pcoqs

Differentially Private Conformal Prediction
An R package for producing differentially private interval and set-valued predictions for regression and classification.

---

🔍 Overview

pcoqs provides functionality for conformal prediction enhanced with differential privacy. It supports:

• Regression prediction intervals
• Classification prediction sets (multiclass + binary)
• Custom nonconformity scores and model/predict wrappers
• Gaussian noise injection for private quantile estimation

---

🧪 Installation

# Install from GitHub (requires devtools)
devtools::install_github("SMAC-Group/pcoqs")

---

🚀 Quick Start

Regression Example

library(pcoqs)

set.seed(123)
X <- matrix(rnorm(100 * 3), ncol = 3)
Y <- X %*% c(1, -2, 1) + rnorm(100)

model <- lm(Y ~ ., data = data.frame(Y, X))

result <- pcoqs(model, X, Y, X, alpha = 0.1, rho = 1.0)
head(result$output)

Classification Example

library(pcoqs)
library(nnet)

set.seed(456)
X <- matrix(rnorm(200 * 3), ncol = 3)
probs <- t(apply(X, 1, function(row) {
  logits <- c(0, row[1], -row[2])
  exp_logits <- exp(logits)
  exp_logits / sum(exp_logits)
}))
Y <- apply(probs, 1, function(p) sample(1:3, 1, prob = p))

model <- multinom(as.factor(Y) ~ ., data = data.frame(Y = as.factor(Y), X), trace = FALSE)

result <- pcoqs(
  model, X, Y, X,
  alpha = 0.1, rho = 1.0,
  predict_fun = function(model, X, ...) predict(model, data.frame(X), type = "probs"),
  score_fun = function(y, probs) sapply(1:length(y), function(i) 1 - probs[i, y[i]]),
  output_fun = function(probs, q) {
    data.frame(predicted_set = apply(probs, 1, function(p) paste(which(p >= 1 - q), collapse = ",")))
  },
  lower_bound = 0, upper_bound = 1
)
head(result$output)

---

📄 Reference & Reproducibility

The methods implemented in this package are based on the paper:

Differentially Private Conformal Prediction via Quantile Binary Search
Ogonnaya Michael Romanus and Roberto Molinari
Transactions on Machine Learning Research (TMLR)

All Python code used to generate the experimental results in the paper (including simulations and real-data experiments) is available in the repository under the folder:

paper_codes/

This folder is provided for full reproducibility of the empirical results reported in the manuscript.

---

📘 Documentation

See function documentation using ?pcoqs, ?priv_quant, and ?noisy_rc once installed.

---

🛡️ License

MIT © Roberto Molinari

---