PYXIDA Institutional Repository
and Digital Library
Collections :

Title :Economics of Information Exchange in Learning systems
Creator :Σιδεράτος, Δημήτρης
Contributor :Κουτσόπουλος, Ιορδάνης (Επιβλέπων καθηγητής)
Publisher :Οικονομικό Πανεπιστήμιο Αθηνών
Type :Text
Extent :59σ.
Language :en
Bibliographic Citation :Βιβλιογραφία : σ. 57-59
Abstract :The ubiquitous adoption of mobile phone applications has transformed the owners of mobile devices into contributors of significant amounts of data, with or without their consent. Such data is usually collected and processed by application providers or other entities, towards various purposes. Data is a valuable asset whose acquisition incurs a benefit to the entity that collects it and a loss to the entity that reveals it. This situation is aggravated by growing concerns about the privacy of data owners. The idea that data owners should control the amount of data they provide and even be compensated for this data finds an increasing number of supporters, and hence, there is growing consensus to treat and trade private data in a market.In this thesis, we design a private-data market geared towards the machine-learning task of building a classifier. Namely, we consider a set of data owners, each with a data item in their possession, and a learner who is interested in buying data in order to train a classifier or to categorize users with an existing classifier.Each data owner has a private cost that quantifies their discomfort for providing their data to the learner. We also consider that at each stage of the learning process, each data owner is characterized by a utility score, which expresses the utility of the data item of that data owner for the learner.We consider the class of market mechanisms in which the learner receives the declared costs of data owners, it computes their utility scores and selects one owner to buy data from and the associated payment. The private cost of the data owners and their inclination to misreport it raises the needs to prevent them from doing so and to incentivize them to participate in the market. The objective of the learner may be either to minimize the privacy invasion to owners or to minimize its expected payment to owners, while maintaining a given expected added value to its task. For the former case, we propose a modified Vickrey-Clarke-Groves (VCG) auction, and for the latter case we formulate an optimal auction. In our approach the learner aims at efficiently building a classifier by enhancing its training data set.We proceed by addressing the multi-round sequential decision version of the problem of learning the classifier, in which at each round, the learner decides whether it will stop the learning process given the current classifier accuracy, or initiate another auction round. The problem amounts to weighing the current cost of classifier- 7 -inaccuracy against the expected (privacy or reimbursement) cost incurred through the auction, plus the expected cost of the new classifier accuracy; we cast this problem as an optimal stopping one. Our framework is generic and fits in a broad range of private-data scenarios.
Subject :Information management
Information services
Information technology
Communication industry
Date Issued :2014

File: Sideratos_2014.pdf

Type: application/pdf