by: npho3
Survey of supervised learning approaches to analysis of family health data from the National Vital Statistics System of the National Center for Health Statistics. See submitted paper npho3-analysis.pdf for full details.
Survey of randomized optimization methods including randomized hill climbing, simulated annealing, and genetic algorithms to determine weights in a neural network. Then contrived problems for which these algorithms will thrive. See submitted paper npho3-analysis.pdf for further details.
Survey of k-means and expectation maximization (EM) clustering algorithms with and without additional feature selection from PCA, ICA, randomized projection, and t-SNE. Both natality and mortality data sets were studied. See submitted paper npho3-analysis.pdf for full details.
The data sets are (1) natality data from every recorded birth in the US in 2017 and (2) the linked mortality statistics from 2012 for every child under 1 year of age who died that year. Each can be processed with the pre-processing scripts in this repository, make sure to edit the scripts to point to the appropriate location of the original data on your local computer. The pre-processing scripts made sure the columns are of the appropriate R data structure (e.g., factor, numeric) as well as manually curating the features by removing redundant or medically irrelevant (to our hypothesis) ones. Alternatively, explore the dat folder within the repository for the cleaned and prepared R objects saved as RData files. Note: make sure git lfs is configured locally to handle these larger files.
All standard R packages used should be able to be installed directly from CRAN without issue. The only exception is MXNet for which additional hacks are detailed below.
Used MXNet for artificial neural networks (ANNs), instructions for installing here. For macOS the instructions say to use homebrew to install OpenCV. However, by default v4 gets installed and MXNet only recognized v3 when I tested. Install opencv@3 then symbolically link the expected OpenCV path.
cran <- getOption("repos")
cran["dmlc"] <- "https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/R/CRAN/"
options(repos = cran)
install.packages("mxnet")For the randomized optimization assignment part 1 and using differing optimization algorithms for the neural network I migrated from R to Python. Use the Docker Jupyter notebooks container which have all the necessary libraries, the RandomizedOptimization.ipynb file installs mlrose automatically.
For the randomized optimization assignment part 2 some ABAGAIL code was required, Java code requires ant to build. I borrowed the implementation of the Traveling Salesman Problem, Countones, and Flipflop from @proudhuma that extends ABAGAIL for these experiments. Copy the java code into the ABAGAIL src folder then build with ant as usual. Run by using one of the various commands:
Each supervised learning approach lives within its own self-contained R script within the base repository. The code can be run by loaded an interactive R session then calling each script via source(). It wasn't explicitly tested but it's likely running it fully from the command line via Rscript or R CMD followed by the script path should work as well. Where possible the file paths are relative but it's possible if the code doesn't run that's the first place to look and modify to fit your local environment.
The final tuned models as determined by a grid-search using 4-fold cross-validation is in the models folder. Hyper-parameter tuning and learning curve plots for each model are within the img folder.
For part 1 in optimizing neural network weights you can run the command below to load the Docker container then open your browser and navigate to this link.
docker run -p 8888:8888 -v "$PWD":/home/jovyan/work/ jupyter/datascience-notebook
For part 2 you can select a test to run after compiling ABAGAIL with ant including the additional custom Java code.
java -cp ABAGAIL.jar mytest.TravelingSalesmanTest
java -cp ABAGAIL.jar mytest.FlipFlopTest
java -cp ABAGAIL.jar mytest.CountOnesTest
Several *.out files will be generated with the run data from which you can generate plots.
The k-means and expectation maximization (EM) clustering codes are available in kmc.R and em.R, respectively. All four feature selection approaches are found within the dimreduction.R file, including the run through the existing artificial neural network (ANN). Respective visualization is included within each file. Code can be run on the command-line or sourced within an interactie R prompt.