analysis/skeleton-modular-conformal.R

In shamindras/robustps: What the Package Does (One Line, Title Case)

# 2 Non-robust prediction sets -------------------------------------------------

## 2.1 work with regular calibration for uncontaminted data

## 2.2 typically fail to provide the right coverage under data contamination

## 2.3 in the worst case could have severe coverage failure

# 3 Naive coverage expansion robust prediction sets ----------------------------

## 3.1 provides the right coverage under contamination

## 3.2 but can be arbitrarily bad in terms of size in the worst case

# 4 Projection (on true class) + expansion robust prediction sets --------------

## 4.1 also provides the right coverage under contamination

## 4.2 in addition, have better sizes compared to naive expansion sets

## 4.3 but projection step is computationally infeasible in general

# 5 Practical projection via outlier removal + coverage expansion --------------

## 5.1 provides the right coverage under structured contamination

## 5.2 and have good sizes compared to true projection + expansion sets

## 5.3 different outlier removal depending on different structural assumptions