graphs

1 files changed, 6 insertions, 9 deletions

diff --git a/paper/sections/experiments.tex b/paper/sections/experiments.tex
index 36e48eb..1b72753 100644
--- a/paper/sections/experiments.tex
+++ b/paper/sections/experiments.tex
@@ -22,16 +22,16 @@
 \captionof{figure}{Figures (a) and (b) report the F$1$-score in $\log$ scale
 for 2 graphs as a function of the number of cascades $m$: (a) Barabasi-Albert graph, $300$ nodes, $16200$ edges. (b)
 Watts-Strogatz graph, $300$ nodes, $4500$ edges. Figure (c) plots the Precision-Recall curve for various values for $\lambda$ ofr a Holme-Kim graph ($200$ nodes, $9772$ edges). Figures (d) and (e) report the
-$\ell_2$-norm $\|\hat \Theta - \Theta\|_2$ for a Kronecker graph which is: (d) exactly sparse (e) non-exactly sparse, as a function of the number of cascades $n$. Figure (f) plots the F$1$-score for the Watts-Strogatz graph as a function of $p_{init}$}
+$\ell_2$-norm $\|\hat \Theta - \Theta\|_2$ for a Kronecker graph which is: (d) exactly sparse (e) non-exactly sparse, as a function of the number of cascades $n$. Figure (f) plots the F$1$-score for the Watts-Strogatz graph as a function of $p_{init}$.}
 \label{fig:four_figs}
 \end{table*}
 
-In this section, we validate empirically the results and assumptions of Section~\ref{sec:results} for different initializations of parameters ($n$, $m$, $\lambda$) and for varying levels of sparsity. We compare our algorithm to two different state-of-the-art algorithms: \textsc{greedy} and \textsc{mle} from \cite{Netrapalli:2012}. As an extra benchmark, we also introduce a new algorithm \textsc{lasso}, which approximates our \textsc{sparse mle} algorithm.
+In this section, we validate empirically the results and assumptions of Section~\ref{sec:results} for for varying levels of sparsity and different initializations of parameters ($n$, $m$, $\lambda$, $p_{\text{init}}$), where $p_{\text{init}}$ is the initial probability of a node being a source node. We compare our algorithm to two different state-of-the-art algorithms: \textsc{greedy} and \textsc{mle} from \cite{Netrapalli:2012}. As an extra benchmark, we also introduce a new algorithm \textsc{lasso}, which approximates our \textsc{sparse mle} algorithm.
 
 \paragraph{Experimental setup}
 We evaluate the performance of the algorithms on synthetic graphs, chosen for their similarity to real social networks. We therefore consider a Watts-Strogatz graph ($300$ nodes, $4500$ edges) \cite{watts:1998}, a Barabasi-Albert graph ($300$ nodes, $16200$ edges) \cite{barabasi:2001}, a Holme-Kim power law graph ($200$ nodes, $9772$ edges) \cite{Holme:2002}, and the recently introduced Kronecker graph ($256$ nodes, $10000$ edges) \cite{Leskovec:2010}. Undirected graphs are converted to directed graphs by doubling the edges.
 
-For every reported data point, we sample edge weights and generate $n$ cascades from the Independent Cascade model for $n \in \{100, 500, 1000, 2000, 5000\}$. We compare for each algorithm the estimated graph $\hat {\cal G}$ with ${\cal G}$. The initial probability of a node being a source is fixed to $0.05$, i.e. an average of $15$ nodes source nodes per cascades for all experiments. All edge weights are chosen uniformly in the interval $[0.2, 0.7]$, except when testing for approximately sparse graphs (see paragraph on robustness). Adjusting for the variance of our experiments, all data points are reported with at most a $\pm 1$ error margin. The parameter $\lambda$ is chosen to be of the order ${\cal O}(\sqrt{\log m / (\alpha n)})$.
+For every reported data point, we sample edge weights and generate $n$ cascades from the Independent Cascade model for $n \in \{100, 500, 1000, 2000, 5000\}$. We compare for each algorithm the estimated graph $\hat {\cal G}$ with ${\cal G}$. The initial probability of a node being a source is fixed to $0.05$, i.e. an average of $15$ nodes source nodes per cascades for all experiments, except for Figure~\label{fig:four_figs} (f). All edge weights are chosen uniformly in the interval $[0.2, 0.7]$, except when testing for approximately sparse graphs (see paragraph on robustness). Adjusting for the variance of our experiments, all data points are reported with at most a $\pm 1$ error margin. The parameter $\lambda$ is chosen to be of the order ${\cal O}(\sqrt{\log m / (\alpha n)})$.
 
 \paragraph{Benchmarks}
 
@@ -39,19 +39,16 @@ We compare our \textsc{sparse mle} algorithm to 3 benchmarks: \textsc{greedy} an
 $$\hat \theta_i \in \arg \min_{\theta} \sum_{t \in {\cal T}} |f(\theta_i\cdot x^t) - x_i^{t+1}|^2 + \lambda \|\theta_i\|_1$$
 \textsc{Lasso} has the merit of being both easier and faster to optimize numerically than the other convex-optimization based algorithms. It approximates the $\textsc{sparse mle}$ algorithm by making the assumption that the observations $x_i^{t+1}$ are of the form: $x_i^{t+1} = f(\theta_i\cdot x^t) + \epsilon$, where $\epsilon$ is random white noise. This is not valid in theory since $\epsilon$ \emph{depends on} $f(\theta_i\cdot x^t)$, however the approximation is validated in practice.
 
-We did not benchmark against other known algorithms (\textsc{netrate} \cite{gomezbalduzzi:2011} and \textsc{first edge} \cite{Abrahao:13}) due to the discrete time assumption. These algorithms also suppose a single-source model, whereas \textsc{sparse mle}, \textsc{mle}, and \textsc{greedy} do not. Learning the graph in the case of a multi-source cascade model is intuitively harder but more realistic, since we rarely have access to ``patient 0'' in practice.
+We did not benchmark against other known algorithms (\textsc{netrate} \cite{gomezbalduzzi:2011} and \textsc{first edge} \cite{Abrahao:13}) due to the discrete time assumption. These algorithms also suppose a single-source model, whereas \textsc{sparse mle}, \textsc{mle}, and \textsc{greedy} do not. Learning the graph in the case of a multi-source cascade model is harder (see Figure~\ref{fig:four_figs} (f)) but more realistic, since we rarely have access to ``patient 0'' in practice.
 
 \paragraph{Graph Estimation}
 In the case of the \textsc{lasso}, \textsc{mle} and \textsc{sparse mle} algorithms, we construct the edges of $\hat {\cal G} : \cup_{j \in V} \{i : \Theta_{ij} > 0.1\}$, \emph{i.e} by thresholding. The true positives are the edges which appear both in ${\cal G}$ and $\hat {\cal G}$ and the true negatives are the edges which appear in neither. Finally, we report the F1-score$=2 \text{precision}\cdot\text{recall}/(\text{precision}+\text{recall})$, which considers the number of true edges recovered by the algorithm over the total number of edges returned by the algorithm (\emph{precision}) with the number of true edges recovered by the algorithm over the total number of edges it should have recovered (\emph{recall}).
 
 Over all experiments, \textsc{sparse mle} achieves higher rates of precision,
-recall, and f1-score. \textsc{sparse mle} is also robust to approximate
-sparsity, and displays a faster convergence of the $\ell2$-norm than any other
-benchmark. Interestingly, both \textsc{mle} and \textsc{sparse mle} perform
-exceptionally well on the Watts-Strogatz graph. The recovery rate converges at
+recall, and f1-score. Interestingly, both \textsc{mle} and \textsc{sparse mle} perform exceptionally well on the Watts-Strogatz graph. The recovery rate converges at
 around $5000$ cascades, which is more than $15$ times the number of nodes. By
 contrast, \textsc{sparse mle} achieves a reasonable F$1$-score of $.75$ for roughly $500$ observed cascades. 
 
 \paragraph{Quantifying robustness}
 
-The previous experiments only considered graphs with strong edges. To test the algorithms in the approximately sparse case, we add sparse edges to the previous graphs according to a bernoulli variable of parameter $1/3$ for every non-edge, and drawing a weight uniformly from $[0,0.1]$. The results are reported in Figure~\ref{fig:four_figs} by plotting the convergence of the $\ell2$-norm error, and show that both the \textsc{lasso}, followed by \textsc{sparse mle} are the most robust to noise.
+The previous experiments only considered graphs with strong edges. To test the algorithms in the approximately sparse case, we add sparse edges to the previous graphs according to a bernoulli variable of parameter $1/3$ for every non-edge, and drawing a weight uniformly from $[0,0.1]$. The results are reported in Figure~\ref{fig:four_figs} (d)-(e) by plotting the convergence of the $\ell2$-norm error, and show that both the \textsc{lasso}, followed by \textsc{sparse mle} are the most robust to noise.