On Training-Conditional Conformal Prediction and Binomial Proportion Confidence Intervals

Authors: Rudi Coppola, Manuel Mazo Espinosa

TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate empirically equation 9 as follows and represent the results graphically in Figure 2. We define a list of values for E by Eq = 0.01 + 0.01 q for q = 0, ..., 98. For every value of Eq we consider an underlying Bernoulli distribution with parameter b1,q = Eq αEq < Eq (right figure) and an underlying Bernoulli distribution with parameter b2,q = E + αEq% > Eq (left figure) with α = 0.005. For every value of q = 0, ..., 98 we examine the two situations b1,q Eq and b2,q > Eq, as mentioned in Example 1 Part 2. The significance level ϵ is set to 2/3. We draw ncal = 5 104 pairs of calibration points {z(i) 1 , z(i) 2 }ncal i=1. For every pair of calibration points z(i) 1 , z(i) 2 we construct the resulting INP as Γϵ (i) .= Γϵ(z(i) 1 , z(i) 2 , ...), draw ntest = 5 104 test points {z(j) N+1}ntest i=j and compute the empirical frequency ˆgi = |{j=1,...ntest:z(j) N+1 Γϵ (i)}| ntest as an approximation for P(ZN+1 Γϵ (i)); finally we compute ˆh = |{i=j,...ncal:ˆgi 1 E}| ncal as an approximation to P2(SE) shown in the plots as the solid red line.
Researcher Affiliation Academia Rudi Coppola EMAIL Department of Mechanical Engineering Delft University of Technology Manuel Mazo Jr. EMAIL Department of Mechanical Engineering Delft University of Technology
Pseudocode No The paper describes methods and theoretical analyses but does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not contain any explicit statement about providing access to source code for the methodology described.
Open Datasets No For every value of Eq we consider an underlying Bernoulli distribution... We draw ncal = 5 104 pairs of calibration points {z(i) 1 , z(i) 2 }ncal i=1. For every pair of calibration points z(i) 1 , z(i) 2 we construct the resulting INP as Γϵ (i) .= Γϵ(z(i) 1 , z(i) 2 , ...), draw ntest = 5 104 test points {z(j) N+1}ntest i=j and compute the empirical frequency... This indicates data is generated for simulation, not from a public dataset.
Dataset Splits Yes We draw ncal = 5 104 pairs of calibration points {z(i) 1 , z(i) 2 }ncal i=1. For every pair of calibration points z(i) 1 , z(i) 2 we construct the resulting INP as Γϵ (i) .= Γϵ(z(i) 1 , z(i) 2 , ...), draw ntest = 5 104 test points {z(j) N+1}ntest i=j and compute the empirical frequency ˆgi = |{j=1,...ntest:z(j) N+1 Γϵ (i)}| ntest as an approximation for P(ZN+1 Γϵ (i)); finally we compute ˆh = |{i=j,...ncal:ˆgi 1 E}| ncal as an approximation to P2(SE) shown in the plots as the solid red line.
Hardware Specification No The paper does not provide specific details about the hardware used for running experiments.
Software Dependencies No The paper does not provide specific ancillary software details with version numbers.
Experiment Setup Yes We define a list of values for E by Eq = 0.01 + 0.01 q for q = 0, ..., 98. For every value of Eq we consider an underlying Bernoulli distribution with parameter b1,q = Eq αEq < Eq (right figure) and an underlying Bernoulli distribution with parameter b2,q = E + αEq% > Eq (left figure) with α = 0.005. The significance level ϵ is set to 2/3. We draw ncal = 5 104 pairs of calibration points {z(i) 1 , z(i) 2 }ncal i=1. For every pair of calibration points z(i) 1 , z(i) 2 we construct the resulting INP as Γϵ (i) .= Γϵ(z(i) 1 , z(i) 2 , ...), draw ntest = 5 104 test points {z(j) N+1}ntest i=j and compute the empirical frequency ˆgi = |{j=1,...ntest:z(j) N+1 Γϵ (i)}| ntest as an approximation for P(ZN+1 Γϵ (i)); finally we compute ˆh = |{i=j,...ncal:ˆgi 1 E}| ncal as an approximation to P2(SE) shown in the plots as the solid red line.