<p>Meta-analysis is a method of synthesizing results of different studies conducted to answer a specific question. Meta-analysis applications have been published in a wide range of disciplines including medicine, education, psychology and many others. However, for many years, only pair-wise and direct comparisons have been made using standard meta-analysis methods. It is only recently that network meta-analysis emerged enabling the comparison of multiple treatments based on estimates from different studies. With network meta-analysis, the relative efficacy (or safety) of a particular intervention versus competing interventions can be obtained even in the absence of head-to-head evidence via a common comparator.</p> <p>An increasing number of methodologies related to network meta-analysis, assessments of underlying assumptions, and strategies for presentation of results have been proposed by several researchers. But only few simulation studies have been done to investigate different characteristics of this emerging statistical method. Hierarchical Bayesian meta-analysis model is commonly used in network meta-analysis to estimate effect of each intervention relative to every other. This model facilitates the calculation of the rank probabilities of a set of alternative treatments. However, various factors can determine the performance of the model which needs to be considered before using results for decision.</p> <p>This project aimed to investigate how the Bayesian hierarchical model estimates the rank probability of the best overall most effective treatment (i.e., the treatment ranked first) under different scenarios for modelling a binary outcome. Different network geometries, numbers of studies per comparison, sets of probabilities of success for treatments and sample sizes were investigated in our simulation study for binary outcome.</p> <p>Our simulation study showed that the estimates of treatments under consideration can be affected by network structures. Similar geometries affect the estimate in similar ways. Unbalanced number of studies per comparison influenced estimates of treatments in the geometries we considered. When a superior treatment is involved in the network, the hierarchical Bayesian mixed treatment model correctly identified it regardless of network patterns, number of studies and individual study sample size.</p> / Master of Science (MSc)

Markov clustering (MCL) is an effective unsupervised pattern recognition algorithm for data clustering in high-dimensional feature space that simulates stochastic flows on a network of sample similarities to detect the structural organization of clusters in the data. However, it presents two main drawbacks: (1) its community detection performance in complex networks has been demonstrating results far from the state-of-the-art methods such as Infomap and Louvain, and (2) it has never been generalized to deal with data nonlinearity. In this work both aspects, although closely related, are taken as separated issues and addressed as such. Regarding the community detection, field under the network science ceiling, the crucial issue is to convert the unweighted network topology into a ‘smart enough’ pre-weighted connectivity that adequately steers the stochastic flow procedure behind Markov clustering. Here a conceptual innovation is introduced and discussed focusing on how to leverage network latent geometry notions in order to design similarity measures for pre-weighting the adjacency matrix used in Markov clustering community detection. The results demonstrate that the proposed strategy improves Markov clustering significantly, to the extent that it is often close to the performance of current state-of-the-art methods for community detection. These findings emerge considering both synthetic ‘realistic’ networks (with known ground-truth communities) and real networks (with community metadata), even when the real network connectivity is corrupted by noise artificially induced by missing or spurious links. Regarding the nonlinearity aspect, the development of algorithms for unsupervised pattern recognition by nonlinear clustering is a notable problem in data science. Minimum Curvilinearity (MC) is a principle that approximates nonlinear sample distances in the high-dimensional feature space by curvilinear distances, which are computed as transversal paths over their minimum spanning tree, and then stored in a kernel. Here, a nonlinear MCL algorithm termed MC-MCL is proposed, which is the first nonlinear kernel extension of MCL and exploits Minimum Curvilinearity to enhance the performance of MCL in real and synthetic high-dimensional data with underlying nonlinear patterns. Furthermore, improvements in the design of the so-called MC-kernel by applying base modifications to better approximate the data hidden geometry have been evaluated with positive outcomes. Thus, different nonlinear MCL versions are compared with baseline and state-of-art clustering methods, including DBSCAN, K-means, affinity propagation, density peaks, and deep-clustering. As result, the design of a suitable nonlinear kernel provides a valuable framework to estimate nonlinear distances when its kernel is applied in combination with MCL. Indeed, nonlinear-MCL variants overcome classical MCL and even state-of-art clustering algorithms in different nonlinear datasets. This dissertation discusses the enhancements and the generalized understanding of how network geometry plays a fundamental role in designing algorithms based on network navigability.

Graphs are a great tool for representing complex physical and social systems, where the interactions among many units, from tens of animal species in a food-web, to millions of users in a social network, give rise to emergent, complex system behaviours. In the field of network science this representation, which is usually called a complex network, can be complicated at will to better represent the real system under study. For instance, interactions may be directed or may differ in their strength or cost, leading to directed weighted networks, but they may also depend on time, like in temporal networks, or nodes (i.e. the units of the system) may interact in different ways, in which case edge-coloured multi-graphs and multi-layer networks represent better the system. Besides this rich repertoire of network structures, we cannot forgot that edges represent interactions and that this interactions are not static, but are, instead, purposely established to reach some function of the system, as for instance, routing people and goods through a transportation network or cognition, through the exchange of neuro-physiological signals in the brain. Building on the foundations of spectral graph theory, of non-linear dimensionality reduction and diffusion maps, and of the recently introduced diffusion distance [Phys. Rev. Lett. 118, 168301 (2017)] we use the simple yet powerful tool of continuous-time Markov chains on networks to model their process-driven geometry and characterise their functional shape. The main results are: (i) the generalisation of the diffusion geometry framework to different types of interconnected systems (from edge-coloured multigraphs to multi-layer networks) and of random walk dynamics [Phys. Rev. E 103, 042301 (2021)] and (ii) the introduction of new descriptors based on the diffusion geometry to quantify and describe the micro- (through the network depth [J. Complex Netw. 8, 4 (2020)]), meso- (functional rich-club) and macro-scale (using statistics of the pairwise distances between the network's nodes [Comm. Phys. 4, 125 (2021)]) of complex networks.

The investigation of the latent geometrical space behind complex network topologies is a fervid topic in current network science and the hyperbolic space is one of the most studied, because it seems associated to the structural organization of many real complex systems. The popularity-similarity-optimization (PSO) generative model is able to grow random geometric graphs in the hyperbolic space with realistic properties such as clustering, small-worldness, scale-freeness and rich-clubness. However, it misses to reproduce an important feature of real complex systems, which is the community organization. Here, we introduce the nonuniform PSO (nPSO) generative model, a generalization of the PSO model with a tailored community structure, and we provide an efficient algorithmic implementation with a O(EN) time complexity, where N is the number of nodes and E the number of edges. Meanwhile, in recent years, the inverse problem has also gained increasing attention: given a network topology, how to provide an accurate mapping into its latent geometrical space. Unlike previous attempts based on a computationally expensive maximum likelihood optimization (whose time complexity is between O(N^3) and O(N^4)), here we show that a class of methods based on nonlinear dimensionality reduction can solve the problem with higher precision and reducing the time complexity to O(N^2).