Maximum sampled conditional likelihood for informative subsampling
Authors: HaiYing Wang, Jae Kwang Kim
JMLR 2022 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Numerical experiments are provided to evaluate the practical performance of the proposed method. |
| Researcher Affiliation | Academia | Hai Ying Wang EMAIL Department of Statistics University of Connecticut Storrs, CT 06269, USA Jae Kwang Kim EMAIL Department of Statistics Iowa State University Ames, IA 50011, USA |
| Pseudocode | No | The paper discusses various algorithms such as the Fisher-scoring algorithm and Newton's method, but it does not present any of them in a structured pseudocode or algorithm block. The steps are described within the regular text. |
| Open Source Code | No | The paper states: "We implemented all the algorithm in Julia (Bezanson et al., 2017) on a Desktop running Ubuntu 20.04." It mentions using Flux.jl but does not provide any explicit statement or link to the source code for the methodology described in the paper. |
| Open Datasets | Yes | Real data example: cover type data (Blackard and Dean, 1999). This dataset contains N = 581,012 observations on ten quantitative variables... To demonstrate the performance of the MSCLE, we applied it to the cover type data (Blackard and Dean, 1999)... In this section, we illustrate the advantage of the MSCLE over the IPW estimator using the famous MNIST data that is available at http://yann.lecun.com/exdb/mnist/. |
| Dataset Splits | Yes | We set the full data sample size N = 10^6, and let the subsample sizes be n = 500; 1000; 1500; and 2000... We used a smaller number of iterations here because the variations of the computational costs across different repetitions are much smaller than that of the estimators... The data has a training set with 60,000 instances and a testing set with 10,000 instances... We use a subsample of average size n = 5,000 out of N = 60,000 (about 8.3% of the training data) to train the model. The pilot probabilities p(xi, θplt) s are obtained from a pilot model trained with 2,000 uniformly selected instances from the training set. |
| Hardware Specification | Yes | We implemented all the algorithm in Julia (Bezanson et al., 2017) on a Desktop running Ubuntu 20.04. We restricted all the calculations to use one thread of the CPU with a base frequency of 2,200 megahertz and a maximum boosted frequency of 4,549 megahertz. |
| Software Dependencies | No | We implemented all the algorithm in Julia (Bezanson et al., 2017) on a Desktop running Ubuntu 20.04... We implement the convolutional neural network Le Net-5 (Le Cun et al., 1998) with Flux.jl (Innes, 2018). While Julia and Flux.jl are mentioned, specific version numbers for these software components are not provided. |
| Experiment Setup | Yes | We set the full data sample size N = 10^6, and let the subsample sizes be n = 500; 1000; 1500; and 2000. We assume that the responses have three possible categories (K = 3), and let the dimension of the covariates xi = (1, x T 1,i)T s be d = 4 where the first element of one is for the intercept parameters... We repeat the simulation for R = 1000 times and calculate the empirical mean squared error (MSE)... We used a smaller number of iterations here because the variations of the computational costs across different repetitions are much smaller than that of the estimators. Results for case (a) with multivariate normal covariates are reported in Table 1. |