/lichen-revised-pls

Revised pls code and work flow.

Primary LanguageR

lichen-revised-pls

Revised pls code and work flow for the aging lichen project.

Scripts

generatePLSData is the main function that calls all other functions in this package. It takes in a spectra object, a className (unit of classification such as Family), a boolean for whether age should be included as covariate in the model, the initial number of components to start with, the number of iterations to run each PLS procedure for, and a directory for where files should be stored. This directory needs a setup as follows:

output
| - class (eg. family)
  |- no-age
    | - accuracies (accuracy rds files that are used to calculate optimal number of components)
    | - cm-final (confusion matrix figure)
    | - confusion-matrices (raw confusion matrices)
    | - metrics (rds file that stores average raw confusion matrix and confusion matrix of standard deviations, average overall accuracy, and number of components used)
    | - upsampled-modles (model objects generated during upsampling procedure)
    | - variable-importance (variable importance plots)
  | - with-age (same stucture as 'no-age' directory)

This function runs PLSDA using parallelization so that multiple cores on the machine can be used at a time in contrast to single threading. The function is set to use half of the available cores on your machine. The main products are the summary metrics (overall accuracy, number of components used), a confusion matrix, and variable importance plots that show the importance of the top 5 and bottom 5 components.

The functions called within this function are as follows:

  • plsda: This function runs PLSDA with 10-fold cross validation repeated three times. Seventy percent of the data is used to train the model, and 30% is used to test the model. The sampling is conducted with replacement. This means that each iteration of the PLSDA function (not to be confused by the 10-fold cross validation) splits the data into a random 70/30 split. This function is called twice within the generatePLSData function. The first time, the plsda function is set to run with downsampled data. This step generates accuracy metrics that are used to determine the optimal number of components to use in the next call of the plsda function with upsampled data. This second call generates model objects which are then used to make the final interpretations about the data.

  • getNComps: This function runs after the downsampled PLS and before the upsampled PLS. The main point of this function is to find the first component for which the average overall model accuracy is within two standard deviations of the component with the highest overall accuracy. The resulting number is the optimal number of components to use in the upsampled PLS, and is injected into the second pls function accordingly.

  • getData: This function loops through the model objects generated in the upsampled PLS procedure and returns:

    1. accuracies - a list of overall accuracies from each iteration
    2. overallAccuracy - an average overall accuracy
    3. cmMean - a confusion matrix that is the average from each iteration and is used to generate the final confusion matrix plot
    4. cmSD - a confusion matrix that contains the value of one standard deviation for each cell in matrix (standard deviations calculated from the collection of confusion matrices generated by all iterations of upsampled PLS)
    5. ncomps - the number chosen as the optimal number of components in the upsampled PLS

  • plotConfusionMatrix: This function creates a high resolution confusion matrix using the corrplot package and saves it (you won't see a plot in R's plot window).

  • plotVip: This function creates a high resolution bar chart that displays the five most important explanatory variables (wavelengths or the age covariate) and the five least important explanatory variables. This plot is also saved and you will not see it plotted in R's plot window.

The runAll script is simply a script that runs the generatePLSData function for each class with and without age as a covariate.