appendix : adding proof of ident results

This commit is contained in:
Louis Lacoste 2024-08-18 14:47:37 +02:00
parent a4cc5deec8
commit 6d6652375d
4 changed files with 128 additions and 75 deletions

View file

@ -1,7 +1,57 @@
\clearpage
\pagenumbering{arabic}% resets `page` counter to 1
\renewcommand*{\thepage}{A-\arabic{page}}
\renewcommand*{\thepage}{S-\arabic{page}}
\appendix
\chapter{Supplementary for~\nameref{chap:struct-detection}}
\section{Proof of the idenfiability result}
\label{sec:proof-identifiability}
We recall the following
\def\thetheorem{\ref{thm:identifiability-iid}}
\begin{theorem}[Identifiability of $iid$-colBiSBM]
The parameters $(\bm{\pi}, \bm{\rho}, \bm{\alpha})$ are
identifiable up to a label switching of the blocks if those
conditions are achieved:
\begin{itemize}
\item[(1.1)] $\exists m^*\in\{1,\dots,M\} : n^1_{m^*} \geq 2 Q_2 - 1~\text{and}~n^2_{m^*} \geq 2 Q_1 - 1$.
\item[(1.2)] $\forall 1\leq q \leq Q_1, \pi_q > 0$
and the coordinates of vector $\bm{\rho}
{X^{m^*}}^T$ are distinct (where ${X^{m^*}}^T$ is the transpose of $X^{m^*}$).
\item[(1.3)] $\forall 1\leq r \leq Q_2, \rho_r > 0$
and the coordinates of vector $\bm{\pi}
X^{m^*}$ are distinct.
\end{itemize}
\end{theorem}
\begin{proof}
Following the tracks of~\cite{chabert-liddellLearningCommonStructures2024a}
we derive the result in Properties~\ref{thm:identifiability-iid}.
\cite{keribinEstimationSelectionLatent2015} building
on~\cite{celisseConsistencyMaximumlikelihoodVariational2012}, proved that the
parameters $(\bm{\pi}, \bm{\rho}, \bm{\alpha})$ of the
$\mathcal{F}\text{-BiSBM}_{n_1^m,n_2^m}(Q_1^m, Q_2^m, \bm{\pi^m}, \bm{\rho^m}, \bm{\alpha^m})$
are identifiable from the observation of network $X^m$ when $\mathcal{F}$
is the Bernoulli distribution and the following conditions are met:
\begin{enumerate}
\item $ n_1^m \geq 2 Q_2^m - 1~\text{and}~n_2^m \geq 2 Q_1^m - 1$.
\item $\forall 1\leq q \leq Q_1^m, \pi_q^m > 0$
and the coordinates of vector $\bm{\rho^m}
{X^{m^*}}^T$ are distinct (where ${X^{m^*}}^T$ is the transpose of $X^{m^*}$).
\item $\forall 1\leq r \leq Q_2^m, \rho_r^m > 0$
and the coordinates of vector $\bm{\pi^m}
X^{m^*}$ are distinct.
\end{enumerate}
Under the \emph{iid}-colBiSBM model, for all $m=1\dots M$,
$X^m \sim \mathcal{F}\text{-BiSBM}_{n_1^m,n_2^m}(Q_1, Q_2,
\bm{\pi}, \bm{\rho}, \bm{\alpha})$. This means that
following~\cite{keribinEstimationSelectionLatent2015}, the
identifiability of $\bm{\alpha}$, $\bm{\pi}$ and $\bm{\rho}$ is obtained
from the distribution of $X^{m^*}$ under assumptions (1.1), (1.2) and
(1.3).
\end{proof}
\chapter{Supplementary for~\nameref{chap:simulation-studies}}
Below are the supplementary material for the~\nameref{chap:simulation-studies}.
@ -31,14 +81,18 @@ Please note that blank space indicates that among all conditions
the corresponding model was not selected at all.
\begin{landscape}
\pagestyle{empty}
\input{../tables/simulations/model_selection/model-selection.tex}
\end{landscape}
\pagestyle{fancy}
\chapter{Supplementary for~\nameref{chap:applications-ecological-networks}}
\section{Additional information on~\nameref{sec:baldock-clustering}}
\fancypagestyle{fancy}
\renewcommand*{\thepage}{S-\arabic{page}}
Due to report size limitations we included these plots here as they are not crucial to understand what is going on in
the section~\ref{sec:baldock-clustering}.
Yet they are useful to confirm the explanation given.

View file

@ -1,5 +1,6 @@
\addtocounter{customchapter}{1}
\chapter[Structure detection in bipartite collection]{Structure detection in a collection of bipartite networks}
\label{chap:struct-detection}
\section{Definition of a collection}
\label{sec:definition-of-a-collection}
@ -351,49 +352,49 @@ We provide below the expression for the penalties for the 4 models that we
propose.
\begin{description}
\item[\textit{iid}-colBiSBM] For the $\bm\pi$ and $\bm\rho$:
\begin{align*}
\text{pen}_{\pi}(Q_1) = (Q_1 - 1)\log(\sum_{m=1}^{M}n_{1}^{m}) & , &
\text{pen}_{\rho}(Q_2) = (Q_2 - 1)\log(\sum_{m=1}^{M}n_{2}^{m})
\end{align*}
For the $\bm\alpha$:
\[\text{pen}_{\alpha}(Q_1, Q_2) = Q_1 \times Q_2 \log(N_M)\]
with
\[ N_M = \sum_{m = 1}^{M} n_{1}^{m} \times n_{2}^{m} \]
And thus the $\text{BIC-L}$ formula is the following:
\[ \text{BIC-L}(\bm{X},Q_1, Q_2) = \max_{\theta}
\mathcal{J} (\mathcal{\hat{R}}, \bm{\theta})
- \frac{1}{2} [\text{pen}_{\pi}(Q_1) + \text{pen}_{\rho}(Q_2) +
\text{pen}_{\alpha}(Q_1, Q_2)]\]
\begin{align*}
\text{pen}_{\pi}(Q_1) = (Q_1 - 1)\log(\sum_{m=1}^{M}n_{1}^{m}) & , &
\text{pen}_{\rho}(Q_2) = (Q_2 - 1)\log(\sum_{m=1}^{M}n_{2}^{m})
\end{align*}
For the $\bm\alpha$:
\[\text{pen}_{\alpha}(Q_1, Q_2) = Q_1 \times Q_2 \log(N_M)\]
with
\[ N_M = \sum_{m = 1}^{M} n_{1}^{m} \times n_{2}^{m} \]
And thus the $\text{BIC-L}$ formula is the following:
\[ \text{BIC-L}(\bm{X},Q_1, Q_2) = \max_{\theta}
\mathcal{J} (\mathcal{\hat{R}}, \bm{\theta})
- \frac{1}{2} [\text{pen}_{\pi}(Q_1) + \text{pen}_{\rho}(Q_2) +
\text{pen}_{\alpha}(Q_1, Q_2)]\]
\item[$\bm{\pi\rho}$-colBiSBM] The support penalties are
\begin{align*}
\text{pen}_{S_1}(Q_1) = -2 \log p_{Q_1} (S_1) & , &
\text{pen}_{S_2}(Q_2) = -2 \log p_{Q_2} (S_2)
\end{align*}
with \begin{align*}
\textstyle \log p_{Q_1}(S_1) = - M \log(Q_1) - \sum_{m=1}^{M} \log {Q_1
\choose Q_1^{(m)}}, \\
\textstyle \log p_{Q_2}(S_2) = - M \log(Q_2) - \sum_{m=1}^{M} \log {Q_2
\choose Q_2^{(m)}}.
\end{align*}
And penalties for the $\bm\rho$ and $\bm\pi$ are
\[ \text{pen}_{\pi}(Q_1, S_1) = \sum_{m=1}^{M} (Q_{1}^{(m)} - 1)
\log n_{1}^{m},
~\text{pen}_{\rho}(Q_2, S_2) = \sum_{m=1}^{M} (Q_{2}^{(m)} - 1)
\log n_{2}^{m}. \]
Penalties for the $\bm\alpha$
\[ \text{pen}_{\alpha}(Q_1, Q_2, S_1, S_2) = (\sum_{q=1}^{Q_1}
\sum_{r=1}^{Q_2} \mathbbb{1}_{(S_1)'S_2 > 0}) \log (N_M). \]
And the corresponding BIC-L formula,
\[
\begin{aligned}
\text{BIC-L}(\bm{X},Q_1, Q_2) =
\max_{S_1,S_2} [
& \max_{\theta_{S_1,S_2} \in \Theta_{S_1,S_2}} \mathcal{J}(\mathcal{\hat{R}},\theta_{S_1,S_2}) \\
- \frac{1}{2} & (\text{pen}_{\pi}(Q_1, S_1) + \text{pen}_{\rho}(Q_2, S_2) \\
& + \text{pen}_{\alpha}(Q_1, Q_2, S_1, S_2) \\
& + \text{pen}_{S_1}(Q_1) + \text{pen}_{S_2}(Q_2))] \\
\end{aligned}
\]
\begin{align*}
\text{pen}_{S_1}(Q_1) = -2 \log p_{Q_1} (S_1) & , &
\text{pen}_{S_2}(Q_2) = -2 \log p_{Q_2} (S_2)
\end{align*}
with \begin{align*}
\textstyle \log p_{Q_1}(S_1) = - M \log(Q_1) - \sum_{m=1}^{M} \log {Q_1
\choose Q_1^{(m)}}, \\
\textstyle \log p_{Q_2}(S_2) = - M \log(Q_2) - \sum_{m=1}^{M} \log {Q_2
\choose Q_2^{(m)}}.
\end{align*}
And penalties for the $\bm\rho$ and $\bm\pi$ are
\[ \text{pen}_{\pi}(Q_1, S_1) = \sum_{m=1}^{M} (Q_{1}^{(m)} - 1)
\log n_{1}^{m},
~\text{pen}_{\rho}(Q_2, S_2) = \sum_{m=1}^{M} (Q_{2}^{(m)} - 1)
\log n_{2}^{m}. \]
Penalties for the $\bm\alpha$
\[ \text{pen}_{\alpha}(Q_1, Q_2, S_1, S_2) = (\sum_{q=1}^{Q_1}
\sum_{r=1}^{Q_2} \mathbbb{1}_{(S_1)'S_2 > 0}) \log (N_M). \]
And the corresponding BIC-L formula,
\[
\begin{aligned}
\text{BIC-L}(\bm{X},Q_1, Q_2) =
\max_{S_1,S_2} [
& \max_{\theta_{S_1,S_2} \in \Theta_{S_1,S_2}} \mathcal{J}(\mathcal{\hat{R}},\theta_{S_1,S_2}) \\
- \frac{1}{2} & (\text{pen}_{\pi}(Q_1, S_1) + \text{pen}_{\rho}(Q_2, S_2) \\
& + \text{pen}_{\alpha}(Q_1, Q_2, S_1, S_2) \\
& + \text{pen}_{S_1}(Q_1) + \text{pen}_{S_2}(Q_2))] \\
\end{aligned}
\]
\end{description}
\subsection{Initialization and pairing of the models}
@ -708,7 +709,7 @@ And the pairwise dissimilarity for networks $(m,m')\in\mathcal{M}^2$ is then:
\begin{figure}[t]
\centering
\begin{tikzpicture}
\begin{tikzpicture}[scale=0.7]
\tikzstyle{instruct}=[font=\small, text justified, rectangle,draw,fill=yellow!50]
\tikzstyle{first_col}=[rectangle, text justified, draw,fill=gray!50]
\tikzstyle{second_col}=[scale=0.55, circle, draw,fill=red!50]
@ -751,12 +752,10 @@ trivial partition in a unique group.
Then using the \emph{Kmeans} we split the collection in two sub-collections
with the dissimilarity matrix. The two sub-collections are fitted and we
compute the score of this new partition $\mathcal{G}^{*} = \{G_1, G_2\}$.
If $Sc(\mathcal{G}^{*}) > Sc(\mathcal{G})$ then we repeat the same procedure on
$G_1$ and $G_2$. Else we return $\mathcal{G}$.
We illustrate our capacity to perform a partition of a collection for all
colBiSBM models in %\ref{sec:network-clustering-of-simulated-networks}.
colBiSBM models in~\ref{sec:network-clustering-of-simulated-networks}.
\section{Model identifiability}
\label{sec:model-identifiability}
@ -764,7 +763,7 @@ colBiSBM models in %\ref{sec:network-clustering-of-simulated-networks}.
The goal here is to prove that if $\ell(\bm{X};\bm{\theta}) = \ell(\bm{X};\bm{\theta}')$ for any collection $\bm{X}$ then $\bm{\theta} = \bm{\theta}'$.
Following the proof proposed by~\cite{chabert-liddellLearningCommonStructures2024a}, that adapted it to the collection case and~\cite{keribinEstimationSelectionLatent2015} that extended the result of~\cite{celisseConsistencyMaximumlikelihoodVariational2012} to the LBM Bernoulli model,
we obtain the following proof of identifiability for the $iid$-colBiSBM:
we obtain the following result of identifiability\footnote{The proof is in appendix. \ref{sec:proof-identifiability}} for the $iid$-colBiSBM:
\begin{theorem}[Identifiability of $iid$-colBiSBM]
\label{thm:identifiability-iid}
The parameters $(\bm{\pi}, \bm{\rho}, \bm{\alpha})$ are
@ -773,11 +772,11 @@ we obtain the following proof of identifiability for the $iid$-colBiSBM:
\begin{itemize}
\item[(1.1)] $\exists m^*\in\{1,\dots,M\} : n^1_{m^*} \geq 2 Q_2 - 1~\text{and}~n^2_{m^*} \geq 2 Q_1 - 1$.
\item[(1.2)] $\forall 1\leq q \leq Q_1, \pi_q > 0$
and the coordinates of vector $\bm{\rho}
{X^{m^*}}^T$ are distinct (where ${X^{m^*}}^T$ is the transpose of $X^{m^*}$).
and the coordinates of vector $\bm{\rho}
{X^{m^*}}^T$ are distinct (where ${X^{m^*}}^T$ is the transpose of $X^{m^*}$).
\item[(1.3)] $\forall 1\leq r \leq Q_2, \rho_r > 0$
and the coordinates of vector $\bm{\pi}
X^{m^*}$ are distinct.
and the coordinates of vector $\bm{\pi}
X^{m^*}$ are distinct.
\end{itemize}
\end{theorem}

Binary file not shown.

View file

@ -26,7 +26,7 @@
hypertexnames=true
}
\newtheorem{theorem}{Theorem}
\newtheorem{theorem}{Properties}
\usepackage{tocbibind} % Pour avoir des index pour table des matières, biblio
\usepackage{geometry}
\geometry{bmargin=25mm}
@ -223,31 +223,31 @@ automata,positioning}
% Pour activer les onglets
\ActivateBG
\begin{selectlanguage}{french}
% \maketitle
\pagenumbering{roman}
\tableofcontents
\include{remerciements}
% \include{chapter1-presentation_UMR}
% \maketitle
\pagenumbering{roman}
\tableofcontents
\include{remerciements}
% \include{chapter1-presentation_UMR}
\end{selectlanguage}
\begin{selectlanguage}{english}
\pagenumbering{arabic}
\include{chapter2-context}
\include{chapter3-structure-detection}
\include{chapter4-simulation-studies}
\pagenumbering{arabic}
\include{chapter2-context}
\include{chapter3-structure-detection}
\include{chapter4-simulation-studies}
% \chapter{Applications}
% \include{Rcodes/real_data/application_dore}
% \include{Rcodes/real_data/CoOPLBM_completion_analyze}
\include{chapter5-applications}
\include{conclusions}
% \chapter{Applications}
% \include{Rcodes/real_data/application_dore}
% \include{Rcodes/real_data/CoOPLBM_completion_analyze}
\include{chapter5-applications}
\include{conclusions}
\addtocounter{maincontentend}{1}
\addtocounter{customchapter}{1}
\printbibliography
\input{appendices.tex}
% \listoffigures
% \listoftables
\addtocounter{maincontentend}{1}
\addtocounter{customchapter}{1}
\printbibliography
\input{appendices.tex}
% \listoffigures
% \listoftables
\end{selectlanguage}
\end{document}