Replication-robust Profit Allocation for Data Exchange in ML
2025 (English)Independent thesis Basic level (degree of Bachelor), 10 credits / 15 HE credits
Student thesis
Abstract [en]
In the context of emerging data marketplaces for machine learning, fair valuation and profit-sharing mechanisms are critical to ensure that data contributors are properly incentivized and protected from strategic manipulations. This thesis explores the challenge of replication manipulation, where malicious participants replicate data to unfairly inflate their profit, undermining fairness in cooperative data-driven environments. Building on cooperative game theory, particularly semivalues such as the Shapley and Banzhaf values, we assess the fairness properties of these methods. To formalize the analysis, we examine the concept of replication robustness as a necessary fairness criterion, ensuring contributors cannot increase their allocated value through mere duplication. Simulations in a set cover game and machine learning data market demonstrate how various semivalues and penalization strategies respond to replication attacks. Our results confirm that while traditional semivalues like the Shapley value are vulnerable, alternatives such as the Banzhaf value, Leave-One-Out, and modified Shapley values achieve replication robustness, but often at the cost of fairness to non-malicious players. To address this, we propose a novel, two-step method utilizing penalty functions and rescaling, preserving replication robustness while protecting fair players. Despite limitations in similarity metrics like cosine similarity, our findings offer a promising direction for designing fair and robust allocation mechanisms in data marketplaces, highlighting the trade-offs between computational efficiency, robustness, and fairness. Future research is recommended to explore more sophisticated similarity measures and scalable approximation techniques.
Place, publisher, year, edition, pages
2025.
Series
TRITA-SCI-GRU ; 2025:193
Keywords [en]
Data marketplaces, semivalue, Shapley value, cooperative game theory
National Category
Mathematical sciences
Identifiers
URN: urn:nbn:se:kth:diva-365900OAI: oai:DiVA.org:kth-365900DiVA, id: diva2:1980199
Subject / course
Optimization and Systems Theory
Educational program
Master of Science in Engineering -Engineering Physics
Supervisors
Examiners
2025-07-012025-07-012025-07-01Bibliographically approved