Periodic Reporting for period 4 - AMDROMA (Algorithmic and Mechanism Design Research in Online MArkets)
Okres sprawozdawczy: 2023-01-01 do 2024-06-30
We consider algorithmic questions that arise in three fields of application: 1. Algorithms and mechanisms for web Internet advertisement; 2. Algorithms for Internet Markets; 3. Algorithms for online labor marketplaces. These problems of great importance for the Internet economy and for society require an interdisciplinary approach and tools for a wide variety of research. In a first project objective we address fundamental and longstanding open questions in auction and market design that are non-trivial due to the very large scale, and numerous practical restrictions. In a second projective objective, we will address the research issues of copying with uncertainty in online markets. In a third Project Objective, we address research issues of large-scale data analysis and optimization for two-sided matching and clustering problems which provide fundamental primitives for many online market applications.
The project led to the publication of approximately 125 papers, including contributions in top-tier conferences such as ACM STOC (6), ACM-SIAM SODA (5), EC (5), ICALP (2), NeurIPS (9), ICML (5), COLT (1), ICLR (1), AAAI (6), IJCAI (7), KDD (1), The Web Conference (4), CIKM (2), SPAA (1), WINE (5), and WSDM (1), as well as in leading academic journals.
A key contribution in the first project objective was the design of algorithms and mechanisms for two-sided markets involving strategic buyers and sellers. Traditionally, economic models assume full market knowledge, but we developed simple, efficient mechanisms that require minimal information—just a single sample from prior distributions representing buyers and sellers. Another notable problem we addressed is devising an online efficient and truthful mechanism for two-sided matching markets, where bidders arrive over time, and items must be allocated immediately. We provided an optimal truthful mechanism for such online matching markets.
The problem of fair division of items has a long history and is often studied under the assumption that agents reveal their true preferences. Recognizing that this isn't always realistic, we examined fair division through simple algorithms like round-robin in scenarios where agents can misreport valuations. Surprisingly, we found that the equilibria of this game still maintained strong fairness properties. Additionally, economic algorithms are increasingly used not only for pricing goods but also for motivating agents, both human and artificial, to complete tasks. We advanced algorithmic contract theory by designing new algorithms to encourage team effort on joint projects, incorporating inspection mechanisms to penalize non-compliant workers.
We also investigated the learning rate of bidding mechanisms for advertisers on online platforms using a first-price auction model. Our work fully characterizes the impact of the type of information disclosed by the platform—such as the valuation of the advertising opportunity and the competing bids—on the learning rate.
In the second project objective, we focused on decision-making algorithms under uncertainty in dynamic online environments. We revisited the classic Pandora’s Box problem, which models search strategies in information markets where there is a cost to explore each alternative. We developed simple, efficient strategies that are optimal or near-optimal when there are precedence constraints among alternatives, or when selecting combinations of alternatives with associated values.
We also tackled the classical scheduling problem of assigning tasks arriving online to a server to minimize average wait times. Traditionally, this problem assumes perfect information on processing times, but we developed the first efficient approximation algorithms that work under uncertainty about job processing times.
Additionally, over a sequence of papers, we studied the bilateral trade problem in two-sided markets where prior information is learned online over a sequence of interactions between buyers and sellers. Our work provides a comprehensive characterization of learning rates for different feedback models in both stochastic and adversarial settings.
For the third project objective, we focused on large-scale problems involving optimal or near-optimal matching, clustering, and submodular optimization, especially in online market applications. We developed efficient algorithms for matching freelance and hired workers to tasks in large-scale crowdsourcing systems. We also introduced the first algorithms capable of maintaining a near-maximum (1 + ε) approximate cardinality matching with O(1) update time during edge insertions. Furthermore, we showed that a data-oblivious random projection onto a smaller number of dimensions can approximate the cost of all k-means clusterings within a (1 ± ε) factor for large datasets in Euclidean space.
Submodular optimization is a critical component of many machine learning algorithms. We demonstrated that a simple randomized greedy approach can deliver fast, near-optimal solutions for maximizing a nonmonotone submodular function with a knapsack constraint, even in adaptive and stochastic settings. We also developed efficient algorithms for submodular optimization in dynamic, consistent, and fairness-oriented contexts.
All the results of the AMDROMA project are available on the project website: https://sites.google.com/a/uniroma1.it/stefanoleonardi/(odnośnik otworzy się w nowym oknie).
A key contribution in the first project objective was the design of algorithms and mechanisms for two-sided markets involving strategic buyers and sellers. Traditionally, economic models assume full market knowledge, but we developed simple, efficient mechanisms that require minimal information—just a single sample from prior distributions representing buyers and sellers. Another notable problem we addressed is devising an online efficient and truthful mechanism for two-sided matching markets, where bidders arrive over time, and items must be allocated immediately. We provided an optimal truthful mechanism for such online matching markets.
The problem of fair division of items has a long history and is often studied under the assumption that agents reveal their true preferences. Recognizing that this isn't always realistic, we examined fair division through simple algorithms like round-robin in scenarios where agents can misreport valuations. Surprisingly, we found that the equilibria of this game still maintained strong fairness properties. Additionally, economic algorithms are increasingly used not only for pricing goods but also for motivating agents, both human and artificial, to complete tasks. We advanced algorithmic contract theory by designing new algorithms to encourage team effort on joint projects, incorporating inspection mechanisms to penalize non-compliant workers.
We also investigated the learning rate of bidding mechanisms for advertisers on online platforms using a first-price auction model. Our work fully characterizes the impact of the type of information disclosed by the platform—such as the valuation of the advertising opportunity and the competing bids—on the learning rate.
In the second project objective, we focused on decision-making algorithms under uncertainty in dynamic online environments. We revisited the classic Pandora’s Box problem, which models search strategies in information markets where there is a cost to explore each alternative. We developed simple, efficient strategies that are optimal or near-optimal when there are precedence constraints among alternatives, or when selecting combinations of alternatives with associated values.
We also tackled the classical scheduling problem of assigning tasks arriving online to a server to minimize average wait times. Traditionally, this problem assumes perfect information on processing times, but we developed the first efficient approximation algorithms that work under uncertainty about job processing times.
Additionally, over a sequence of papers, we studied the bilateral trade problem in two-sided markets where prior information is learned online over a sequence of interactions between buyers and sellers. Our work provides a comprehensive characterization of learning rates for different feedback models in both stochastic and adversarial settings.
For the third project objective, we focused on large-scale problems involving optimal or near-optimal matching, clustering, and submodular optimization, especially in online market applications. We developed efficient algorithms for matching freelance and hired workers to tasks in large-scale crowdsourcing systems. We also introduced the first algorithms capable of maintaining a near-maximum (1 + ε) approximate cardinality matching with O(1) update time during edge insertions. Furthermore, we showed that a data-oblivious random projection onto a smaller number of dimensions can approximate the cost of all k-means clusterings within a (1 ± ε) factor for large datasets in Euclidean space.
Submodular optimization is a critical component of many machine learning algorithms. We demonstrated that a simple randomized greedy approach can deliver fast, near-optimal solutions for maximizing a nonmonotone submodular function with a knapsack constraint, even in adaptive and stochastic settings. We also developed efficient algorithms for submodular optimization in dynamic, consistent, and fairness-oriented contexts.
All the results of the AMDROMA project are available on the project website: https://sites.google.com/a/uniroma1.it/stefanoleonardi/(odnośnik otworzy się w nowym oknie).
Here are some of the most significant advancements beyond the state of the art achieved by the AMDROMA project:
AMDROMA conducted the first systematic investigation into the information-theoretic requirements for approximating social welfare in two-sided markets with minimal prior information. Additionally, it was the first to fully characterize the regret regimes for the fundamental bilateral trade problem under various feedback models and private valuations. The project also provided a complete analysis of the trade-off between learning rate and information transparency in repeated first-price auctions. The AMDROMA project solved for the first time the matching secretary problem with a near optimal truthful algorithm, thus solving a longstanding open problem in the area.
In the area of fair division mechanisms, AMDROMA was the first to demonstrate that fair allocations can be achieved at equilibrium using simple algorithms, even when agents misreport their valuations. Moreover, the AMDROMA provided the first efficient algorithms for incentivizing a set of agents in the algorithmic contract theory setting. Furthermore, the project introduced the first approximation algorithms for weighted flow time minimization that do not require precise knowledge of job processing times.
For Weitzman’s Pandora-box problem, AMDROMA developed novel threshold-based strategies that, surprisingly, work for precedence constraints in trees and forests.
In addressing computationally efficient methods for large-scale data analysis, the project showed that data-oblivious random projections onto a logarithmic number of dimensions are sufficient to approximate the cost of all k-means clusterings within a (1±ε) multiplicative factor, solving a longstanding open problem.
Lastly, the project made advancements in the maximum matching problem in the incremental setting by presenting the first deterministic algorithm that maintains a (1 + ε)-approximate matching with constant amortized update time per insertion.
AMDROMA conducted the first systematic investigation into the information-theoretic requirements for approximating social welfare in two-sided markets with minimal prior information. Additionally, it was the first to fully characterize the regret regimes for the fundamental bilateral trade problem under various feedback models and private valuations. The project also provided a complete analysis of the trade-off between learning rate and information transparency in repeated first-price auctions. The AMDROMA project solved for the first time the matching secretary problem with a near optimal truthful algorithm, thus solving a longstanding open problem in the area.
In the area of fair division mechanisms, AMDROMA was the first to demonstrate that fair allocations can be achieved at equilibrium using simple algorithms, even when agents misreport their valuations. Moreover, the AMDROMA provided the first efficient algorithms for incentivizing a set of agents in the algorithmic contract theory setting. Furthermore, the project introduced the first approximation algorithms for weighted flow time minimization that do not require precise knowledge of job processing times.
For Weitzman’s Pandora-box problem, AMDROMA developed novel threshold-based strategies that, surprisingly, work for precedence constraints in trees and forests.
In addressing computationally efficient methods for large-scale data analysis, the project showed that data-oblivious random projections onto a logarithmic number of dimensions are sufficient to approximate the cost of all k-means clusterings within a (1±ε) multiplicative factor, solving a longstanding open problem.
Lastly, the project made advancements in the maximum matching problem in the incremental setting by presenting the first deterministic algorithm that maintains a (1 + ε)-approximate matching with constant amortized update time per insertion.