The effect of transmission variance on observer placement for sourcelocalization
Abstract
Detecting where an epidemic started, i.e., which node in a network was the source, is of crucial importance in many contexts. However, finding the source of an epidemic can be challenging, especially because the information available is often sparse and noisy. We consider a setting in which we want to localize the source based exclusively on the information provided by a small number of observers – i.e., nodes that can reveal if and when they are infected – and we study where such observers should be placed.
We show that the optimal observer placement depends not only on the topology of the network, but also on the variance of the nodetonode transmission delays. We consider both lowvariance and highvariance regimes for the transmission delays and propose algorithms for observer placement in both cases. In the lowvariance regime, it suffices to only consider the networktopology and to choose observers that, based on their distances to all other nodes in the network, can distinguish among possible sources. However, the highvariance regime requires a new approach in order to guarantee that the observed infection times are sufficiently informative about the location of the source and do not get masked by the noise in the transmission delays; this is accomplished by additionally ensuring that the observers are not placed too far apart.
We validate our approaches with simulations on three realworld networks. Compared to stateoftheart strategies for observer placement, our methods have a better performance in terms of sourcelocalization accuracy for both the low and the highvariance regimes.
Keywords
Source localization Epidemics Sensor placementIntroduction
The difficulties faced in finding the optimal observers for source localization are twofold. First, computing the likelihood of a node being the source conditional on the available observations can be computationally prohibitive (Shah and Zaman 2011; Pinto et al. 2012); evaluating the probability of correct localization given a set of observers is, in general, even harder. Second, the optimal selection of a limited number of observers is NPhard, even when the transmission delays are deterministic. We take a principled approach that begins with considering deterministic transmission delays (zerovariance regime), and we build on this intuition in order to develop heuristics for both lowvariance and highvariance regimes for the transmission delays.^{1}
Model and problem statement
Transmission model. We assume that the epidemic spreads in a known contact network. The transmission delay through edge uv, i.e., the time it takes for a node u to infect a neighbor node v is encoded by the random variable X _{ uv }.
We assume a transmission model which is both natural and versatile as it comprises deterministic transmissions, which we call zerovariance, and arbitrary random independent transmission models. We study, in particular, how the amount of randomness (i.e., the variance of X _{ uv }) in the transmission delays affects the choice of observers for source localization. Towards this, we are the first to separately analyze two different regimes for the amount of randomness of the transmission delays: lowvariance and highvariance. A dichotomy exists between the two, and our approach for observer placement differs.
We use the SI epidemic model adopted, e.g., in (Pinto et al. 2012; Luo and Tay 2012). Nonetheless, since our methods for source localization only uses the time at which the sensors are first infected (no assumption on recovery or reinfection dynamics is made), they can be applied to any epidemic model, including the well known SIS or SIR (provided that nodes do not recover before infecting their neighbors).
Source localization. We assume that there is a single source that initiates the epidemic, an extension of our results to the case the case of multiple sources could use the recent work by Zhang et al. (2015) on a related problem and is left for future work.
Let \(\mathcal {O} \subseteq V\) be the set of observer nodes (which we will select). We assume we know the time at which each observer is infected, and we refer to this vector of infection times as \(T_{\mathcal {O}}\). Knowing \(T_{\mathcal {O}}\) is a standard and realistic assumption (Netrapalli and Sanghavi 2012). We want to identify the source using only the information contained in \(T_{\mathcal {O}}\).
which is a widely used metric for source localization (see, e.g., (Shah and Zaman 2011; Pinto et al. 2012; Louni and Subbalakshmi 2014)). In our experiments we also evaluate another important metric, the expected distance between the estimated source and the real source (Celis et al. 2015; Louni et al. 2015), i.e., \(\mathbf {E}[\!d(s^{\star },\hat {s})]\), where d denotes the distance between two nodes in the network.
In “Metrics for source localization” section we present several alternatives to these two metrics, including worstcase metrics, and show that optimizing different metrics can require different sets of observers.
Main contributions
Lowvariance regime. When the variance in the transmission delays is low (see “The lowvariance regime” section), we prove that the set of optimal observers is exactly the optimal set for the zerovariance regime. In the zero and low variance regime, both the probability of success \(\mathcal {P}_{s}\) (as well as other possible metrics of interest) can be explicitly computed. Despite this seeming simplicity, the problem remains NPhard. We tackle the problem by using its connection with the wellstudied related Double Resolving Set (DRS) problem (Cáceres et al. 2007) that minimizes the number of observers for correct localization. This minimum number is, in many cases, still prohibitively large, and can be as much as n−1, hence we cannot use this approach directly. However, from the connection between observer placement and DRS, we find inspiration for our algorithm which, by selecting one observer at a time until the budget is exhausted in order to reach a DRS set, greedily improves \(\mathcal {P}_{s}\).
Highvariance regime. When the noise in the transmission delays is high, it is no longer negligible and it poses an additional challenge to source localization; in effect, the accumulation of noise from node to node as the epidemic spreads might no longer enable us to distinguish between two potential sources, especially when they are both far from all observers. Hence, we must strengthen the requirements for observer placement in order to ensure that the nodes can be distinguished by observers that are near to them; this nearness is a function of the noise, of the budget k, and of the network topology. We define a novel objective function that both maximizes the success probability and imposes a uniform spread of observers in the network. Taking inspiration from the lowvariance regime, we design an algorithm that greedily maximizes this new objective (see “The highvariance regime” section).
Empirical results. In “Empirical results” section, we evaluate our algorithms on three different realworld datasets that represent different application areas for source localization and different network topologies. First, we take a community of people living in the proximity of a university campus (Aharony et al. 2011), a typical network for the transmission of airborne diseases. Second, we take a community of students exchanging messages over a Facebooklike social network (Opsahl and Panzarasa 2009) through which ideas and trends can propagate. Finally, we consider the road network of the state of California (California Road Network): this captures geographical networks that can model the transmission of a disease between connected communities or the diffusion of contaminants, e.g., through a hydrological network. We show that our methods perform favourably against stateoftheart approaches in both the low and the highvariance regimes (see “Comparison against benchmarks” section). For the lowvariance regime, we further compare our method against two other natural greedy heuristics for observer placement (see “Comparison with benchmarks” section); we show that our approach outperforms the rest. Moreover, in the empirical results, the dichotomy between the low and highvariance regimes becomes apparent.
Preliminaries
Model
Let \(\mathcal {G} = (V, E, w)\) be a weighted network. For ease of presentation we assume the graph is undirected and w _{ uv }=w _{ vu }; however our definitions and approach extend straightforwardly to the directed case. Assuming u is infected, the weight \(w_{uv}\in \mathbb {R}_{+}\) of edge u v∈E represents the expected time it takes for u to infect v. The edge weights induce a weighteddistance metric d on \(\mathcal {G}\): d(u,v) is the length of the shortest path from u to v. We also sometimes consider the minimum number of edges on a path connecting two nodes, which we call the hopsdistance.
We assume that the epidemic is initiated by a single unknown source s ^{⋆} at an unknown time t ^{⋆}. The fact that the time t ^{⋆} at which an epidemic starts is unknown adds a significant difficulty to the problem because a single observation is not per se informative. Instead, in order to localize the source, we must use the differences between the observed infection times.
If a node u gets infected at time t _{ u }, a noninfected neighbor v of u will become infected at time t _{ v }=t _{ u }+X _{ uv } where X _{ uv } is a random variable. A large part of the epidemic literature models transmission delays with exponential random variables. However we make a different modeling choice for two reasons. First, we are interested in decoupling the transmission variance and the average transmission time (for exponential random variables, mean and variance cannot be tuned independently). Second, in many applications it has been suggested that the transmission delays can be lessskewed than exponential random variables (Cha et al. 2009; Lessler et al. 2009; Vergu et al. 2010). For every edge uv we assume X _{ uv } to be a symmetric and nonnegative^{2} random variable. We do not make any strong assumption on the distribution of the transmission delays X _{ uv }: we only assume that their mean is equal to the edge weights, i.e., E[X _{ uv }]=w _{ uv } for every u v∈E, and that their variance is an increasing function of both the edge weight and of a variance parameter σ, that is, Var(X _{ uv })=g(w _{ uv },σ), where g depends on the particular distribution of X _{ uv } and g(x,0)=0 for all \(x\in \mathbb {R}^{+}\).
If the variance is zero, or if it is low compared to edge weights, network distances are a good proxy for time delays (see “Identification of the source class” section). We refer to this setting as a lowvariance regime, as opposed to the highvariance regime in which time delays are very noisy and network distances no longer work as a proxy for time delays.
Distance vectors and node equivalence
We start with a few definitions. Our setting is similar to that of Celis et al. (2015).
Definition 1
which means that \([\!u]_{\mathcal {O}} \neq [\!v]_{\mathcal {O}}\).
The problem of finding the minimumsize set of nodes S, such that for every u,v in a network there exist s _{ i },s _{ j }∈S for which d(u,s _{ i })−d(u,s _{ j })≠d(v,s _{ i })−d(v,s _{ j }) is known as the Double Resolving Set (DRS) Problem (Cáceres et al. 2007), while the minimum size of a DRS is known as the Double Metric Dimension (DMD) of the network. Our problem differs from DRS because we focus on the more realistic context in which, due to limited resources, we want to allocate a finite budget in order to optimize source localization^{3} (as opposed to minimizing the number of observers for perfect localization, which is, in many cases, still prohibitively large). However, the connection between our problem and DRS paves the way for a principled approach to observer placement.
We now define, for every v∈V, a distance vector, which, as we will see in Lemma 1, mathematically captures equivalence in a manner that is easy to work with.
Definition 2
(Distance Vector) Let \(\mathcal {G} = (V,E)\), \(\mathcal {O}\subseteq {V}\) with \(\mathcal {O}=k \geq 2\) and \(o_{1}\in \mathcal {O}\). For each node v∈V the distance vector of v with respect to o _{1} is \(\mathbf {d}_{s, o_{1}} \in \mathbb {R}^{k1}\) with entries d(v,o _{ i+1})−d(v,o _{1}) for 1≤i≤k−1.
The following lemma, similar in spirit to Lemma 3.1 in (Chen et al. 2014), shows that the equality between distance vectors of different nodes does not depend on the choice of the reference observer o _{1}.
Lemma 1
Let \(\mathcal {G} = (V,E)\) and \(\mathcal {O} \subseteq V\) with \(\mathcal {O}=k \geq 2\) and let u,v∈V. Then, \([u]_{\mathcal {O}}\)=\([v]_{\mathcal {O}}\) if and only if \(\mathbf {d}_{u, o_{1}} = \mathbf {d}_{v, o_{1}}\), independently of the choice of the reference observer o _{1}.
Metrics for source localization
In this section we define some possible metrics of interest for the sourcelocalization problem and we show that optimizing these metrics can effectively require different sets of observers.
For ease of exposition, we restrict ourselves to the zerovariance regime and we assume that the prior distribution on the position of the source is uniform.
In the zerovariance regime, the partition in equivalence classes is effectively the only factor for the localization of the source: if [s ^{⋆}] is a singleton, it is always possible to localize the source exactly based on the observed infection time; if it is not a singleton, we can only correctly identify the class to which s ^{⋆} belongs and we produce an estimated source \(\hat {s} \in [s^{\star }]\) sampling from [s ^{⋆}] uniformly.
We adopt two metrics to evaluate the performance of our algorithms: the success probability \(\mathcal P_{s}\) and the expected error distance \(\mathcal {D}\).
Note that \(\mathcal {P}_{s} = 1\) if and only if all equivalence classes are singletons.
where again \(\mathcal {D} = 0\) if and only if all equivalence classes are singletons. An analogous expression for the hopsdistance (instead of the weighted distance as in (3)) is also considered in the experimental evaluation in “Empirical results” section.
The lowvariance regime
Identification of the source class
We formalize how we can localize the source in the zerovariance setting, i.e., when X _{ uv }=w _{ uv } for every edge (u,v).
For every observer \(o_{i} \in \mathcal {O}\), denote by t _{ i } the time at which o _{ i } gets infected. In the zerovariance setting, the observed infection times of nodes o _{2},…,o _{ K } with respect to observer o _{1}, i.e., the vector τ=deft _{2}−t _{1},…,t _{ k }−t _{1}, is exactly the distance vector of the unknown source s ^{⋆} with respect to o _{1}. Then, if for every \(u,v \in V [u]_{\mathcal {O}}\neq [v]_{\mathcal {O}}\), the source can be always correctly identified by finding the node whose distance vector matches the observed infection times. Theorem 1 proves that this is true also in a more general lowvariance framework where we are always able to identify the equivalence class to which the real source belongs by looking at the distances between the distance vectors \(\{\mathbf {d}_{v, o_{1}}, v \in V\}\) and the vectors of infection times τ.
Theorem 1
If the transmission delays are such that for each u v∈E, X _{ uv }∈ [w _{ uv }(1−ε),w _{ uv }(1+ε)] with \(\varepsilon <\varepsilon _{0} \triangleq \frac {\delta }{4D}\) then for every \(v \in \ [s^{\star }] \\mathbf {d}_{v, o_{1}}  \boldsymbol \tau \_{\infty }\leq 2\varepsilon D\) and for every \(v \notin \ [s^{\star }] \\mathbf {d}_{v, o_{1}}  \boldsymbol \tau \_{\infty } > 2\varepsilon D\).
Proof
Since for every \(v \in [s^{\star }] \mathbf {d}_{v, o_{1}}=\mathbf {d}_{s^{\star }, o_{1}}\), we conclude that for every v∈ [s ^{⋆}], \(\\mathbf {d}_{v, o_{1}}  \boldsymbol \tau \_{\infty }\leq 2\varepsilon D\).
Note that here ε _{0} plays the role of σ _{0} in Fig. 1 in the sense that it is an upperbound on a regime in which the delays are effectively deterministic and the variance of the transmission delays does not affect the accuracy of source localization.
If additional conditions on the weights or on the network topology are made, more refined versions of Theorem 1 can be proven. For example, in a tree with integer weights, due to the uniqueness of the path between two any vertices, it can be shown that δ≥2 and Theorem 1 holds for \(\varepsilon <\varepsilon _{0} \triangleq \frac {1}{2D}\).
For the remainder of this section, we will assume ε<δ/4D, which we call the lowvariance regime.
Estimation of the source
Assume that a prior probability distribution on the identity of the source is given, i.e., that we know π(v)=defP(s ^{⋆}=v). After the source class \([s^{\star }]_{\mathcal {O}}\) is identified based on τ as described in “Identification of the source class” section, we let our estimated source \(\hat {s}\) be chosen at random from the conditional probability \(\phantom {\dot {i}\!}\pi _{[s^{\star }]}(u)\stackrel {\text {def}}{=}\mathbf {P}(s^{\star }=vv \in [s^{\star }])\). If a prior π is not known, we select the estimated source uniformly at random from [s ^{⋆}], which is equivalent to having a uniform prior π.
For ease of exposition, we focus on the case in which the prior distribution on the position of the source is uniform, hence π(v)=1/n for all v∈V. Our algorithms and observations can be easily extended to general priors.
Observer placement
Independently of the topology of the network \(\mathcal {G}\), the success probability \(\mathcal {P}_{s}\), as well as other possible metrics of interest, can be computed exactly in polynomial time (see, e.g., Eqs. (2) and (3)). In fact, due to Lemma 1 and Theorem 1, it is enough to compute the distance vector of Definition 1 for all the nodes. Nonetheless, if we have a budget k≥2 of nodes that we can choose as observers, finding the configuration that maximizes \(\mathcal {P}_{s}\) is an NPhard problem. This is a direct consequence of the hardness result of Chen et al. (2014).
Theorem 2
Let k≥2 be the budget on the number of nodes we can select as observers. Finding \(\mathcal {O} \subseteq V\) such that \(\mathcal {O} \in \text {argmax}_{\mathcal {O}=k}\mathcal {P}_{s}(\mathcal {O})\) is NPhard.
The proof follows straightforwardly with a reduction from the DRS problem (see Appendix B).
Our first main contribution in this paper is a solution to the budgeted observerplacement problem for general graphs.
For trees, the optimal observer placement can be find in polynomial time using dynamic programming techniques (Celis et al. 2015). In a general graph (with loops) the problem of source localization is made more challenging by the multiplicity of paths through which the epidemic can spread and for the same reason also finding an optimal observer set becomes much harder.
A first idea to solve observer placement on a general graph could be to use the latter result on a BFSapproximation of the graph. However, as mentioned in “Metrics for source localization” section, on a tree the optimal observer placement is contained in the leaves set. If we consider a nontree graph and take a BFSapproximation, the leaves of the BFS tree depend on where the BFStree is rooted. Hence using the result of (Celis et al. 2015) on a tree approximation it is not possible to guarantee high probability of success independently of the position of the source.
Our approach, presented in Algorithm 1, does not rely on a graph approximation. Moreover, it is specifically designed for the source localization problem and has a simple greedy structure: for every node v∈V, initialize \(\mathcal {O}\leftarrow \{v\}\) and iteratively add to \(\mathcal {O}\) the node u that maximizes the gain with respect to the success probability until we either run out of budget or \(\mathcal {P}_{s}=1\). Eq. 2 ensures that greedily maximizing the success probability is equivalent to greedily maximizing the number q of equivalence classes. When adding an element to the observer set, the partition in equivalence classes can be updated in linear time, total running time of our algorithm is O(k n ^{3}). Despite bypassing the NPhardness of the problem, this might not be sufficiently fast for very large networks. However, the procedure is extremely parallelizable (see, for example, the main for loop and the a r g m a x in the w h i l e loop).
The osbserver placement obtained through Algorithm 1 will be denoted LVOBS to emphasize the fact that it is designed for the case in which the variance is absent or very small (LV stands for lowvariance regime).
Comparison with benchmarks
As budgeted observer placement (even in the zerovariance setting) is NPhard, there is no optimal algorithm to compare against. Instead, we evaluate the performance of our algorithm against a set of natural benchmarks that have shown to have good performance in other works (Seo et al. 2012; Berry et al. 2006; Zhang et al. 2016) (see “Comparison against benchmarks” section for a discussion of these benchmarks, Figs. 1012 for the results).
Alternative objective functions. We further compare LVOBS against two other natural heuristics that also optimize an objective function greedily.
The first is an adapted version of the approximation algorithm for the DRS problem proposed by Chen et al. (2014) and described in Appendix A.
By stopping the greedy process after it selects k nodes, we can adapt in a natural way this approximation algorithm and create a heuristic for the budgeted version that we denote by Φ _{ ent }. We want to check if LVOBS actually reaches smaller values of \(\mathcal {P}_{s}\) compared to Φ _{ ent }.
The second is a direct minimization of the expected error distance \(\mathcal {D}= \mathbf {E}\, [\!d(s^{\star }, \hat {s})]\) of Eq. (3) that we denote by Φ _{ dist }. Even if LVOBS is not directly minimizing \(\mathcal {D}\), we want to compare the results we obtain in terms of \(\mathcal {D}\) with those obtain to Φ _{ dist } in order to check if, at least in some budget regimes, we can use the maximization of \(\mathcal {P}_{s}\) as a proxy for the minimization of \(\mathcal {D}\).
The results of our empirical evaluation are presented in Table 2 in Appendix C.
The results achieved by Φ _{ ent } and Φ _{ dist } are, on average, worse than those of Algorithm 1 both in terms of \(\mathcal {P}_{s}\) and of \(\mathcal {D}\), independently of the graph topology. We observe two exceptions. First, when k is very small: Φ _{ dist } reaches smaller values of \(\mathcal {D}\) compared to LVOBS, which can explained by the fact that Φ _{ dist } directly minimizes \(\mathcal {D}\) and that, when fewer observers are available the difference between the observer placements that maximize \(\mathcal {P}_{s}\) and minimize \(\mathcal {D}\) is greater. Second, for large k, on the Barabàsi Albert networks Φ _{ ent } gives, in average, larger \(\mathcal {P}_{s}\) than LVOBS. This is probably due to the fact that, for this class of graphs, the DMD is small, hence with a large value of k we approach the regime in which the objective function of Φ _{ ent }, designed to minimize the DMD of the network, is optimal.
The highvariance regime
When the variance is not guaranteed to be low, as defined in “The lowvariance regime” section, computing analytically the success probability  or other metrics of interest  is unfortunately not possible (except for very simple graphs, like the path network of Fig. 2, and for particular transmission delays, e.g., Gaussiandistributed).
When the variance is high, also the localization of the source is more challenging because the observed infection delays t _{ i }−t _{ j } can be misleading, especially if the corresponding observers o _{ i } and o _{ j } are far from the source. Take, for example, a path of length L where the two leaves are the only two observers and all edges have weight equal to 1. Figure 5a shows how the success probability \(\mathcal P_{s}\) decays faster for increasing values of L. Building on this observation, we propose a strategy for observer placement that enforces a controlled distance from a general source node to the observer set.
Source localization. For the highvariance case we localize the source using an adapted version of the algorithm proposed by Pinto et al. (2012) (see Appendix D for details). This adapted algorithm can be seen as a generalization to the highvariance regime of the source localization method presented in “Identification of the source class” section for the lowvariance regime.
Observer placement
Although we cannot control σ, we can control the path length between observers.
We make use of the following sufficient condition for a set to be a DRS, i.e., for an observer set to guarantee correct source localization.
Lemma 2
Let \(\mathcal {G}=(V, E)\) be a network, \(\mathcal {O} \subseteq V\). If for every u∈V there exist \(o_{1}, o_{2} \in \mathcal {O}\) such that there is a unique shortest path \(\mathcal {P}(o_{1}, o_{2})\) between o _{1} and o _{2} and \(u \in \mathcal {P}(o_{1}, o_{2})\), then \(\mathcal {O}\) is a DRS for G.
Proof
Since c>0, d(v,o _{3})>d(u,o _{3}) and d(v,o _{4})>d(u,o _{4}) giving a contradiction with v (and not u) being on the shortest path \(\mathcal {P}(o_{3}, o_{4})\). We conclude that (u,v) are resolved by either (o _{1},o _{3}) or (o _{1},o _{4}). □
The converse of this lemma is not true: If \(\mathcal {O}\) double resolves \(\mathcal {G}\), it is not even true that for every node u there must exist \(o_{1}, o_{2} \in \mathcal {O}\) such that u is contained in some shortest path between o _{1} and o _{2} of (see Fig. 5b).
Path covering strategy. We take Lemma 2 as a basis for deriving a path covering strategy for observer placement. In practice, the condition about the uniqueness of the shortest path is too strong and excludes many potentially useful observer nodes. Experimentally we see that in many practical situations two shortest paths differ only by a few nodes and the majority of nodes on the path are resolved by the two extreme nodes. This is why we relax the condition of Lemma 2 and we prefer, when the shortest path is not unique, to select one arbitrarily. Let S⊆V be a set of observers and L a positive integer: We call P _{ L }(S) the set of nodes that lie on a shortest path of length at most L between any two observers in the set S. Given a budget k, and a positive integer L, we denote by \(S^{\star }_{k, L}\) the set of k vertices that maximize the cardinality of P _{ L }(S). We call L the length constraint for the observer placement because we consider an observer to be useful for source localization only if it is within distance L from another observer. \(S^{\star }_{k, L}\) can be approximated greedily as in Algorithm 2. The running time of Algorithm 2 is O(n ^{2} k ^{2}), however, as Algorithm 1, this algorithm is highly parallelizable and hence tractable even for large networks.
We will refer to the observer placement produced by Algorithm 2 as HVOBS (L) to emphasize that it is designed for the highvariance case.
Unfortunately also for Algorithm 2 we cannot use a submodularity argument to derive approximation guarantees. In fact, the function P _{ L } is not submodular. Consider the path \(\mathcal {P}\) of 7 nodes in Fig. 6b, fix L=3 and set \(\mathcal {O}_{1}=\{1\}\). If we add node 7 to \(\mathcal {O}_{1}\) no node lies on a path of length smaller than L=3 among the two observers 1 and 7, hence the gain is 0. Consider now \(\mathcal {O}_{2}=\{1, 4\}\supseteq \mathcal {O}_{1}\). If we add node 7 to \(\mathcal {O}_{2}\), the gain is 3 because node 5,6 and 7, that did not lie on any path of length smaller than L connecting two observers before, now lie on the path connecting 4 and 7, hence P _{ L } is not submodular.
Comparison with Algorithm 1. Note that taking L equal to the maximum weighted distance Δ between two nodes in \(\mathcal {G}\) does not make Algorithm 2 equivalent to Algorithm 1, i.e., we do not obtain LVOBS. To see how the two algorithms could give different results, take a cycle of odd length d with a leaf node ℓ added as a neighbor to an arbitrary node v and assume to start the algorithm with initial set {v}. At the first step, the two algorithms will make the same choice, choosing one of the two nodes that is at distance (d−1)/2 from v. At the second step however, LVOBS will add ℓ (a DRS contains all leaves (Chen et al. 2014)), whereas Algorithm 2 will add a node on the cycle. This observation is key to our results because it explains why Algorithm 2 results in a more uniform (and hence varianceresistant) observer placement with respect to LVOBS. HVOBS operates a tradeoff between the average distance to the observers and the maximization of \(\mathcal {P}_{s}\).
Choice of the L parameter. How could one optimally set L? Needless to say, the optimal L depends on the network topology and on the available budget: Clearly, for a larger budget a smaller L is preferred.
Empirical results
Datasets
Displays statistics for the networks examined
V  E  min(w _{ uv })  avg(w _{ uv })  max(w _{ uv })  Avg Degree  Diameter  Avg Dist  Avg Clust.  

Friends & Families  120  563  4  5.58  7  9.38  6  17.5  0.67 
Facebook Messages  1020  6205  1  2.97  5  12.16  5  6.69  0.09 
California Roads  1259  1801  1  1.71  9  2.86  66  55.3  0.2 

Friend & Families (F & F). This is a dataset containing phone calls, SMS exchanges and bluetooth proximity, among a community living in the proximity of a university campus (Aharony et al. 2011). We select the largest connected component of individuals who took part in the experiment during its whole duration. The edges are weighted, according to the number of phone calls, SMSs, and bluetooth contacts.

Facebooklike Message Exchange (FB) (Opsahl and Panzarasa 2009). As the individuals included in this dataset were living on the same universitycampus, the number of messages exchanged is likely to be a good measure of inperson interaction. We selected links on which at least one message was sent in both directions and individuals that had a contact with at least one other individual.

California Road Network (CR) (California Road Network). In order to obtain a single connected component and remove points that effectively represent the same location, we collapsed the points falling within a distance of 2 km. Moreover we iteratively deleted all leaves. In fact, the roads that cross the state border are not completely tracked in this dataset and terminate with a leaf. Some other leaves might represent remote locations, not necessarily close to the borders, but their influence on the epidemic should anyway be very low.
The diameter of the CR network is very large compared with that of the other two networks.
The edges are weighted according to a rescaled version of the real distance (measured in km).
In all three networks, edges are given (nonunit) integer weights, which is realistic in many applications as the expected transmission delays are known only up to some level of precision. Integer weights do not simplify the localization of the source; in fact, this makes it more difficult to distinguish between vertices. For example, if the edges of the CR network were weighted according to the Euclidean distance between the two endpoints, LVOBS would use only a very small portion of the budget and the comparison with other observer placements would not be meaningful.
Comparison against benchmarks
 ABC (Adaptive Betweenness Centrality): Betweenness Centrality (BC) is a popular method for placing observers for sourcelocalization (see, e.g., (Louni and Subbalakshmi 2014) and (Seo et al. 2012), where it emerges as the best heuristic for observer placement among those tested). It consists of the k nodes having the largest BC, which is defined, for all u∈V aswhere σ _{ x,y } is the number of shortest paths between x and y and σ _{ x,y }(u) is the number of those paths that passes through u. Here we consider an adaptive version of BC (ABC) which iteratively chooses the node that maximizes the betweenness centrality without considering the shortest paths that pass by alreadychosen vertices (Yoshida 2014). ABC, with respect to the basic BC, gives less clustered, and hence more efficient, observer sets.$$\text{BC}(u) = \sum_{x, y \in V, x \neq y} \frac{\sigma_{x,y}(u)}{\sigma_{x,y}} $$
 Coveragerate (Coverage) (Zhang et al. 2016): This approach maximizes the number of nodes that have an observer as a neighbor, i.e.,where N _{ o } denotes the set of neighbors of o. It has been shown to outperform several heuristics with a diffusion model and a sourcelocalization setting that are very similar to ours (Zhang et al. 2016).$$\mathrm{C}(\mathcal{O}) = \cup_{o \in \mathcal{O}} N_{o}/n $$
 kMedian: this is the optimal placement for the closelyrelated problem of maximizing the detectability of a flow (Berry et al. 2006). The kMedian placement is the set of k nodes \(\mathcal {O}\) such that$$\mathcal{O} = \text{argmin}_{\mathcal{O}=k} \sum_{s \in V} (\min_{o \in \mathcal{O}} d(s, o)). $$
Determining the kMedians of a network is NPhard (Kariv and Hakimi 1979), hence we approximate kMedians with a greedy heuristic.
Transmission delays
Unless otherwise specified, we sample the transmission delays X _{ uv } from truncated Gaussian random variables with parameters (w _{ uv },σ w _{ uv },[^{ w } u v/2,^{3w } u v/2]). More precisely, if \(Y_{uv} \sim \mathcal {N}(w_{uv}, \sigma w_{uv})\) is a Gaussian random variable, X _{ uv } is obtained by conditioning Y _{ uv } with Y _{ uv }∈[^{ w } u v/2,^{3w } u v/2]. With respect to the delay distribution assumed by Pinto et al. (Pinto et al. 2012) i.e., \({X_{uv} \sim \mathcal {N}(w_{uv}, \sigma w_{uv})}\), the distribution we assume has the advantage of admitting only strictly positive infection delays. Furthermore, different values of the parameter σ result in different regimes for the transmission delays, making our model very versatile. When σ=0, we are in the zerovariance regime; when σ is large, the distribution of X _{ uv } becomes closer to a uniform random variable U([^{ w } u v/2,^{3w } u v/2]). Finally, when σ is strictly positive but small, \({X_{uv} \approx \mathcal {N}(w_{uv}, (\sigma w_{uv})^{2})}\).
To assess the robustness of our approach for source localization and observer placement, we also experiment with uniformly distributed transmission delays, i.e., for every edge u v∈E, we take X _{ uv }∼Unif([(1−ε)w _{ uv },(1+ε)w _{ uv }]). The uniform distribution is, among the unimodal distributions on a bounded support, the one that maximizes the variance (Gray and Odell 1967). Hence, uniform delays are a very challenging setting for source localization.
Experimental results
We estimate the probability of success \(\mathcal {P}_{s}\) and the expected distance \(\mathcal {D}\) for different values of the variance parameter σ. Our estimations are computed averaging the results obtained choosing each node in turn as the source and generating synthetic epidemics. For the FB and CR datasets, we run 5 simulations per node and value of σ; for the F & F dataset, as the network is smaller, we run 20 simulations per node and value of σ. For the FB and CR datasets, we localize the source based on the first 20 observations only: Given the large size of these networks, it would be unrealistic to wait for all the nodes to get infected before running the algorithm.
We first take as budget for the observers the minimum budget for which \(\mathcal {P}_{s}(\textsc {lvObs}) = 1\). This corresponds to k∼10% for the F & F dataset, k∼9% for the CR network and k∼5% for the FB dataset. This is the setting in which we expect the improvement of HVOBS over LVOBS to be especially strong: For smaller values of k we expect LVOBS to be nearly optimal even in the highvariance regime because we do not have enough budget to contrast both the topological undistinguishability among nodes (what LVOBS is designed for) and the accumulation of variance (what HVOBS is designed for).
For the F & F and the CR networks, we also experiment with smaller percentages of observers and consistently find an improvement of HVOBS over LVOBS in the highvariance regime: Below a certain amount of variance σ _{1} LVOBS performs better than HVOBS for any choice of the parameter L, whereas above σ _{1} a calibrated choice of L leads to a significant improvement. Such L stays constant for all σ>σ _{1}, i.e., with the notation of Fig. 1 we have σ _{1}=σ _{ F }.
For the FB dataset instead, probably due to the low diameter with respect to the number of nodes, we observe that HVOBS does not improve on LVOBS for any value of L.
Both LVOBS and HVOBS systematically outperform the baseline heuristics for observer placement that we described in “Comparison against benchmarks” section. For the CR dataset the performance of Adaptive Betweenness Centrality is particularly poor. The Coverage Rate heuristic outperforms Adaptive Betweenness Centrality on all three networks (confirming what found by by Zhang et al. (2016)) but is consistently less effective than KMedians and than our methods.
Finally in Fig. 12, we consider uniform transmission delays, and we measure whether, without making any changes, our observer placement still performs well. We find comparable results which suggest that our observer placement is not dependant on the exact transmission model and that the variance of the transmission delays is really a key factor for a good observer placement.
Related work
The problem of source localization has been widely studied in recent years, we survey the works that are more relevant to ours and refer the reader to the survey by Jiang et al. (2014) for a more complete review of the different approaches.
Transmission delays. Many transmission models for epidemics have been studied (Lelarge 2009) and considered for source localization. Although discretetime transmission delays are common (Luo et al. 2014; Prakash et al. 2012; Altarelli et al. 2014), in order to better approximate realistic settings, much work (including ours) adopt continuoustime models with varying distributions for the transmission delays; e.g., exponential (Shah and Zaman 2011; Luo and Tay 2012) or Gaussian (Pinto et al. 2012; Louni and Subbalakshmi 2014; Louni et al. 2015; Zhang et al. 2016). In the same line of the latter class of works, we use truncated Gaussian variables, which gives us the advantage of ensuring that infection delays are strictly positive.
Source localization. Many approaches (Zheng and Tan 2015; Prakash et al. 2012; Sundareisan et al. 2015), beginning with the seminal work by Shah and Zaman Shah and Zaman (2011), rely on knowing the state of the entire network at a fixed point in time t; this is often called a complete observation of the epidemic. These models use maximum likelihood estimation (MLE) to estimate the source. The results of (Shah and Zaman 2011) have been extended in many ways, for example in the case of multiple sources (Luo and Tay 2012) or to obtain a local source estimator (Dong et al. 2013). An alternate line of work considers a complete observation of the epidemic, except that the observed states are noisy, i.e., potentially inaccurate (Zhu and Ying 2013; Sundareisan et al. 2015). As assuming the knowledge of the state of all the nodes is often not realistic, partial observation settings have also been studied. In such a setting, only a subset of nodes \(\mathcal {O}\) reveal their state. In this line of work, the observers are mainly given, either arbitrarily or via a random process, and the problem of selecting observers is not addressed. For example, when a fraction x of nodes are randomly selected, Lokhov et al. (2014) propose an approach which relies on the knowledge of the state (S, I or R) of a fraction of the nodes in the graph at a given moment in time and in which the starting time of the epidemic, if unknown, can be inferred from the data available. When the nodes are independently selected to be observers, an approach to source estimation based on the notion of Jordan center was proposed (Luo et al. 2014) and has since been used for source estimation, especially with regard to a game theoretic version of epidemics (Fanti et al. 2015). This line of work does not assume infection times are known, which we believe is, in many cases, an unnecessary limitation. Indeed by using infection times we can achieve exact source localization in the zerovariance setting with sufficiently many observers (Chen et al. 2014), whereas this is not true otherwise.
Observer placement. Natural heuristics for observer placement (e.g., using highdegree vertices or optimizing for distance centrality) were first evaluated under the additional assumption that infected nodes know which neighbor infected them (Pinto et al. 2012). Later, Louni and Subbalakshmi (2014) proposed, for a similar model, to place the observers using a BetweennessCentrality criterion (which we use as a benchmark, see “Comparison against benchmarks” section), and extended it to noisy observations (Louni et al. 2015). These and other heuristic approaches for observer placement are evaluated empirically by Seo et al. (2012); they reach the conclusion that, among the placements they evaluate, the BetweennessCentrality criterion performs the best. In their work the source is estimated by ranking candidates according to their distance to the set of observers, without using the time at which the observers became infected. Once again, this approach is inherently limited by the fact that it does not make use of the time of infection.
The problem of minimizing the number of observers required to detect the precise source (as opposed to maximizing the performance given a budget of observers) has been considered in the zerovariance setting. For trees, given the time at which the epidemic starts, the minimization problem was solved by Zejnilovic et al. (2013). Without assuming a tree topology and a known starting time, approximation algorithms have been developed towards this end (Chen et al. 2014) (still in a zerovariance setting). However, in a network of size n, the number of observers required, even if minimized, can be up to n−1, hence, a budgeted setting is practically more interesting. For trees, the budgeted placement of observers was solved by using techniques different from ours (Celis et al. 2015). However these techniques heavily rely on the tree structure of the network and do not seem to be extendible to other topologies. In a recent work, Zhang et al. (2016) consider selecting a fixed number of observers using several heuristics such as BetweennessCentrality, DegreeCentrality and ClosenessCentrality and they show that none of these methods are satisfactory. They introduce a new heuristic for the choice of observers, called CoverageRate, which is linked to the total number of nodes neighboring observers, and show that an approximated optimization of this metric yields better performance. Connecting the budgeted placement problem to the unbudgeted minimization problem, we provably outperform their approach in lowvariance settings. For example, in the lowvariance setting, on cycles of oddlength d with budget k=2, any two nodes at distance more than 2 are equivalent with respect to CoverageRate, but they maximize \(\mathcal {P}_{s}\) only if they are at distance (d−1)/2; our approach instead, selects this optimal placement. Moreover, the effect of the variance in the transmission delays is neglected by Zhang et al., leaving open the question of whether their approach works in general. We consider CoverageRate as one of our baselines.
Conclusion and future work
In this work, we have taken a principled approach towards budgeted observer placement for source localization, which shows a dichotomy between the low and highvariance regimes. We developed complementary approaches to handle both regimes. We evaluated our approaches against stateoftheart and alternative heuristics showing a better performance of the algorithms proposed in this paper.
A direction for future work would be to measure the performance with worst case rather than average case metrics: if we can handle (adversarially chosen) source distributions where the epidemic starts at the leastobserved location, then this gives a bound on the performance with an arbitrary prior distribution.
A natural extension of our model was recently studied in a work by Spinelli et al. (2017) which accounts for two stages of observation. In the first stage, as in this work, a small set of observers are selected to monitor the network. In the next stage, once an epidemic begins, additional observers are deployed in the relevant region of the network to localize the source. The latter work does not address interesting questions such as the impact of the initial budget deployed and of the position of the observers chosen in the first stage. The techniques and the results of this paper pave the way for answering these questions which we consider of high practical importance.
Endnotes
^{1} A preliminary version of this work was presented at the 54^{ th } Annual Allerton Conference on Communication, Control, and Computing (Spinelli et al. 2016).
^{2} Note that in Figs. 2 and 5a we compute the value of the success probability \(\mathcal {P}_{s}\) assuming Gaussian distributed delays (and ignoring that, with low probability, negative delays could appear) because this is the only distribution that makes the exact computation of this value feasible. However, in all experiments we only consider nonnegative distributions for X _{ uv }.
^{3} See “Metrics for source localization” section for a discussion of alternative metrics for source localization.
^{4} Call \(\mathcal {O}_{opt}\) the optimal observer placement for any of the metrics considered and \(\mathcal {L}\) the leaves set. If \(\mathcal {O}_{opt} \nsubseteq \mathcal L\) there would be observer \(o\in \mathcal {O}_{opt}\) equivalent to a leaf \(\ell \notin \mathcal {O}_{opt}\) and by substituting o with ℓ we would break [ o] in two or more smaller equivalence classes. In this way the value of the metric considered would get closer to its optimum.
^{5} The standard error of measurement is not reported for the sake of readability but it was checked to be small.
^{6} Lyapunov condition with δ=1 is easily verified for a sequence of independent and uniformly bounded random variables (see Example 27.4 in (Billingsley 1995) for more details).
^{7} https://github.com/bmspinelli/observers_for_source_loc.
Appendix A: Double Resolving Sets
The problem of minimizing the required number of observers in order to perfectly identify the source in the zerovariance setting has been studied (Chen et al. 2014); an observer set \(\mathcal {O}\) such that \(\mathcal {P}_{s}(\mathcal {O}) = 1\) is called a Double Resolving Set (DRS). While the original formulation of the DRS problem is slightly different, this version follows straightforwardly from our observations in “The lowvariance regime” section.
Definition 3
(Double Resolving Set) Given a network \(\mathcal {G}\), S⊆V is said to be a Double Resolving Set of \(\mathcal {G}\) if for any x,y∈V there exist u,v∈S s.t. d(x,u)−d(x,v)≠d(y,u)−d(y,v).
Finding a Double Resolving Set of minimum size is known to be NPhard (Kratica et al. 2009). An approximation algorithm, based on a greedy minimization of an entropy function, has been studied. Note that this has no connection to true informationtheoretic entropy.
Definition 4
Note that \(H_{\mathcal {O}}\) is minimized if and only if each equivalence class consists of only one node and hence if and only if \(\mathcal {P}_{s}=1\). However, despite the fact that \(H_{\mathcal {O}}\) is minimized when \(\mathcal {P}_{s}\) is maximized and that both act on the same set of equivalence classes for a given \(\mathcal {O}\), the greedy processes that minimize \(H_{\mathcal {O}}\) and maximize \(\mathcal {P}_{s}\) are not the same. This can be seen by rewriting both objective functions in the following way. Let [c _{1},…,c _{ q }] be the sequence of equivalence class sizes. Then \(H_{\mathcal {O}}\) can be written as \(H_{\mathcal {O}}([c_{1},.., c_{q}]) = \sum _{i=1}^{l} \sum _{j=2}^{c_{i}} \log (j) = \sum _{i=2}^{\max {c_{j}}} \log (i) \#\{c_{j} \geq i\}.\) Analogously we have the following equality for the success probability \(\mathcal {P}_{s}([c_{1}, \ldots, c_{q}])\): \(n (1\mathcal {P}_{s}([c_{1}, \ldots, c_{q}])) = n  q = \sum _{i=2}^{\max {c_{j}}} \#\{c_{j} \geq i\}\)
Hence, though similar in spirit, a greedy minimization of \(H_{\mathcal {O}}\) is not related to a greedy optimization of \(\mathcal P_{s}\) (or \(\mathbf {E}[d(s^{*}, \hat s)]\)).
Appendix B: Hardness of Budgeted Observer Placement
Theorem 3
Given a network \(\mathcal {G}=(V,E)\) and a budget k, finding an observer set \(\mathcal {O}\) which maximizes \(\mathcal {P}_{s}\) is NPhard.
Proof
We will prove that the budgeted observer placement is NPhard with a reduction from the DRS problem (see Appendix A: Double Resolving Sets section), i.e., given a polynomialtime algorithm for the budgeted observer placement problem, we will prove that we can solve the DRS problem in polynomial time.
Assume that we have a polynomialtime algorithm that takes as input a network \(\mathcal {G} = (V,E)\) and a budget k, and outputs a set \(\mathcal {O} \subseteq V\) of size k such that \(\mathcal {P}_{s}\) is maximized. Recall from “The lowvariance regime” section that given a network \(\mathcal {G}\) and a set \(\mathcal {O}\), the probability \(\mathcal {P}_{s}\) can be calculated in time O(n) where n=V (it is enough to compute the n distances vector with respect to \(\mathcal {O}\) and any reference observer \(o_{1}\in \mathcal {O}\)). Hence, we will construct an algorithm for the DRS problem.
Since the full set V always resolves the network, the program is well defined (i.e., it always returns some k). Moreover, it returns precisely the minimum budget k required in order to attain \(\mathcal {P}_{s} = 1\). Lastly, it is clear that the runtime is at most O(n(p(n)+n)) where p(n) is the running time of algorithm. Hence, we have a polynomialtime algorithm for the DRS problem. □
Appendix C: Alternative objective functions for Algorithm 1
We present the results of the experiment described in Comparison with benchmarks section. Let us here denote LVOBS with Φ for consistency of notation.
Comparison of LVOBS (Φ) with the greedy algorithms that minimize the entropy function of (Chen et al. 2014) (Φ _{ ent }) and the expected distance (Φ _{ dist })
\(\rho (\mathcal {P}_{s}, \Phi, \Phi _{dist})\)  \(\rho (\mathcal {D}, \Phi _{dist}, \Phi)\)  \(\rho ({\mathcal P_{s}}, \Phi, \Phi _{ent})\)  

Random Geometric Network, N=100, r=0.2  
k=2  0.205  0.101  0.033 
k=4  0.014  0.003  0.007 
k=8  0.003  0.002  0.003 
Barabàsi Albert Network, N=100, m=3  
k=2  0.168  0.023  0.037 
k=4  0.039  0.025  0.028 
k=8  0.004  0.003  0.005 
Appendix D: Source Localization in the HighVariance Regime
has a tractable closed form (Pinto et al. 2012). Note that the model of (Pinto et al. 2012) additionally assumed infected observers knew the neighbor that infected them; this assumption is not essential for the derivation of the ML estimator and it is not required in our work.
In this work, we adopt (10) as the source estimator in the noisy case. In fact, even if our edge delays are not Gaussiandistributed, under the hypothesis of sparse observations, we can apply the Central Limit Theorem (CLT) to approximate the sum of the edge delays with Gaussian random variables: if all edges have the same weight we can apply the CLT for i.i.d. random variables; if this is not the case, we can apply Lyapunov’s version of CLT.^{6}Using (10) to compute the ML estimator, the likelihood of nodes in the same equivalence class can result to be different as an artefact of the BFStree approximation. Hence, for consistency with our sourcelocalization method in the lowvariance case, we compute an average likelihood and estimate that the source is in the class with the higher average likelihood. Then, once an equivalence class for the source is estimated, we select \(\hat {s}\) by sampling the prior probability on the position of the source (if available) or by uniform sampling from the estimated equivalence class.
Notes
Acknowledgements
The authors wish to thank the anonymous reviewers for their helpful comments and suggestions.
B. Spinelli was partially supported by the Bill & Melinda Gates Foundation, under Grant No. OPP1070273.
Availability of data and materials
The datasets and the code used for all the experiments presented in this section are publicly available on GitHub.^{7}
Authors’ contributions
All authors participated in the conception and design of the work. BS prepared the experimental setup, analysed the experimental results and drafted the manuscript. All authors participated in a critical revision of the manuscript and approved it in its final form. All authors read and approved the final manuscript.
Competing interests
The authors declare that they have no competing interests.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
 Aharony, N, Pan W, Ip C, Khayal I, Pentland A (2011) Social fMRI: Investigating and shaping social mechanisms in the real world. Pervasive Mob Comput 7(6).Google Scholar
 Altarelli, F, Braunstein A, Dall’Asta L, LageCastellanos A, Zecchina R (2014) Bayesian inference of epidemics on networks via belief propagation. Phys Rev Lett 112(11).Google Scholar
 Berry, J, Hart WE, Phillips CE, Uber JG, Watson J (2006) Sensor placement in municipal water networks with temporal integer programming models. J Water Resour Plan Manag 132(4).Google Scholar
 Billingsley, P (1995) Probability and Measure. John Wiley & Sons, New York.zbMATHGoogle Scholar
 Cáceres, J, Hernando MC, Mora M, Pelayo IM, Puertas ML, Seara C, Wood DR (2007) On the metric dimension of cartesian products of graphs. SIAM J Discret Math 21(2).Google Scholar
 California Road Network. http://www.census.gov/geography.html.
 Celis, LE, Pavetić F, Spinelli B, Thiran P (2015) Budgeted sensor placement for source localization on trees In: LatinAmerican Algorithms, Graphs and Optimization Symposium (LAGOS).Google Scholar
 Cha, M, Mislove A, Gummadi KP (2009) A measurementdriven analysis of information propagation in the flickr social network In: Int Conf World Wide Web (WWW).Google Scholar
 Chen, X, Hu X, Wang C (2014) Approximability of the minimum weighted doubly resolving set problem In: Int Computing and Combinatorics Conf (COCOON).Google Scholar
 Dong, W, Zhang W, Tan CW (2013) Rooting out the rumor culprit from suspects In: IEEE Int. Symposium on Information Theory (ISIT).Google Scholar
 Fanti, GC, Kairouz P, Oh S, Viswanath P (2015) Spy vs. spy: Rumor source obfuscation In: SIGMETRICS Conf.Google Scholar
 Gray, HL, Odell PL (1967) On least favorable density functions. SlAM Rev9.Google Scholar
 Jiang, J, Wen S, Yu S, Xiang Y, Zhou W (2014) Identifying propagation sources in networks: Stateoftheart and comparative studies. IEEE Commun Surv Tutorials.Google Scholar
 Kariv, O, Hakimi SL (1979) An algorithmic approach to network location problems. II: The pmedians. SIAM J Appl Math37.Google Scholar
 Kratica, J, Čangalović M, KovačevićVujčić V (2009) Computing minimal doubly resolving sets of graphs. Comput Oper Res 36(7).Google Scholar
 Lelarge, M (2009) Efficient control of epidemics over random networks In: SIGMETRICS/Performance Conf.Google Scholar
 Lessler, J, Reich NG, Brookmeyer R, Perl TM, Nelson KE, Cummings DAT (2009) Incubation periods of acute respiratory viral infections: a systematic review. Lancet Infect Dis 9(5).Google Scholar
 Luo, W, Tay WP (2012) Identifying infection sources in large tree networks In: IEEE Conf on Sensor, Mesh and Ad Hoc Communications and Networks (SECON).Google Scholar
 Luo, W, Tay W, Leng M (2014) How to identify an infection source with limited observations. IEEE J Sel Top Signal Process 8(4).Google Scholar
 Lokhov, AY, Mézard M, Ohta H, Zdeborová L (2014) Inferring the origin of an epidemic with a dynamic messagepassing algorithm. Phys Rev E 90(1).Google Scholar
 Louni, A, Santhanakrishnan A, Subbalakshmi KP (2015) Identification of source of rumors in social networks with incomplete information In: Int Conf on Social Computing (SocialCom).Google Scholar
 Louni, A, Subbalakshmi KP (2014) A twostage algorithm to estimate the source of information diffusion in social media networks In: IEEE INFOCOM Workshop on Dynamic Social Networks.Google Scholar
 Netrapalli, P, Sanghavi S (2012) Learning the graph of epidemic cascades. SIGMETRICS Perform Eval Rev 40(1).Google Scholar
 Opsahl, T, Panzarasa P (2009) Clustering in weighted networks. Soc Networks 31(2).Google Scholar
 Pinto, P, Thiran P, Vetterli M (2012) Locating the source of diffusion in largescale networks. Phys Rev Lett109.Google Scholar
 Prakash, BA, Vreeken J, Faloutsos C (2012) Spotting culprits in epidemics: How many and which ones? In: IEEE Int Conf on Data Mining (ICDM).Google Scholar
 Seo, E, Mohapatra P, Abdelzaher T (2012) Identifying rumors and their sources in social networks In: SPIE Defense, Security, and Sensing.Google Scholar
 Shah, D, Zaman T (2011) Rumors in a network: Who’s the culprit?IEEE Trans Inf Theory57.Google Scholar
 Spinelli, B, Celis LE, Thiran P (2016) Observer placement for source localization: the effect of budgets and transmission variance In: Allerton Conf on Communication, Control & Computing.Google Scholar
 Spinelli, B, Celis LE, Thiran P (2017) Back to the source: An online approach for sensor placement and source localization In: Int Conf World Wide Web (WWW). https://arxiv.org/abs/1702.01056.
 Sundareisan, S, Vreeken J, Prakash BA (2015) Hidden hazards: Finding missing nodes in large graph epidemics In: Int Conf Data Mining (SDM).. SIAM.Google Scholar
 Vergu, E, Busson H, Ezanno P (2010) Impact of the infection period distribution on the epidemic spread in a metapopulation model. PLoS ONE 5(2).Google Scholar
 Yoshida, Y (2014) Almost lineartime algorithms for adaptive betweenness centrality using hypergraph sketches In: ACM SIGKDD Int Conf on Knowledge discovery and data mining (KDD).Google Scholar
 Zejnilovic, S, Gomes JP, Sinopoli B (2013) Network observability and localization of the source of diffusion based on a subset of vertices In: Allerton Conf. on Communication, Control & Computing.Google Scholar
 Zhang, X, Zhang Y, Lv T, Yin Y (2016) Identification of efficient observers for locating spreading source in complex networks. Physica A Stat Mech Appl 442: 100–109.CrossRefGoogle Scholar
 Zhang, Z, Xu W, Wu W, Du DZ (2015) A novel approach for detecting multiple rumor sources in networks with partial observations. J Comb Optim.Google Scholar
 Zheng, L, Tan CW (2015) A probabilistic characterization of the rumor graph boundary in rumor source detection In: IEEE Int Conf on Digital Signal Processing (DSP).Google Scholar
 Zhu, K, Ying L (2013) Information source detection in the SIR model: A sample path based approach In: Information Theory and Applications Workshop (ITA).Google Scholar
Copyright information
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License(http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.