Types Buyers have different prior estimations, and the seller cannot tell buyers apart. 𝑃𝑟𝑜𝑏 𝜔= =𝟎.𝟕, 𝑃𝑟𝑜𝑏 𝜔= =𝟎.𝟑 𝑃𝑟𝑜𝑏 𝜔= =𝟎.𝟒, 𝑃𝑟𝑜𝑏 𝜔= =𝟎.𝟔 Distribution of buyer types 𝐹 𝜽 ∈ΔΘ is public information.

Second Degree Price Discrimination A separate pricing scheme 𝐼 𝜃 , 𝑡 𝜃 for each type 𝜃∈Θ. 𝑃𝑟𝑜𝑏 𝜔= =𝟎.𝟕, 𝑃𝑟𝑜𝑏 𝜔= =𝟎.𝟑 𝑃𝑟𝑜𝑏 𝜔= =𝟎.𝟒, 𝑃𝑟𝑜𝑏 𝜔= =𝟎.𝟔 Incentive compatible (I.C.): 𝑣 𝜽, 𝐼 𝜽 − 𝑡 𝜽 ≥𝑣 𝜽, 𝐼 𝜽 ′ − 𝑡 𝜽 ′ , ∀𝜽, 𝜽 ′ .

The MGeneral Mechanism Theorem 2: For multiple buyer types, the revenue maximizing menu can be solved in polynomial time of |Ω| and Θ , by solving a convex program.

Back to the Example The revenue-maximizing menu: 𝑃𝑟𝑜𝑏 𝜔= =𝟎.𝟕, 𝑃𝑟𝑜𝑏 𝜔= =𝟎.𝟑 𝑃𝑟𝑜𝑏 𝜔= =𝟎.𝟒, 𝑃𝑟𝑜𝑏 𝜔= =𝟎.𝟔 Extracting full surplus from , but not from . The optimal revenue is 0.6

Practical Considerations The weaknesses of MGeneral? Price discrimination is unfair. Heavy computation. Menu size as big as |Θ|.

Outline Motivation Preliminaries Pricing Mechanisms Evaluation Results Simple Setting: Single Buyer Type General Setting: Multiple Buyer Types Practical Considerations Evaluation Results Summary

The MPractical Mechanism Consider a constant-size menu: where 𝑡 is the optimal fixed price:

Performance of MPractical Theorem 3: MPractical extracts at least a logarithmic fraction Ω( 1 log|Θ| ) of the full surplus even in the worst case. Theorem 4: No constant-size menu can extract more than a logarithmic fraction 𝑂( 1 log|Θ| ) of the full surplus in the worst case.

Back to the Example The optimal menu: 𝑃𝑟𝑜𝑏 𝜔= =𝟎.𝟕, 𝑃𝑟𝑜𝑏 𝜔= =𝟎.𝟑 𝑃𝑟𝑜𝑏 𝜔= =𝟎.𝟒, 𝑃𝑟𝑜𝑏 𝜔= =𝟎.𝟔 Extracting full surplus from , but not from . The optimal revenue is 0.6

Outline Motivation Preliminaries Pricing Mechanisms Evaluation Results Simple Setting: Single Buyer Type General Setting: Multiple Buyer Types Practical Considerations Evaluation Results Summary

Evaluation Results Varying |Θ|:

Outline Motivation Preliminaries Pricing Mechanisms Evaluation Results Simple Setting: Single Buyer Type General Setting: Multiple Buyer Types Practical Considerations Evaluation Results Summary

Summary We characterize the unique economic properties of IoT data, and propose a market model from an information design perspective to capture these properties. We extract full surplus in the simple setting, and formulate the revenue maximization problem in the general setting as a polynomial convex program. We consider a more practical setting, and propose a constant size mechanism that achieves tight logarithmic approximation ratio.

Thank you!