Strategyproof Linear Regression

Designing machine learning algorithms that are robust to noise in training data has lately been a subject of intense research. A large body of work addresses stochastic noise [12, 7], while another one studies adversarial noise [11, 2] in which errors are introduced by an adversary with the explicit purpose of sabotaging the algorithm. This is often too pessimistic, and leads to negative results. The literature on game theory and mechanism design offers an interesting middle ground: strategic noise. In this paradigm, training data is provided by strategic sources that purposefully introduce errors for maximizing their own benefit. This is less pessimistic than adversarial noise where the errors are introduced for simply harming the algorithm.