LDA on Julia using MultivariateStats

I'm learning classification methods using the book of Brunton & Kutz "Data-Driven Science and Engineering", but instead of use only the MATLAB and Python code resources, i rewriting the textbook examples using Julia, because is my main programming language.

I can't find why fitting a MulticlassLDA model to the data doesn't work, it returns a DimensionMismatch("Inconsistent array sizes."), but as far as i can tell my arrays are dispatched to the fit function as indicated in the documentation.

This is my code:

using MAT, LinearAlgebra, Statistics, MultivariateStats

# Load data in MATLAB format. Abailible in http://www.databookuw.com/

dogs = read(matopen("../DATA/dogData_w.mat"),"dog_wave") 
cats = read(matopen("../DATA/catData_w.mat"),"cat_wave")
CD = hcat(dogs,cats)

u, s, v = svd(CD .- mean(CD)) #SVD decomposition

xtrain = vcat(v[1:60,2:2:4],v[81:140,2:2:4])   #training data array, dims 120x2
label = Int.(vcat(ones(60),-ones(60)))         #label's vector, length 120
xtest = vcat(v[61:80,2:2:4],v[141:160,2:2:4])  

classf= fit(MulticlassLDA,2,xtrain,label)

1 answer

  • answered 2021-06-23 06:49 Bogumił Kamiński

    You have two issues which are fixed this way:

    label = [fill(1, 60); fill(2, 60)] # labels must range from 1 to n
    fit(MulticlassLDA,2,permutedims(xtrain),label) # observations in xtrain must be stored in columns (not rows)

    See the comment in https://multivariatestatsjl.readthedocs.io/en/stable/index.html:

    All methods implemented in this package adopt the column-major convention of JuliaStats: in a data matrix, each column corresponds to a sample/observation, while each row corresponds to a feature (variable or attribute).

    And an explanation about y argument to fit https://multivariatestatsjl.readthedocs.io/en/stable/mclda.html#data-analysis:

    y – the vector of class labels, of length n. Each element of y must be an integer between 1 and nc.

    I hope this helps.