This vignette illustrates the standard use of the
PLNmixture
function and the methods accompanying the R6
Classes PLNmixturefamily
and
PLNmixturefit
.
The packages required for the analysis are PLNmodels plus some others for data manipulation and representation:
The main function PLNmixture
integrates some features of
the future package to perform parallel computing: you
can set your plan to speed the fit by relying on 2 workers as
follows:
We illustrate our point with the trichoptera data set, a full description of which can be found in the corresponding vignette. Data preparation is also detailed in the specific vignette.
The trichoptera
data frame stores a matrix of counts
(trichoptera$Abundance
), a matrix of offsets
(trichoptera$Offset
) and some vectors of covariates
(trichoptera$Wind
, trichoptera$Temperature
,
etc.)
PLN-mixture for multivariate count data is a variant of the Poisson Lognormal model of Aitchison and Ho (1989) (see the PLN vignette as a reminder) which can be viewed as a PLN model with an additional mixture layer in the model: the latent observations found in the first layer are assumed to be drawn from a mixture of \(K\) multivariate Gaussian components. Each component \(k\) has a prior probability \(p(i \in k) = \pi_k\) such that \(\sum_k \pi_k = 1\). We denote by \(C_i\in \{1,\dots,K\}\) the multinomial variable \(\mathcal{M}(1,\boldsymbol{\pi} = (\pi_1,\dots,\pi_K))\) describing the component to which observation \(i\) belongs to. Introducing this additional layer, our PLN mixture model is as follows
\[ \begin{array}{rcl} \text{layer 2 (clustering)} & \mathbf{C}\_i \sim \mathcal{M}(1,\boldsymbol{\pi}) \\ \text{layer 1 (Gaussian)} & \mathbf{Z}\_i | \, \mathbf{C}\_i = k \sim \mathcal{N}({\boldsymbol\mu}^{(k)}, {\boldsymbol\Sigma}^{(k)}), \\ \text{observation space } & Y_{ij} \| Z_{ij} \quad \text{indep.} & \mathbf{Y}_i | \mathbf{Z}_i\sim\mathcal{P}\left(\exp\{\mathbf{Z}_i\}\right). \end{array} \]
Just like PLN, PLN-mixture generalizes to a formulation where the main effect is due to a linear combination of \(d\) covariates \(\mathbf{x}_i\) and to a vector \(\mathbf{o}_i\) of \(p\) offsets in sample \(i\) in each mixture component. The latent layer then reads
\[ \mathbf{Z}_i | \mathbf{C}_i = k \, \sim \mathcal{N}({\mathbf{o}_i + \mathbf{x}_i^{\top}{\mathbf{B}} + \boldsymbol\mu}^{(k)},{\boldsymbol\Sigma}^{(k)}), \]
where \({\mathbf{B}}\) is a \(d\times p\) matrix of regression parameters common to all the mixture components.
When using parametric mixture models like Gaussian mixture models, it is generally not recommended to have covariances matrices \({\boldsymbol\Sigma}^{(k)}\) with no special restriction, especially when dealing with a large number of variables. Indeed, the total number of parameters to estimate in such unrestricted model can become prohibitive.
To reduce the computational burden and avoid over-fitting the data,
two different, more constrained parametrizations of the covariance
matrices of each component are currently implemented in the
PLNmodels
package (on top of the general form of \(\Sigma_k\)):
The diagonal structure assumes that, given the group membership of a site, all variable abundances are independent. The spherical structure further assumes that all species have the same biological variability. In particular, in both parametrisations, all observed covariations are caused only by the group structure.
For readers familiar with the mclust
R
package (Fraley and
Raftery 1999), which implements Gaussian mixture models with
many variants of covariance matrices of each component, the spherical
model corresponds to VII
(spherical, unequal volume) and
the diagonal model to VVI
(diagonal, varying volume and
shape). {Using constrained forms of the covariance matrices enables}
PLN-mixture to {provide a clustering} even when the number of sites
\(n\) remains of the same order, or
smaller, than the number of species \(p\).
Just like with all models fitted in PLNmodels, we adopt a variational strategy to approximate the log-likelihood function and optimize the consecutive variational surrogate of the log-likelihood with a gradient-ascent-based approach. In this case, it is not too difficult to show that PLN-mixture can be obtained by optimizing a collection of weighted standard PLN models.
In the package, the PLN-mixture model is adjusted with the function
PLNmixture
, which we review in this section. This function
adjusts the model for a series of value of \(k\) and provides a collection of objects
PLNmixturefit
stored in an object with class
PLNmixturefamily
.
The class PLNmixturefit
contains a collection of
components constituting the mixture, each of whom inherits from the
class PLNfit
, so we strongly recommend the reader to be
comfortable with PLN
and PLNfit
before using
PLNmixture
(see the PLN
vignette).
We fit a collection of \(K=5\) models with one iteration of forward smoothing of the log-likelihood as follows:
mixture_models <- PLNmixture(
Abundance ~ 1 + offset(log(Offset)),
data = trichoptera,
clusters = 1:4
)
##
## Initialization...
##
## Adjusting 4 PLN mixture models.
## number of cluster = 1
number of cluster = 2
number of cluster = 3
number of cluster = 4
##
## Smoothing PLN mixture models.
## Going backward +++
Going forward +++
## Post-treatments
## DONE!
Note the use of the formula
object to specify the model,
similar to the one used in the function PLN
.
PLNmixturefamily
The mixture_models
variable is an R6
object
with class PLNmixturefamily
, which comes with a couple of
methods. The most basic is the show/print
method, which
outputs a brief summary of the estimation process:
## --------------------------------------------------------
## COLLECTION OF 4 POISSON LOGNORMAL MODELS
## --------------------------------------------------------
## Task: Mixture Model
## ========================================================
## - Number of clusters considered: from 1 to 4
## - Best model (regarding BIC): cluster = 2
## - Best model (regarding ICL): cluster = 4
One can also easily access the successive values of the criteria in the collection
param | nb_param | loglik | BIC | ICL |
---|---|---|---|---|
1 | 18 | -1158.502 | -1193.529 | -2152.930 |
2 | 37 | -1104.249 | -1176.248 | -1982.117 |
3 | 56 | -1070.270 | -1179.241 | -1887.654 |
4 | 75 | -1041.142 | -1187.085 | -1790.688 |
A quick diagnostic of the optimization process is available via the
convergence
field:
param | nb_param | objective | convergence | outer_iterations | |
---|---|---|---|---|---|
out | 1 | 18 | 1158.502426 | 0.000000 | 2.000000 |
elt | 2 | 37 | 1104.249040 | 0.000019 | 2.000000 |
elt.1 | 3 | 56 | 1070.269876 | 0.000001 | 2.000000 |
elt.2 | 4 | 75 | 1041.142223 | 0.000000 | 2.000000 |
A visual representation of the optimization can be obtained be representing the objective function
Comprehensive information about PLNmixturefamily
is
available via ?PLNmixturefamily
.
The plot
method of PLNmixturefamily
displays evolution of the criteria mentioned above, and is a good
starting point for model selection:
Note that we use the original definition of the BIC/ICL criterion
(\(\texttt{loglik} -
\frac{1}{2}\texttt{pen}\)), which is on the same scale as the
log-likelihood. A popular
alternative consists in using \(-2\texttt{loglik} + \texttt{pen}\) instead.
You can do so by specifying reverse = TRUE
:
From those plots, we can see that the best model in terms of BIC is
obtained for a number of clusters of 2. We may extract the corresponding
model with the method getBestModel()
. A model with a
specific number of clusters can also be extracted with the
getModel()
method:
PLNmixturefit
Object myMix_BIC
is an R6Class
object with
class PLNmixturefit
which in turns has a couple of methods.
A good place to start is the show/print
method:
## Poisson Lognormal mixture model with 2 components and spherical covariances.
## * Useful fields
## $posteriorProb, $memberships, $mixtureParam, $group_means
## $model_par, $latent, $latent_pos, $optim_par
## $loglik, $BIC, $ICL, $loglik_vec, $nb_param, $criteria
## $component[[i]] (a PLNfit with associated methods and fields)
## * Useful S3 methods
## print(), coef(), sigma(), fitted(), predict()
The user can easily access several fields of the
PLNmixturefit
object using active binding or
S3
methods:
## [1] 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 1 2 2 2 2 2 2 2 1 1 1 1 1 1 1
## [39] 1 1 1 1 1 1 1 1 1 1 2
## [1] 0.8163265 0.1836735
1 | 0 |
1 | 0 |
1 | 0 |
1 | 0 |
1 | 0 |
1 | 0 |
## [[1]]
## Che Hyc Hym Hys Psy Aga Glo Ath
## 0.7184859 0.7184859 0.7184859 0.7184859 0.7184859 0.7184859 0.7184859 0.7184859
## Cea Ced Set All Han Hfo Hsp Hve
## 0.7184859 0.7184859 0.7184859 0.7184859 0.7184859 0.7184859 0.7184859 0.7184859
## Sta
## 0.7184859
##
## [[2]]
## Che Hyc Hym Hys Psy Aga Glo Ath
## 0.9626126 0.9626126 0.9626126 0.9626126 0.9626126 0.9626126 0.9626126 0.9626126
## Cea Ced Set All Han Hfo Hsp Hve
## 0.9626126 0.9626126 0.9626126 0.9626126 0.9626126 0.9626126 0.9626126 0.9626126
## Sta
## 0.9626126
coef(myMix_BIC, 'main') # equivalent to myMix_BIC$model_par$Theta
coef(myMix_BIC, 'mixture') # equivalent to myMix_BIC$model_par$Pi, myMix_BIC$mixtureParam
coef(myMix_BIC, 'means') # equivalent to myMix_BIC$model_par$Mu, myMix_BIC$group_means
coef(myMix_BIC, 'covariance') # equivalent to myMix_BIC$model_par$Sigma, sigma(myMix_BIC)
Intercept | Intercept.1 | |
---|---|---|
Che | -7.16 | -21.31 |
Hyc | -8.33 | -7.71 |
Hym | -2.68 | -4.87 |
Hys | -6.59 | -8.40 |
Psy | -0.52 | -0.81 |
Aga | -3.72 | -6.97 |
In turn, each component of a PLNmixturefit
is a
PLNfit
object (see the corresponding vignette)
## A multivariate Poisson Lognormal fit with spherical covariance model.
## ==================================================================
## nb_param loglik BIC ICL
## 18 -828.201 -863.227 -1388.21
## ==================================================================
## * Useful fields
## $model_par, $latent, $latent_pos, $var_par, $optim_par
## $loglik, $BIC, $ICL, $loglik_vec, $nb_param, $criteria
## * Useful S3 methods
## print(), coef(), sigma(), vcov(), fitted()
## predict(), predict_cond(), standard_error()
The PLNmixturefit
class also benefits from two important
methods: plot
and predict
.
plot
methodWe can visualize the clustered latent position by performing a PCA on the latent layer:
We can also plot the data matrix with samples reordered by clusters to check whether it exhibits strong pattern or not. The limits between clusters are highlighted by grey lines.
predict
methodFor PLNmixture, the goal of predict
is to predict the
membership based on observed newly species counts.
By default, the predict
use the argument
type = "posterior"
to output the matrix of posterior
probabilities \(\hat{\pi}_k\)
predicted.class <- predict(myMix_BIC, newdata = trichoptera)
## equivalent to
## predicted.class <- predict(myMIX_BIC, newdata = trichoptera, type = "posterior")
predicted.class %>% head() %>% knitr::kable(digits = 2)
1 | 0 |
1 | 0 |
1 | 0 |
1 | 0 |
1 | 0 |
1 | 0 |
Setting type = "response"
, we can predict the most
likely cluster \(\hat{k} = \arg\max_{k =
1\dots K} \{ \hat{\pi_k}\}\) instead:
predicted.class <- predict(myMix_BIC, newdata = trichoptera,
prior = myMix_BIC$posteriorProb, type = "response")
predicted.class
## 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26
## 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 1 2 2
## 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49
## 2 2 2 2 2 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2
## Levels: 1 2
We can assess that the predictions are quite similar to the real group (this is not a proper validation of the method as we used data set for both model fitting and prediction and are thus at risk of overfitting).
Finally, we can get the coordinates of the new data on the same graph
at the original ones with type = "position"
. This is done
by averaging the latent positions \(\hat{\mathbf{Z}}_i + \boldsymbol{\mu}_k\)
(found when the sample is assumed to come from group \(k\)) and weighting them with the \(\hat{\pi}_k\). Some samples, have
compositions that put them very far from their group mean.
predicted.position <- predict(myMix_BIC, newdata = trichoptera,
prior = myMix_BIC$posteriorProb, type = "position")
prcomp(predicted.position) %>%
factoextra::fviz_pca_ind(col.ind = predicted.class)
When you are done, do not forget to get back to the standard sequential plan with future.