228 lines
No EOL
11 KiB
TeX
228 lines
No EOL
11 KiB
TeX
\section{VEM}
|
|
|
|
\begin{frame}{Developed formula of variational EM}
|
|
\begin{multline*}
|
|
\ell (\bm{Y};\theta) \geq \color{red}\sum_{m=1}^{M} \bigg( \color{black} \sum_{i = 1}^{n_1^m}\sum_{j=1}^{n_2^m}\sum_{q \in \mathcal{Q}_{1,m}} \sum_{r \in \mathcal{Q}_{2,m}} \tau^{1,m}_{i,q} \tau^{2,m}_{j,r} \log f(Y^{m}_{ij}; \alpha_{qr}) \\
|
|
+ \sum_{i=1}^{n_1^m} \sum_{q \in \mathcal{Q}_{1,m}} \tau^{1,m}_{i,q} \log \pi_{\color{black}q}^{\color{gray}m} + \sum_{j=1}^{n_2^m} \sum_{r \in \mathcal{Q}_{2,m}} \tau^{2,m}_{j,r} \log \rho_{\color{black}r}^{\color{gray}m} \\
|
|
- \sum_{i=1}^{n_1} \tau^{1,m}_{i,q} \log \tau^{1,m}_{i,q} - \sum_{j=1}^{n_2} \tau^{2,m}_{j,r} \log \tau^{2,m}_{j,r} \color{red}\bigg) \color{black} \eqcolon
|
|
\mathcal{J}(\tau;\theta),
|
|
\end{multline*}
|
|
|
|
\begin{block}{Variational approximation}
|
|
$\tau_{iq}^{1,m} = \mathcal{R}^1_{Y^m,\tau}(Z_{iq}^m = 1)$
|
|
and $\tau_{jr}^{2,m} = \mathcal{R}^2_{Y^m,\tau}(W_{jr}^m = 1)$
|
|
\end{block}
|
|
\end{frame}
|
|
|
|
\begin{frame}{\emph{Variational Expectation} Step}
|
|
\[
|
|
\widehat{\tau}^{(t+1)} = \arg \max_{\tau}
|
|
\mathcal{J}(\mathcal{\tau},\bm{\widehat{\theta}}^{(t)})
|
|
\Leftrightarrow \arg\min_{\tau\in\mathcal{T}} \mathbf{KL}[\mathcal{R}_{\mathbf{Y},\tau}, \mathbb{P}(.|\mathbf{Y})]
|
|
\]
|
|
|
|
\begin{equation*}
|
|
\begin{cases}
|
|
\widehat{\tau}_{iq}^{1,m} \propto \widehat{\pi}_{q}^{m(t)} \prod_{j=1}^{n_2^m}\prod_{r\in\mathcal{Q}_2^m} f(Y_{ij}^m;\widehat{\alpha}_{qr}^{(t)})^{\widehat{\tau}_{jr}^{2,m(t+1)}} & \forall i = 1, \dots , n_1^m, q \in \mathcal{Q}_1^m \\
|
|
\widehat{\tau}_{jr}^{2,m} \propto \widehat{\rho}_{r}^{m(t)} \prod_{i=1}^{n_1^m}\prod_{q\in\mathcal{Q}_1^m} f(Y_{ij}^m;\widehat{\alpha}_{qr}^{(t)})^{\widehat{\tau}_{iq}^{1,m(t+1)}} & \forall j = 1, \dots , n_2^m, r \in \mathcal{Q}_2^m
|
|
\end{cases}
|
|
\end{equation*}
|
|
\footnotetext[2]{Initialization of $\widehat{\tau}$ with a
|
|
\emph{spectral clustering} on the networks.}
|
|
\end{frame}
|
|
|
|
\begin{frame}{\emph{Maximization} Step}
|
|
\[
|
|
\widehat{\theta}^{(t+1)} = \arg \max_{\theta} \mathcal{J}(\mathcal{\bm{\widehat{\tau}}}^{(t+1)},\theta)
|
|
\]
|
|
|
|
\begin{block}{Connectivity parameters}
|
|
\begin{align*}
|
|
\widehat{\alpha}_{qr} = \frac{\sum_{m=1}^{M} \sum_{i=1}^{n_1^m} \sum_{j=1}^{n_2^m} \tau_{iq}^{1,m} \tau_{jr}^{2,m} \alert<2>{Y_{ij}^m}}{\sum_{m=1}^{M} \sum_{i=1}^{n_1^m} \sum_{j=1}^{n_2^m} \tau_{iq}^{1,m} \tau_{jr}^{2,m}}
|
|
\end{align*}
|
|
\end{block}
|
|
\only<1>{
|
|
\begin{block}{Proportions for \emph{iid}}
|
|
\begin{align*}
|
|
\widehat{\pi}_q = \frac{\sum_{m=1}^{M} \sum_{i=1}^{n_1^m} \tau_{iq}^{1,m}}{\sum_{m=1}^{M} n_1^m} & &
|
|
\widehat{\rho}_r = \frac{\sum_{m=1}^{M} \sum_{j=1}^{n_2^m} \tau_{jr}^{2,m}}{\sum_{m=1}^{M} n_2^m}
|
|
\end{align*}
|
|
\end{block}
|
|
}
|
|
\only<2>{
|
|
\begin{block}{Proportions for $\pi\rho$}
|
|
\begin{align*}
|
|
\widehat{\pi}^{\color{red}m}_q = \frac{\sum_{i=1}^{n_1^m} \tau_{iq}^{1,m}}{n_1^m} & &
|
|
\widehat{\rho}^{\color{red}m}_r = \frac{\sum_{j=1}^{n_2^m} \tau_{jr}^{2,m}}{n_2^m}
|
|
\end{align*}
|
|
\end{block}
|
|
}
|
|
|
|
\end{frame}
|
|
|
|
\begin{frame}
|
|
\frametitle{Why does VE minimizes KL ?}
|
|
\begin{align*}
|
|
\ell_c(\bY,\bZ,\bW;\theta) & = \log \Prob(\bZ, \bW|\bY;\theta) + \ell(\bY;\theta) \\
|
|
\Leftrightarrow \ell(\bY;\theta) & = \ell_c(\bY,\bZ,\bW;\theta) - \log \Prob(\bZ, \bW|\bY;\theta) \\
|
|
\Leftrightarrow \Esp_{\Ryt}[\ell(\bY;\theta)] & = \Esp_{\Ryt}[\ell_c(\bY,\bZ,\bW;\theta)] - \Esp_{\Ryt}[\log \Prob(\bZ,\bW|\bY;\theta)] \\
|
|
\Leftrightarrow \ell(\bY;\theta) & = \Esp_{\Ryt}[\ell_c(\bY,\bZ,\bW;\theta)] - \Esp_{\Ryt}[\log \Prob(\bZ,\bW|\bY;\theta)] \\
|
|
\end{align*}
|
|
\begin{align*}
|
|
\text{But }\KL{\Ryt}{\log \Prob(\bZ,\bW|\bY;\theta)} & = - \Esp_{\Ryt} [\log \frac{\Prob(\bZ,\bW|\bY;\theta)}{\Ryt}] \\
|
|
= - \Esp_{\Ryt} [\log \Prob(\bZ,\bW|\bY;\theta)] + & \underbrace{\Esp_{\Ryt[\log \Ryt]}}_{-\Hshannon(\Ryt)} \\
|
|
\Leftrightarrow \KL{\Ryt}{\log \Prob(\bZ,\bW|\bY;\theta)} + \Hshannon(\Ryt) & = - \Esp_{\Ryt} [\log \Prob(\bZ,\bW|\bY;\theta)]
|
|
\end{align*}
|
|
Thus $\ell(\bY;\theta) - \KL{\Ryt}{\log \Prob(\bZ,\bW|\bY;\theta)} = \mathcal{J}(\tau;\theta) \qed$
|
|
\end{frame}
|
|
|
|
\begin{frame}
|
|
\frametitle{On the BIC-L}
|
|
Raconter l'histoire dans l'ordre suivant :
|
|
\begin{itemize}
|
|
\item ICL = Méthode BIC (approx Laplace) sur la log complète, fait apparaître la
|
|
pénalité de complexité et pénalise l'entropie
|
|
\item ICLv = ICL mais avec les paramètres variationnels et l'entropie variationnelle
|
|
\item BIC-L = ICLv mais sans la pénalité sur l'entropie et la rajoutant à la fin
|
|
\end{itemize}
|
|
\begin{align*}
|
|
\text{BIC}(\hat{\theta}) & = \log p(\mathbf{Y};\hat{\theta}) - \frac{1}{2} \text{pen}(\dots) \\
|
|
& = \Esp_{\mathbf{Z}, \mathbf{W}|\mathbf{Y}} [\underbrace{\log p(\mathbf{Y},\mathbf{Z},\mathbf{W};\hat{\theta})}_{\ell_c(\mathbf{Y},\mathbf{Z},\mathbf{W};\hat{\theta})}] + \mathcal{H}(p(\mathbf{Z},\mathbf{W}|\mathbf{Y})) - \frac{1}{2} \text{pen}(\dots) \\
|
|
\text{ICL}(\hat{\theta}) & = \Esp_{\mathbf{Z}, \mathbf{W}|\mathbf{Y}} [\ell_c(\mathbf{Y},\mathbf{Z},\mathbf{W};\hat{\theta})] - \frac{1}{2} \text{pen}(\dots) \\
|
|
\text{BIC-L}(\hat{\theta}, \hat{\tau}) & = \Esp_{\mathcal{R}_{\mathbf{Y}, \hat{\tau}}}[\ell_c(\mathbf{Y},\mathbf{Z},\mathbf{W};\hat{\theta}^{\text{var}})] + \mathcal{H}(\mathcal{R}_{\mathbf{Y}, \hat{\tau}}) - \frac{1}{2} \text{pen}(\dots) \\
|
|
\end{align*}
|
|
\end{frame}
|
|
|
|
\section{Model selection}
|
|
|
|
\begin{frame}
|
|
\frametitle{Choice of $(Q_1,Q_2)$ - Greedy approach}
|
|
\begin{columns}
|
|
\begin{column}{0.5\linewidth}
|
|
\begin{tikzpicture}
|
|
\input{tikz/greedy-exploration.tex}
|
|
\end{tikzpicture}
|
|
\end{column}
|
|
\begin{column}{0.35\linewidth}
|
|
\begin{itemize}
|
|
\item Initial model~: \\ \vspace{0.125\baselineskip}
|
|
\begin{tikzpicture}
|
|
\draw[fill=gray, draw=gray] circle [radius=0.225cm];
|
|
\end{tikzpicture}
|
|
\onslide<2->{
|
|
\item Model after \emph{split}~:
|
|
\begin{tikzpicture}
|
|
\draw[fill=blueind, draw=blueind] circle [radius=0.225cm];
|
|
\end{tikzpicture}
|
|
\item Model maximizing the criterion~:\\ \vspace{0.125\baselineskip}
|
|
\begin{tikzpicture}
|
|
\draw[fill=white, draw=green, very thick] circle [radius=0.225cm];
|
|
\end{tikzpicture}
|
|
}
|
|
\onslide<3->{
|
|
\item Model after \emph{merge}~:
|
|
\begin{tikzpicture}
|
|
\draw[fill=red, draw=red] circle [radius=0.225cm];
|
|
\end{tikzpicture}
|
|
}
|
|
\end{itemize}
|
|
\end{column}
|
|
\end{columns}
|
|
\end{frame}
|
|
|
|
\begin{frame}
|
|
\frametitle{Choice of $(Q_1,Q_2)$ - Sliding window}
|
|
\begin{columns}
|
|
\begin{column}{0.6\textwidth}
|
|
\begin{figure}
|
|
\input{tikz/moving-window}
|
|
\caption{Sliding window}
|
|
\end{figure}
|
|
\end{column}
|
|
\begin{column}{0.4\textwidth}
|
|
\only<3>{\begin{block}{}
|
|
Initialization of the model if necessary
|
|
\end{block}}
|
|
\only<9>{\begin{block}{}
|
|
Localization of the new mode
|
|
\end{block}}
|
|
\only<10>{\begin{block}{}
|
|
Move to the new mode then iterate
|
|
\end{block}}
|
|
\end{column}
|
|
\end{columns}
|
|
\end{frame}
|
|
|
|
\section{Clustering}
|
|
|
|
\begin{frame}{Clustering algorithm}
|
|
\centering
|
|
\vspace{0.25\baselineskip}
|
|
\begin{tikzpicture}[scale=0.85]
|
|
\input{tikz/clustering.tex}
|
|
\end{tikzpicture}
|
|
\[
|
|
D_{\mathcal{M}}(m,m') = \sum_{q = 1}^{Q_1} \sum_{r = 1}^{Q_2} \max(\widetilde{\pi}_{q}^{m}, \widetilde{\pi}_{q}^{m'}) \left( \widetilde{\alpha}_{qr}^{m} - \widetilde{\alpha}_{qr}^{m'}\right)^{2} \max(\widetilde{\rho}_{r}^{m}, \widetilde{\rho}_{r}^{m'})
|
|
\]
|
|
\end{frame}
|
|
|
|
\section{Results~\cite{baldockSystemsApproachReveals2019,baldockDailyTemporalStructure2011}}
|
|
\begin{frame}[allowframebreaks]
|
|
\begin{figure}[ht]
|
|
\centering
|
|
\begin{subfigure}[t]{0.5\textwidth}
|
|
\centering
|
|
\includegraphics[width=0.45\textwidth]{tikz/applications/baldock/mat-Baldock2019_Bristol.pdf}
|
|
\caption{Donnée}
|
|
\end{subfigure}\hfil
|
|
\begin{subfigure}[t]{0.5\textwidth}
|
|
\centering
|
|
\includegraphics[width=0.45\textwidth]{tikz/applications/baldock/colbisbm-mat-Baldock2019_Bristol.pdf}
|
|
\caption{Reordered}
|
|
\end{subfigure}
|
|
\caption{Bristol}
|
|
\end{figure}
|
|
|
|
\begin{figure}[ht]
|
|
\centering
|
|
\begin{subfigure}[t]{0.5\textwidth}
|
|
\centering
|
|
\includegraphics[width=0.45\textwidth]{tikz/applications/baldock/mat-Baldock2019_Edinburgh.pdf}
|
|
\caption{Donnée}
|
|
\end{subfigure}\hfil
|
|
\begin{subfigure}[t]{0.5\textwidth}
|
|
\centering
|
|
\includegraphics[width=0.5\textwidth]{tikz/applications/baldock/colbisbm-mat-Baldock2019_Edinburgh.pdf}
|
|
\caption{Reordered}
|
|
\end{subfigure}
|
|
\caption{Edinburgh}
|
|
\end{figure}
|
|
|
|
\begin{figure}
|
|
\begin{subfigure}[ht]{0.5\textwidth}
|
|
\centering
|
|
\includegraphics[width=0.5\textwidth]{tikz/applications/baldock/mat-Baldock2019_Leeds.pdf}
|
|
\caption{Donnée}
|
|
\end{subfigure}\hfil
|
|
\begin{subfigure}[ht]{0.5\textwidth}
|
|
\centering
|
|
\includegraphics[width=0.5\textwidth]{tikz/applications/baldock/colbisbm-mat-Baldock2019_Leeds.pdf}
|
|
\caption{Réordonnée}
|
|
\end{subfigure}
|
|
\caption{Leeds}
|
|
\end{figure}
|
|
|
|
\begin{figure}
|
|
\begin{subfigure}[ht]{0.5\textwidth}
|
|
\centering
|
|
\includegraphics[width=0.5\textwidth]{tikz/applications/baldock/mat-Baldock2019_Reading.pdf}
|
|
\caption{Donnée}
|
|
\end{subfigure}\hfil
|
|
\begin{subfigure}[ht]{0.5\textwidth}
|
|
\centering
|
|
\includegraphics[width=0.5\textwidth]{tikz/applications/baldock/colbisbm-mat-Baldock2019_Reading.pdf}
|
|
\caption{Réordonnée}
|
|
\end{subfigure}
|
|
\caption{Reading}
|
|
\end{figure}
|
|
\end{frame} |