Classification Under Human Assistance

Most supervised learning models are trained for full automation. However, their predictions sometimes worse than those by human experts on some specific instances. Motivated this empirical observation, our goal is to design classifiers that optimized operate under different automation levels. More specifically, we focus convex margin-based and first show the problem NP-hard. Then, further that, support vector machines, corresponding objective function can be expressed as difference of two functions f = g - c, where monotone, non-negative gamma-weakly submodular, c modular. This representation allows a recently introduced deterministic greedy algorithm, well more efficient randomized variant enjoy approximation guarantees at solving problem. Experiments synthetic real-world data from several applications in medical diagnosis illustrate theoretical findings demonstrate assistance, levels outperform humans operating alone.