HurdleDMR.jl is a Julia implementation of the Hurdle Distributed Multinomial Regression (HDMR), as described in:
Kelly, Bryan, Asaf Manela, and Alan Moreira (2018). Text Selection. Working paper.
It includes a Julia implementation of the Distributed Multinomial Regression (DMR) model of Taddy (2015).
This tutorial explains how to use this package.
First, install Julia itself. The easiest way to do that is from the download site https://julialang.org/downloads/. An alternative is to install JuliaPro from https://juliacomputing.com
Once installed, open julia in a terminal (or in Juno), press ] to activate package manager and add the following packages:
pkg> add HurdleDMRAdd parallel workers and make package available to workers
using Distributed
addprocs(4)
import HurdleDMR; @everywhere using HurdleDMR
Setup your data into an n-by-p covars matrix, and a (sparse) n-by-d counts matrix. Here we generate some random data.
if not installed, install following package with
pkg> add CSV GLM DataFrames Distributionsusing CSV, GLM, Lasso, DataFrames, Distributions, Random, LinearAlgebra, SparseArrays
n = 100
p = 3
d = 4
Random.seed!(13)
m = 1 .+ rand(Poisson(5),n)
covars = rand(n,p)
ηfn(vi) = exp.([0 + i*sum(vi) for i=1:d])
q = [ηfn(covars[i,:]) for i=1:n]
rmul!.(q,ones(n)./sum.(q))
counts = convert(SparseMatrixCSC{Float64,Int},hcat(broadcast((qi,mi)->rand(Multinomial(mi, qi)),q,m)...)')
covarsdf = DataFrame(covars,[:vy, :v1, :v2])
The Distributed Multinomial Regression (DMR) model of Taddy (2015) is a highly scalable
approximation to the Multinomial using distributed (independent, parallel)
Poisson regressions, one for each of the d categories (columns) of a large counts matrix,
on the covars.
To fit a DMR:
m = dmr(covars, counts)
or with a dataframe and formula
mf = @model(c ~ vy + v1 + v2)
m = fit(DMR, mf, covarsdf, counts)
in either case we can get the coefficients matrix for each variable + intercept as usual with
coef(m)
By default we only return the AICc maximizing coefficients. To also get back the entire regulatrization paths, run
paths = fit(DMRPaths, mf, covarsdf, counts)
we can now select, for example the coefficients that minimize 10-fold CV mse (takes a while)
coef(paths, MinCVKfold{MinCVmse}(10))
For highly sparse counts, as is often the case with text that is selected for
various reasons, the Hurdle Distributed Multinomial Regression (HDMR) model of
Kelly, Manela, and Moreira (2018), may be superior to the DMR. It approximates
a higher dispersion Multinomial using distributed (independent, parallel)
Hurdle regressions, one for each of the d categories (columns) of a large counts matrix,
on the covars. It allows a potentially different sets of covariates to explain
category inclusion ($h=1{c>0}$), and repetition ($c>0$).
Both the model for zeroes and for positive counts are regularized by default,
using GammaLassoPath, picking the AICc optimal segment of the regularization
path.
HDMR can be fitted:
m = hdmr(covars, counts; inpos=1:2, inzero=1:3)
or with a dataframe and formula
mf = @model(h ~ vy + v1 + v2, c ~ vy + v1)
m = fit(HDMR, mf, covarsdf, counts)
where the h ~ equation is the model for zeros (hurdle crossing) and c ~ is the model for positive counts
in either case we can get the coefficients matrix for each variable + intercept as usual with
coefspos, coefszero = coef(m)
By default we only return the AICc maximizing coefficients. To also get back the entire regulatrization paths, run
paths = fit(HDMRPaths, mf, covarsdf, counts)
coef(paths, AllSeg())
A sufficient reduction projection summarizes the counts, much like a sufficient
statistic, and is useful for reducing the d dimensional counts in a potentially
much lower dimension matrix z.
To get a sufficient reduction projection in direction of vy for the above example
z = srproj(m,counts,1,1)
Here, the first column is the SR projection from the model for positive counts, the second is the the SR projection from the model for hurdle crossing (zeros), and the third is the total count for each observation.
Counts inverse regression allows us to predict a covariate with the counts and other covariates. Here we use hdmr for the backward regression and another model for the forward regression. This can be accomplished with a single command, by fitting a CIR{HDMR,FM} where the forward model is FM <: RegressionModel.
cir = fit(CIR{HDMR,LinearModel},mf,covarsdf,counts,:vy; nocounts=true)
where the nocounts=true means we also fit a benchmark model without counts.
we can get the forward and backward model coefficients with
coefbwd(cir)
coeffwd(cir)
The fitted model can be used to predict vy with new data
yhat = predict(cir, covarsdf[1:10,:], counts[1:10,:])
We can also predict only with the other covariates, which in this case is just a linear regression
yhat_nocounts = predict(cir, covarsdf[1:10,:], counts[1:10,:]; nocounts=true)
Kelly, Manela, and Moreira (2018) show that the differences between DMR and HDMR can be substantial in some cases, especially when the counts data is highly sparse.
Please reference the paper for additional details and example applications.