Online Learning of Acyclic Conditional Preference Networks from Noisy Data

The complexity of exact learning of acyclic conditional preference networks from swap examples

“…The algorithm assumes the user to provide explicit answers of the form "object o is preferred over object o ′ ", "object o ′ is preferred over object o", or "neither of the two objects is preferred over the other" to any query (o, o ′ ). Labernia et al [13] proposed another online learning algorithm based on swap observations where the latter can be noisy. It is assumed that the target CP-net represents the global preference for a group of users and the noise is due to variations of a user's preference compared to the global one.…”

“…To the best of our knowledge, the only studies of learning CP-nets in Angluin's query model, where the target concept is identified exactly, are one by Koriche and Zanuttini [7] and one by Labernia et al [22]. Koriche and Zanuttini assumed perfect oracles and investigated the problem of learning complete and incomplete bounded CP-nets from membership and equivalence queries over the swap instance space.…”

“…The setting considered in their work is more general than ours and exhibits the power of membership queries when it comes to learning CP-nets. Labernia et al [22] investigated the problem of learning an average CP-net from multiple users using equivalence queries alone. However, neither study addresses the problem of learning complete acyclic CP-nets from membership queries alone, whether corrupted or uncorrupted.…”

“…[12]. CP-nets have also been studied in models of passive learning from examples, both for batch learning [8,9,10,4] and for online learning [11,13].…”

“…A swap is a pair of objects that differ in the value of only a single attribute. Learning CP-nets over swap examples is an often studied scenario [7,8,22,23,13], which we adopt here for various reasons detailed in Section 4.…”