path: root/finale/mid_report.tex

\documentclass[10pt]{article}
\usepackage[utf8x]{inputenc}
\usepackage{amsmath, amssymb, amsthm, microtype}
\usepackage{algpseudocode, algorithm, algorithmicx}
\usepackage[pagebackref=false,breaklinks=true,
            colorlinks=true,citecolor=blue]{hyperref}
\usepackage[capitalize, noabbrev]{cleveref}
\usepackage{graphicx, subfig}
\usepackage{bbm}
\usepackage{fullpage}
\input{def} 
\DeclareMathOperator*{\argmax}{arg\,max}
\DeclareMathOperator*{\argmin}{arg\,min}
\DeclareMathOperator{\E}{\mathbb{E}}
\let\P\relax
\DeclareMathOperator{\P}{\mathbb{P}}
\newcommand{\ex}[1]{\E\left[#1\right]}
\newcommand{\prob}[1]{\P\left[#1\right]}
\newcommand{\inprod}[1]{\left\langle #1 \right\rangle}
\newcommand{\R}{\mathbf{R}}
\newcommand{\N}{\mathbf{N}}
\newcommand{\C}{\mathcal{C}}
\newcommand{\eps}{\varepsilon}
\newcommand{\bt}{\boldsymbol{\theta}}
\newcommand{\bx}{\mathbf{x}}
\newcommand{\cl}[1]{\text{\textbf{#1}}}
\newcommand{\eqdef}{\mathbin{\stackrel{\rm def}{=}}}

\newtheorem{theorem}{Theorem}
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{corollary}[theorem]{Corollary}
\theoremstyle{definition}
\newtheorem{definition}[theorem]{Definition}
\theoremstyle{remark}
\newtheorem*{example}{Example}
\newtheorem*{remark}{Remark}

\title{Network Inference from Cascades}
\author{Thibaut Horel \and Jean Pouget-Abadie}

\begin{document}

\maketitle

\begin{abstract}
    The Network Inference Problem (NIP) is the machine learning challenge of
    recovering the edges and edge weights of an unknown weighted graph from the
    observations of a random contagion process propagating over this graph.
    While estimators for the edge weights with provable convergence guarantees
    have been obtained under various formulations of the NIP, a rigorous
    statistical treatment of the problem is still lacking. In this work, we
    build upon the unified NIP formulation of \cite{pouget} to explore the
    connections between the topological properties of the graph to be learnt
    and the resulting quality of the estimators. Specifically, we analyze which
    properties of the graph render NIP unfeasible or hard, and which properties
    can be exploited to improve the quality of the estimators.
\end{abstract}

\section{Introduction}

A random contagion process propagating over a graph provides valuable
information about the presence and weights of its edges. This fact is exploited
in many real-life applications, including learning how people influence each
other by observing how they communicate or share information in social
networks, learning protein activation graphs in biology or learning correlations
between financial assets.

Formally, this problem can be formulated as a machine learning task where one
specifies a probabilistic contagion model whose parameters are the edge weights
of the unknown graph. Then, given observations drawn from this random process,
one seeks to construct an estimator of the edge weights. This problem is known
in the literature as the Network Inference Problem

A growing body of works has focused on obtaining convergence guarantees for
estimators of the edge weights under various formulations of the Network
Inference Problem. In \cite{pouget}, we noted that many of these formulations
could be interpreted as specific instances of a common \emph{Generalized Linear
Cascade Model} (GLMC) and showed how to apply results from the sparse recovery
literature to obtain close-to-optimal convergence guarantees.

However, in \cite{pouget} as in previous works, the convergence guarantees rely
on many assumptions about the non-degeneracy of the probabilistic model or the
observations. These assumptions are hard to interpret and weaken possible
applications of the obtained results to real-life applications. Furthermore,
previous works focused on obtaining generic convergence guarantees and it is
not obvious that the developed frameworks can be adapted in presence of domain
specific knowledge about the graph to be learned. The goal of the current work
is to address these shortcomings by asking the following question: \emph{how do
    properties of the graph to be learned impact the task of solving the
    Network Inference Problem for this graph?}. Specifically, we wish to
    understand:
\begin{itemize}
    \item which properties of the graph render the Network Inference Problem
        unfeasible or hard to solve. This relates to the question of
        identifiability of the contagion model, and understanding how the
        constants which appear on convergence rate bounds for edge weights
        estimators relate to topological properties of the graph.
    \item how domain specific knowledge about the graph to be learned can be
        exploited to make the Network Inference Problem easier to solve.
        A natural framework for this is a Bayesian one, where the domain
        specific knowledge is encoded as a Bayesian prior. To the best of our
        knowledge, this approach has not been explored beyond MAP estimation
        under a sparse (Laplace) prior in the context of discrete-time models.
\end{itemize}

The organization of the current report is as follows: we conclude the
introduction with a review of related works. The Generalized Linear Cascade
(GLC) model from \cite{pouget} is presented in Section 2. Section 3. explores
the question of identifiablity of the model. Section 4. presents both the
frequentist and Bayesian approach for inference of the GLC model as well as
a novel active learning formulation of the problem. The code used in our
experiments can be found in the Appendix.

\paragraph{Related works.}
The study of edge prediction in graphs has been an active field of research for
over a decade~\cite{Nowell08, Leskovec07, AdarA05}.~\cite{GomezRodriguez:2010}
introduced the {\scshape Netinf} algorithm, which approximates the likelihood of
cascades represented as a continuous process.  The algorithm, relying on
Maximum-Likelihood and convex relaxations was improved in later
work~\cite{gomezbalduzzi:2011}, but is not known to have any theoretical
guarantees beside empirical validation on synthetic networks.

Let us denote by $m$ the number of nodes in the graph and for a given node, we
denote by $s$ its in-degree. In what follows, we call a \emph{recovery
guarantee} an upper bound on the number of observations required (expressed as
a function of $s$ and $m$) to learn the incoming edges of a node to a fixed
arbitrary accuracy level. \cite{Netrapalli:2012} studied the discrete-time
version of the independent cascade model and obtained the first ${\cal O}(s^2
\log m)$ recovery guarantee on general networks, by using unregularized MLE and
making a {\it correlation decay\/} assumption, which limits the number of new
infections at every step.  They also suggest a {\scshape Greedy} algorithm,
with provably good performance in tree graphs, which maintains a counter of
each node's antecedents, i.e. the set of nodes infected at a prior time step to
its infection time. The work of~\cite{Abrahao:13} studies the same
continuous-model framework as \cite{GomezRodriguez:2010} and obtains an ${\cal
O}(s^9 \log^2 s \log m)$ support recovery algorithm, without the
\emph{correlation decay} assumption.  \cite{du2013uncover,Daneshmand:2014}
propose Lasso regularization of MLE for recovering the weights of the graph
under a continuous-time independent cascade model. The work
of~\cite{du2014influence} is slightly orthogonal to ours since they suggest
learning the \emph{influence} function, rather than the parameters of the
network directly.

More recently, the continuous process studied in previous work has been
reformulated as a Hawkes process, with recent papers
\cite{linderman2014discovering, linderman2015scalable, simma2012modeling},
focusing on Expectation-Maximization, and Variational Inference methods to learn
these processes. Because of the discrete nature of the GLC model, we hope to
bring a better understanding to the link between the properties of a graph and
its \emph{learnability}. We wish to further explore the idea of non-product
priors raised in \cite{linderman2014discovering, linderman2015scalable}, since
the experimental validation of their work focused on simple graph priors.
Finally, the \emph{Active Learning} formulation is, to the best of the
authors' knowledge, novel in this context.

\section{Model}

The GLC model is described over a directed graph $G = (V, \Theta)$. Denoting by
$k=|V|$ the number of nodes in the graph, $\Theta\in\R_{+}^{k\times k}$ is the
matrix of edge weights. Note that $\Theta$ implicitly defines the edge set $E$ of
the graph through the following equivalence:
\begin{displaymath}
    (u,v)\in E\Leftrightarrow \Theta_{u,v} > 0,\quad
    (u,v)\in V^2
\end{displaymath}

The time is discretized and indexed by a variable $t\in\N$. The nodes can be in
one of three states: \emph{susceptible}, \emph{infected} or \emph{immune}.
Let us denote by $S_t$ the set of nodes susceptible at the beginning of time
step $t\in\N$ and by $I_t$ the set of nodes who become infected at this time
step. The following dynamics:
\begin{displaymath}
    S_0 = V,\quad S_{t+1} = S_t \setminus I_t
\end{displaymath}
expresses that the nodes infected at a time step are no longer susceptible
starting from the next time step (they become part of the immune nodes).

The dynamics of $I_t$ are described by a random Markovian process. Let us
denote by $x_t$ the indicator vector of $I_t$ at time step $t\in\N$. $x_0$ is
drawn from a \emph{source distribution} $p_s:\{0,1\}^n\to[0,1]$. For $t\geq 1$,
we have:
\begin{equation}
    \label{eq:markov}
    \forall i\in S_t,\quad
    \P\big(x_{i}^{t} = 1\,|\, x^{t-1}\big) = f(\bt_i\cdot x^{t-1})
\end{equation}
where $\bt_i$ is the $i$th column of $\Theta$. The function $f:\R\to[0,1]$ can
be interpreted as the inverse link function of the model. Finally, the
transitions in \cref{eq:markov} occur independently for each $i$. A cascade
continues until no infected nodes remains.

We refer the reader to \cite{pouget} for a more complete description of the
model and examples of common contagion models which can be interpreted as
specific instances of the GLC model.

\section{Identifiability}

It follows from Section 2, that a source distribution $p_s$ and
\cref{eq:markov} together completely specify the distribution $p$ of a cascade
$\mathbf{x} = (x_t)_{t\geq 0}$:
\begin{equation}
    \label{eq:dist}
    p(\bx;\Theta) = p_s(x^0)\prod_{t\geq 1}\prod_{i\in S_t}
    f(\bt_i\cdot x^{t-1})^{x^t_i}\big(1-f(\theta_i\cdot x^{t-1})\big)^{1-x_i^t}
\end{equation}

In what follows, we consider that $p_s$ and $f$ are fixed. Hence the only
parameter of the model is the matrix $\Theta$ of parameters that we wish to
learn. We note that the problem of learning $\Theta$ might be ill-defined in
case of unidentifiability. We recall the definition of identifiability below.

\begin{definition}
    We say that the model $\mathcal{P}
    = \big\{p(\cdot\,;\,\Theta)\,|\,\Theta\in\R_+^{k\times k}\big\}$ is
    identifiable iff:
    \begin{displaymath}
        \forall \Theta,\Theta'\in\R_+^{k\times k},\quad
        p(\cdot\,;\,\Theta) = p(\cdot\, ;\, \Theta') \Rightarrow \Theta = \Theta'
    \end{displaymath}
\end{definition}

For the GLC model, identifiability is not guaranteed. To develop an intuition
about why this might be the case, we present below examples of
unidentifiability because of two obstructions: lack of information and parent
ambiguity.

\begin{figure}
\begin{center}
    \includegraphics[scale=0.8]{example.pdf}
\end{center}
\caption{Example of a graph which can be unidentifiable, both because of lack
    of information or parent ambiguity if the source distribution is chosen
adversarially.}
\label{fig:ident}
\end{figure}

\vspace{.5em}

\begin{example}[Lack of information]
    Let us consider the graph in \cref{fig:ident} and the source distribution
    $p_s$ defined by $p_s\big((0, 1, 0)\big) = 1$. In other words, node $2$ is
    always the only infected node at $t=0$. Then, it is clear that all
    casacades die at $t=1$ at which node 3 becomes infected with probability
    $f(\theta_{2,3})$. This probability does not depend on $\theta_{1,3}$.
    Hence all matrices $\Theta$ compatible with the graph (defining the same
    edge set) and for which $\theta_{2,3}$ is constant yield the same cascade
    distribution according to \cref{eq:dist}.
\end{example}

\vspace{.5em}

\begin{example}[Parent ambiguity]
Let us consider again the graph in \cref{fig:ident} and let us consider the
case where the source distribution is such that $p_s\big((1,1,0)\big) = 1$,
\emph{i.e} nodes 1 and 2 are always jointly infected at time $t=0$. Then it is
clear that cascades will all die at time $t=1$ where node 3 becomes infected
with probability $f(\theta_{1,3} + \theta_{2,3})$. This implies that all
matrices $\Theta$ which are compatible with the graph in \cref{fig:ident}
(defining the same edge set) and for which $\theta_{1,3} + \theta_{2,3} = c$
for some $c\in \R_+$ define the same cascade probability distribution.
\end{example}

\vspace{.5em}

It appears from these examples show that the question of identifiability is
closely related to the question of reachability of nodes from the source. The
following proposition shows that with the right assumptions on the source
distribution we can guarantee identifiability of the model.

We will slightly abuse the notation $p_s$ and for $u\in V$ write:
\begin{displaymath}
    p_s(w) \eqdef \sum_{\bx: \bx_w = 1} p_s(\bx)
\end{displaymath}
and similarly for a set of vertics $S$,  $p_s(S)$ is defined by replacing the
condition $\bx_w = 1$ by $\bx_S = 1$ in the summation condition in the above
equation. Furthermore, we write $u\leadsto v$ if there is a directed path from
$u$ to $v$ in $G$ and $u\leadsto v\leadsto w$ if there is a directed path from
$u$ to $w$ passing through $v$ in $G$.

\begin{proposition}
Under the assumptions:
\begin{enumerate}
    \item $\forall z\in \mathbf{\R_+},\; f(z)< 1$
    \item for all $S\subseteq V$ with $|S|\geq 2$, $p_s(S)<1$.
    \item $\forall (u,v)\in V^2$, there exists $w$ such that $p_s(w)\neq 0$ and
        $w\leadsto u\leadsto v$.
\end{enumerate}
then the GLC model is identifiable.
\end{proposition}

\begin{proof}[Proof sketch] We only give a quick proof sketch here and defer
    the details to the final version of this report. Let us consider $\Theta
    \neq \Theta'$. The goal is to prove that $p(\cdot\,;\,\Theta)\neq
    p(\cdot\,\;\, \Theta')$. Since $\Theta \neq \Theta'$ there exists a pair
    $(u,v)$ such that $\Theta_{u,v} \neq \Theta'_{u,v}$. Using assumption 3. we
    are guaranteed that with non-zero probability there will exist a time-step
    $t$ such that $x^t_u = 1$. We then distinguish two cases:
    \begin{itemize}
        \item $t=0$, in this case assumption 2. guarantees that with non-zero
            probability $u$ will be the only infected node at $t=0$.
        \item $t\geq 1$, in this case assumption 1. guarantees that with
            non-zero probability $u$ will be the only infected node at
            time-step $t$ (because the transitions in \cref{eq:markov} occur
            independently).
    \end{itemize}
    In both cases, it then follows from the form of \cref{eq:dist} after
    conditioning on $x^{t'}$ for $t' < t$ that the probability that $v$ will be
    infected at $t+1$ is different under $\Theta$ and under $\Theta'$.
\end{proof}


\begin{remark}
Intuitively, the assumptions of the proposition can be interpreted as follows:
assumption 2. prevents parent ambiguity at the source; assumption 1. prevents
parent ambiguity from occurring at later time-steps in the cascade even when
there is no ambiguity at the source; finally, assumption 3. prevents lack of
information.

When the graph $G$ is strongly connected, examples of simple source
distributions which satisfy the assumptions of the proposition include a single
source chosen uniformly at random among the vertices of the graph, or a source
where each node in $G$ is included independently with some probability $p>0$.

Finally, assumption 1 can be relaxed to a weaker assumption. Under this weaker
assumption, Proposition 2. becomes an exact characterization of the
identifiability of the model. Again, we defer the details to the final version
of this report.
\end{remark}

\section{Inference}

\subsection{Maximum-Likelihood Estimation}
\label{MLEsec}

It follows from the form of \cref{eq:dist} that the log-likelihood of $\Theta$
given cascade data can be written as the sum of $k$ terms, each term $i$ only
depending on $\bt_i$. Hence, each $\bt_i$ can be learnt separately by solving
a node-specific optimization problem. Specifically, for a given node $i$, if we
concatenate together all the time steps $t$ where this node was susceptible and
write $y^t = x^{t+1}_i$, \emph{i.e} whether or not the node became infected at
the next time step, the MLE estimator for $\bt_i$ is obtained by solving the
following optimization problem:
\begin{equation}
    \label{eq:mle}
    \hat{\theta}\in \argmax_\theta \sum_{t} y^t\log f(\theta\cdot x^t)
    + (1-y^t) \log \big(1 - f(\theta\cdot x^t)\big)
\end{equation}
It is interesting to note that at the node-level, doing MLE inference for the
GLC model is exactly amounts to fitting a Generalized Linear Model. When $f$ is
log-concave as is the case in most examples of GLC models, then the above
optimization problem becomes a convex optimization problem which can be solved
exactly and efficiently. The code to perform MLE estimation can be found in the
appendix, file \textsf{mle.py}.

Furthermore, as usual for an MLE estimator, we obtain that $\hat{\theta}$ is
asymptotically consistent under mild assumptions.

\begin{proposition}
    Under the assumptions of Proposition 2., and further assuming that
    $\lim_{z\to\infty} f(z) = 1$, the optimization program \eqref{eq:mle} has
    a unique solution $\hat{\theta}$ which is an asymptotically consistent
    estimator of $\bt_i$.
\end{proposition}

\subsection{Bayesian Approach}

In this section, we adopt a Bayesian approach to the Network Inference problem.
Since we assume that the state of the source $x^0$ is observed, the choice of
$p_s$ is not relevant to the inference of $\Theta$ (as long as it satisfies the
conditions of Proposition 2.). We enrich the model described in \cref{eq:dist}
by placing a prior $\mathcal{D}$ on $\Theta$.

\paragraph{Choice of a Graph Prior.} Following the standard assumption that in
many applications, graphs are sparse (\emph{i.e} contain few edges), prior
works have focused on MLE estimation with LASSO regularization ($\ell_1$-norm
penalization). This can be interpreted in the Bayesian world as doing MAP
estimation with an independent Laplace prior on $\Theta$:
\begin{displaymath}
\Theta\sim \prod_{i,j} Laplace(0,\lambda)
\end{displaymath}

However, in many domains of applications, we know much more about the possible
structure of the graph $G$ to be learned and by placing more informative priors on
$\Theta$ we can:
\begin{itemize}
\item take into account prior knowledge of edges' existence (or lack thereof)
\item take into account knowledge of the graph structure, such as density of
    triangles, diameter, degree distribution, clustering coefficient, etc.
\end{itemize}

A commonly used prior for graphs is the Exponential Random Graph Model (ERGM),
which allows flexible representations of networks and Bayesian inference. The
distribution of an ERGM family is defined by a feature vector $s(G)$ and by the
probability distribution:
\begin{displaymath}
P(G | w) \propto \exp \left( s(G)\cdot w \right)
\end{displaymath}

\paragraph{Inference.} If the prior has a product form, \emph{i.e} can be
factorized node-by-node, then the MAP estimation problem has the same
decomposability property as the MLE estimator described in Section 4.1: it is
still possible to learn the columns $\bt_i$ of $\Theta$ by solving independent
optimization problems.

However for a general ERGM, the prior might not have a product form. For
example a prior could encourage the presence of triangles or stars by including
the number of triangles or stars in the feature vectors $s(G)$. It is easy to
see that in this case the prior won't have a product form. Though
straightforward MCMC could be applied here, recent
work~\cite{caimo2011bayesian, koskinen2010analysing, robins2007recent} has
shown that ERGM inference has slow convergence and lack of robustness and
develop better alternatives to the naive MCMC formulation.


\paragraph{Experiments.}
Experiments using an ERGM prior are ongoing, hence we only present results for
MCMC on product-type priors. We used the library PyMC to sample from the
posterior distribution using MCMC\@ (see paragraph \emph{Results} below). This
method is very slow however, and scales badly in the number of cascades and
number of nodes. We could greatly benefit from using an alternative method:
\begin{itemize}
\item EM\@. This approach was used in \cite{linderman2014discovering,
    simma2012modeling} to learn the parameters of a Hawkes process, a closely
    related inference problem.
\item Variational Inference. This approach was used
    in~\cite{linderman2015scalable} as an extension to
    \cite{linderman2014discovering}. Considering the scalabilty of their
    approach, we hope to apply their method to our problem here, due to the
    similarity of the two processes, and to the computational constraints of
    running MCMC over a large parameter space.
\end{itemize}

\begin{figure}
\subfloat[][50 cascades]{
\includegraphics[scale=.4]{../simulation/plots/2015-11-05_22:52:30.pdf}}
\subfloat[][100 cascades]{
\includegraphics[scale=.4]{../simulation/plots/2015-11-05_22:52:47.pdf}}\\
\subfloat[][150 cascades]{
\includegraphics[scale=.4]{../simulation/plots/2015-11-05_22:53:24.pdf}}
\subfloat[][200 cascades]{
\includegraphics[scale=.4]{../simulation/plots/2015-11-05_22:55:39.pdf}}\\
\subfloat[][250 cascades]{
\includegraphics[scale=.4]{../simulation/plots/2015-11-05_22:57:26.pdf}}
\subfloat[][1000 cascades]{
\includegraphics[scale=.4]{../simulation/plots/2015-11-05_22:58:29.pdf}}
\caption{Bayesian Inference of $\Theta$ with MCMC using a $Beta(1, 1)$ prior on
each edge.  For each figure, the plot $(i, j)$ on the $i^{th}$ row and $j^{th}$
column represent a histogram of samples taken from the  posterior of the
corresponding edge $\Theta_{i, j}$. The red line indicates the true value of the
edge weight.  If an edge does not exist (has weight $0$) the red line is
confounded with the y axis.}
\label{betapriorbayeslearning}
\end{figure}

\paragraph{Results.} We ran some experiments on a simple network with 4 nodes
with $\binom{4}{2}=6$ parameters to learn with the MCMC package PyMC\@. We plot
in Figure~\ref{betapriorbayeslearning} the progressive learning of $\Theta$ for
increasing numbers of observations. Of note, since the GLMC model does not
include self-loops, the diagonal terms are never properly learned, which is
expected but not undesirable. We notice that the absence of an edge is
(relatively) quickly learned: the posterior on edges with no weight converges
to $0$ after $100$ cascades. For existing edges though, the posterior visually
concentrates only after $1000$ cascades, which is unreasonably high considering
the small number of parameters that we are learning in this experiment.

\subsection{Active Learning}

Finally, we propose an Active Learning formulation of the Network Inference
problem. In this setup, the source $x^0$ is no longer drawn from a random
distribution $p_s$ but a single source $i$ is chosen by the designer of the
experiment. Once a source is drawn, a cascade is drawn from the Markovian
process $\mathcal{D'}$ described in \eqref{eq:markov}.

Our suggested selection stratgey is to select at each time step the source is
to select the node which maximises the information gain (mutual information) on
the parameter $\Theta$ resulting from observing the cascade $\bx$. The mutual
information can be computed node-by-node by sampling:
\begin{equation*}
I(\bx ,\Theta | x^0 = i) = \mathbb{E}_{\substack{\Theta \sim D_t \\
x | \Theta, i \sim D'}}\log p(x|\Theta) - \mathbb{E}_{\substack{\Theta \sim D_t
\\ x' | \Theta, i \sim D'}} \log p(x')
\end{equation*}
where $D_t$ denotes the posterior distribution on $\Theta$ after the first $t$
cascades.  Not that the second term involves marginalizing over $\Theta$:
$p(x') = \mathbb{E}_{\Theta \sim D_t} p(x'| \Theta)$. The full specification of
our Active Learning procedure is described in Algorithm 1. For the final
version of this project, we plan to test experimentally the learning speedup
induced by using Active Learning compared to independent observations from
a random source.

\begin{algorithm}
\caption{Active Bayesian Learning through Cascades (ABC)}
\begin{algorithmic}[1]
\State $\Theta \sim \mathcal{D}_0 = \mathcal{D}$ \Comment{Initial prior on
$\Theta$}
\While{True}
\State $i \leftarrow \argmin_{i} I(\Theta, \bx | x_0 = i)$
\Comment{Maximize mutual information}
\State Observe $(x_t)_{t\geq1} |x_0 = i$ \Comment{Observe cascade from source}
\State $\mathcal{D}_{t+1} \gets \text{posterior of $\Theta \sim \mathcal{D}_t$
given $(x_t)_{t\geq1}$}$ \Comment{Update posterior on $\theta$}
\EndWhile
\end{algorithmic}
\end{algorithm}

\bibliography{sparse}{}
\bibliographystyle{abbrv}
\appendix
\section{\textsf{mle.py}}
\input{mle}
\section{\textsf{bayes.py}}
\input{bayes}
\end{document}