Strategic Classification Framework Accounts for Data Preferences Over Outcomes
-
Introduces the concept of strategic classification, which accounts for data points having preferences over classification outcomes. Presents a framework for modeling different types of preferences.
-
Formalizes the idea of data points manipulating feature vectors to achieve desired classification, balanced by cost functions limiting manipulation. Defines data point best response based on this balance.
-
Motivates strategic classification by noting real-world examples where data points influence outcomes, like spam filtering, disease testing, and college admissions.
-
Ultimately aims to find a strategic classifier that minimizes loss, where loss uses the data point best response instead of the original feature vector.
-
Lays groundwork for defining strategic VC dimension in next article, which generalizes the canonical VC dimension concept.