Santiago Segarra

Sampling of Graph Signals With Successive Local Aggregations

A new scheme to sample signals defined on the nodes of a graph is proposed. The underlying assumption is that such signals admit a sparse representation in a frequency domain related to the structure of the graph, which is captured by the so-called graph-shift operator. Instead of using the value of the signal observed at a subset of nodes to recover the signal in the entire graph, the sampling scheme proposed here uses as input observations taken at a single node. The observations correspond to sequential applications of the graph-shift operator, which are linear combinations of the information gathered by the neighbors of the node. When the graph corresponds to a directed cycle (which is the support of time-varying signals), our method is equivalent to the classical sampling in the time domain. When the graph is more general, we show that the Vandermonde structure of the sampling matrix, critical when sampling time-varying signals, is preserved. Sampling and interpolation are analyzed first in the absence of noise, and then noise is considered. We then study the recovery of the sampled signal when the specific set of frequencies that is active is not known. Moreover, we present a more general sampling scheme, under which, either our aggregation approach or the alternative approach of sampling a graph signal by observing the value of the signal at a subset of nodes can be both viewed as particular cases. Numerical experiments illustrating the results in both synthetic and real-world graphs close the paper.

Stationary Graph Processes and Spectral Estimation

Stationarity is a cornerstone property that facilitates the analysis and processing of random signals in the time domain. Although time-varying signals are abundant in nature, in many practical scenarios, the information of interest resides in more irregular graph domains. This lack of regularity hampers the generalization of the classical notion of stationarity to graph signals. This paper proposes a definition of weak stationarity for random graph signals that takes into account the structure of the graph where the random process takes place, while inheriting many of the meaningful properties of the classical time domain definition. Provided that the topology of the graph can be described by a normal matrix, stationary graph processes can be modeled as the output of a linear graph filter applied to a white input. This is shown equivalent to requiring the correlation matrix to be diagonalized by the graph Fourier transform; a fact that is leveraged to define a notion of power spectral density (PSD). Properties of the graph PSD are analyzed and a number of methods for its estimation are proposed. This includes generalizations of nonparametric approaches such as periodograms, window-based average periodograms, and filter banks, as well as parametric approaches, using moving-average, autoregressive, and ARMA processes. Graph stationarity and graph PSD estimation are investigated numerically for synthetic and real-world graph signals.

Reconstruction of Graph Signals Through Percolation from Seeding Nodes

New schemes to recover signals defined in the nodes of a graph are proposed. Our focus is on reconstructing bandlimited graph signals, which are signals that admit a sparse representation in a frequency domain related to the structure of the graph. Most existing formulations focus on estimating an unknown graph signal by observing its value on a subset of nodes. By contrast, in this paper, we study the problem of inducing a known graph signal using as input a graph signal that is nonzero only for a small subset of nodes. The sparse signal is then percolated (interpolated) across the graph using a graph filter. Alternatively, one can interpret graph signals as network states and study graph-signal reconstruction as a network-control problem where the target class of states is represented by bandlimited signals. Three setups are investigated. In the first one, a single simultaneous injection takes place on several nodes in the graph. In the second one, successive value injections take place on a single node. The third one is a generalization where multiple nodes inject multiple signal values. For noiseless settings, conditions under which perfect reconstruction is feasible are given, and the corresponding schemes to recover the desired signal are specified. Scenarios leading to imperfect reconstruction, either due to insufficient or noisy signal value injections, are also analyzed. Moreover, connections with classical interpolation in the time domain are discussed. Specifically, for time-varying signals, where the ideal interpolator after uniform sampling is a (low-pass) filter, our proposed approach and the reconstruction of a sampled signal coincide. Nevertheless, for general graph signals, we show that these two approaches differ. The last part of the paper presents numerical experiments that illustrate the results developed through synthetic and real-world signal reconstruction problems.

Axiomatic construction of hierarchical clustering in asymmetric networks

We present an axiomatic construction of hierarchical clustering in asymmetric networks where the dissimilarity from node a to node b is not necessarily equal to the dissimilarity from node b to node a. The theory is built on the axioms of value and transformation which encode desirable properties common to any clustering method. Two hierarchical clustering methods that abide to these axioms are derived: reciprocal and nonreciprocal clustering. We further show that any clustering method that satisfies the axioms of value and transformation lies between reciprocal and nonreciprocal clustering in a well defined sense. We apply this theory to the formation of circles of trust in social networks.

Distributed implementation of linear network operators using graph filters

A signal in a network (graph) can be defined as a vector whose elements represent the value of a given magnitude at the different nodes. A linear network (graph) operator is then a linear transformation whose input and output are graph signals. This paper investigates how to use graph filters to implement generic network linear operators in a distributed manner, so that nodes only need to exchange a finite number of messages with their neighbors. The schemes are designed within the framework of shift-invariant graph filters, which are polynomials of the so-called graph-shift operator. The graph-shift operator is a matrix that accounts for the topology of the network, and the filter coefficients - which are the same across nodes - are the coefficients of that polynomial. First, we identify conditions under which the linear operator can be computed exactly. Then, we provide approximate designs for cases where perfect implementation is unfeasible. Setups where each node is allowed to use a different set of filter coefficients are briefly discussed. Finally, we apply this framework to the problem of finite-time consensus and analyze the graph-filter approximation performance for general linear graph operators.

Authorship Attribution Through Function Word Adjacency Networks

A method for authorship attribution based on function word adjacency networks (WANs) is introduced. Function words are parts of speech that express grammatical relationships between other words but do not carry lexical meaning on their own. In the WANs in this paper, nodes are function words and directed edges from a source function word to a target function word stand in for the likelihood of finding the latter in the ordered vicinity of the former. WANs of different authors can be interpreted as transition probabilities of a Markov chain and are therefore compared in terms of their relative entropies. Optimal selection of WAN parameters is studied and attribution accuracy is benchmarked across a diverse pool of authors and varying text lengths. This analysis shows that, since function words are independent of content, their use tends to be specific to an author and that the relational data captured by function WANs is a good summary of stylometric fingerprints. Attribution accuracy is observed to exceed the one achieved by methods that rely on word frequencies alone. Further combining WANs with methods that rely on word frequencies, results in larger attribution accuracy, indicating that both sources of information encode different aspects of authorial styles.

Interpolation of graph signals using shift-invariant graph filters

New schemes to recover signals defined in the nodes of a graph are proposed. Our focus is on reconstructing bandlimited graph signals, which are signals that admit a sparse representation in a frequency domain related to the structure of the graph. The schemes are designed within the framework of linear shift-invariant graph filters and consider that the seeding signals are injected only at a subset of interpolating nodes. After several sequential applications of the graph-shift operator - which computes linear combinations of the information available at neighboring nodes - the seeding signals are diffused across the graph and the original bandlimited signal is eventually recovered. Conditions under which the recovery is feasible are given, and the corresponding schemes to recover the signal are proposed. Connections with the classical interpolation in the time domain are also discussed.

Network Topology Inference from Spectral Templates

We address the problem of identifying the structure of an undirected graph from the observation of signals defined on its nodes. Fundamentally, the unknown graph encodes direct relationships between signal elements, which we aim to recover from observable indirect relationships generated by a diffusion process on the graph. The fresh look advocated here leverages concepts from convex optimization and stationarity of graph signals, in order to identify the graph shift operator (a matrix representation of the graph) given only its eigenvectors. These spectral templates can be obtained, e.g., from the sample covariance of independent graph signals diffused on the sought network. The novel idea is to find a graph shift that, while being consistent with the provided spectral information, endows the network with certain desired properties such as sparsity. To that end, we develop efficient inference algorithms stemming from provably tight convex relaxations of natural nonconvex criteria, particularizing the results for two shifts: the adjacency matrix and the normalized Laplacian. Algorithms and theoretical recovery conditions are developed not only when the templates are perfectly known, but also when the eigenvectors are noisy or when only a subset of them are given. Numerical tests showcase the effectiveness of the proposed algorithms in recovering synthetic and real-world networks.

Diffusion and Superposition Distances for Signals Supported on Networks

We introduce the diffusion and superposition distances as two metrics to compare signals supported in the nodes of a network. Both metrics consider the given vectors as initial temperature distributions and diffuse heat through the edges of the graph. The similarity between the given vectors is determined by the similarity of the respective diffusion profiles. The superposition distance computes the instantaneous difference between the diffused signals and integrates the difference over time. The diffusion distance determines a distance between the integrals of the diffused signals. We prove that both distances define valid metrics and that they are stable to perturbations in the underlying network. We utilize numerical experiments to illustrate their utility in classifying signals in a synthetic network as well as in classifying ovarian cancer histologies using gene mutation profiles of different patients. We also utilize diffusion as part of a label propagation method in semi-supervised learning to classify handwritten digits.

Authorship attribution using function words adjacency networks

We present an authorship attribution method based on relational data between function words. These are content independent words that help define grammatical relationships. As relational structures we use normalized word adjacency networks. We interpret these networks as Markov chains and compare them using entropy measures. We illustrate the accuracy of the method developed through a series of numerical experiments including comparisons with frequency based methods. We show that accuracy increases when combining relational and frequency based data, indicating that both sources of information encode different aspects of authorial styles.