These datasets naturally come in two different types. First, from the time evolution of some financial indicator or the irregular motion of turbulent fluids to the waveform signal of speech, complex systems produce incredibly complicated univariate/multivariate time series, whose hidden structure should be processed and analysed using fast and novel approaches. Second, the intertwined architecture of the interaction patterns of complex systems is naturally represented and modeled in terms of graphs -a paradigmatic of this approach being the brain, modeled by single units (neurons) connected by edges that model synaptic connections. These distributed processing systems usually lay at the edge between order and randomness (the so-called complex network paradigm) and come in different flavours (undirected/directed, static/temporal, monolayer/multilayer). Each of these two families of datasets have its own mathematical corpus that deals with the description and characterisation of these data, namely signal processing and network science.

The working hypothesis of this project is that information encoded or hidden in a data set can be retrieved by mapping such data set into an alternative mathematical representation, where the extraction of information may be eventually simpler. As such, we aim to explore what new information can be extracted by mapping time series into graphs and therefore using network science to characterise signals and their underlying dynamics: in short, to make graph-theoretical time series analysis. We are also interested in the dual problem, namely extracting time series from graphs and therefore using the tools of time series analysis and signal processing to describe, compare and classify networks of many kinds: a signal processing of graphs.

We will consider specific methods (visibility algorithms, Markov chain theory, fluctuation analysis) and will be able to define and validate new graph-theoretical measures to describe signals and new signal-theoretic measures to describe graphs, as well as to build a mathematically sound and solid theory to relate these two approaches.

Ultimately, the results of our research will be implemented in a software whose input is a time series/complex network and whose output is a set of key features which describe the object under study from several angles (both the signal processing and graph theoretic angle). These features will then feed automatic classifiers for pattern recognition and data analytics.

Submitted for publication

Submitted for publication

Physical Review E 96, 012318 (2017)

Network Neuroscience (in press 2017)

Submitted for publication

Nature Scientific Reports 7, 43862 (2017)

Back to Home

Previous page