@article{HakeBodenbergerGroth2023, author = {Hake, Tim and Bodenberger, Bernhard and Groth, Detlef}, title = {In Python available: St. Nicolas House Algorithm (SNHA) with bootstrap support for improved performance in dense networks}, series = {Human biology and public health}, volume = {1}, journal = {Human biology and public health}, publisher = {Universit{\"a}tsverlag Potsdam}, address = {Potsdam}, issn = {2748-9957}, doi = {10.52905/hbph2023.1.63}, pages = {16}, year = {2023}, abstract = {The St. Nicolas House Algorithm (SNHA) finds association chains of direct dependent variables in a data set. The dependency is based on the correlation coefficient, which is visualized as an undirected graph. The network prediction is improved by a bootstrap routine. It enables the computation of the empirical p-value, which is used to evaluate the significance of the predicted edges. Synthetic data generated with the Monte Carlo method were used to firstly compare the Python package with the original R package, and secondly to evaluate the predicted network using the sensitivity, specificity, balanced classification rate and the Matthew's correlation coefficient (MCC). The Python implementation yields the same results as the R package. Hence, the algorithm was correctly ported into Python. The SNHA scores high specificity values for all tested graphs. For graphs with high edge densities, the other evaluation metrics decrease due to lower sensitivity, which could be partially improved by using bootstrap,while for graphs with low edge densities the algorithm achieves high evaluation scores. The empirical p-values indicated that the predicted edges indeed are significant.}, language = {en} } @article{NovineMattssonGroth2022, author = {Novine, Masiar and Mattsson, Cecilie Cordua and Groth, Detlef}, title = {Network reconstruction based on synthetic data generated by a Monte Carlo approach}, series = {Human biology and public health}, volume = {2021}, journal = {Human biology and public health}, number = {3, Summer School Supplement}, publisher = {Universit{\"a}tsverlag Potsdam}, address = {Potsdam}, issn = {2748-9957}, doi = {10.52905/hbph2021.3.26}, pages = {23}, year = {2022}, abstract = {Background: Network models are useful tools for researchers to simplify and understand investigated systems. Yet, the assessment of methods for network construction is often uncertain. Random resampling simulations can aid to assess methods, provided synthetic data exists for reliable network construction. Objectives: We implemented a new Monte Carlo algorithm to create simulated data for network reconstruction, tested the influence of adjusted parameters and used simulations to select a method for network model estimation based on real-world data. We hypothesized, that reconstructs based on Monte Carlo data are scored at least as good compared to a benchmark. Methods: Simulated data was generated in R using the Monte Carlo algorithm of the mcgraph package. Benchmark data was created by the huge package. Networks were reconstructed using six estimator functions and scored by four classification metrics. For compatibility tests of mean score differences, Welch's t-test was used. Network model estimation based on real-world data was done by stepwise selection. Samples: Simulated data was generated based on 640 input graphs of various types and sizes. The real-world dataset consisted of 67 medieval skeletons of females and males from the region of Refshale (Lolland) and Nordby (Jutland) in Denmark. Results: Results after t-tests and determining confidence intervals (CI95\%) show, that evaluation scores for network reconstructs based on the mcgraph package were at least as good compared to the benchmark huge. The results even indicate slightly better scores on average for the mcgraph package. Conclusion: The results confirmed our objective and suggested that Monte Carlo data can keep up with the benchmark in the applied test framework. The algorithm offers the feature to use (weighted) un- and directed graphs and might be useful for assessing methods for network construction.}, language = {en} } @phdthesis{Cecchini2019, author = {Cecchini, Gloria}, title = {Improving network inference by overcoming statistical limitations}, doi = {10.25932/publishup-42670}, url = {http://nbn-resolving.de/urn:nbn:de:kobv:517-opus4-426705}, school = {Universit{\"a}t Potsdam}, pages = {124}, year = {2019}, abstract = {A reliable inference of networks from data is of key interest in many scientific fields. Several methods have been suggested in the literature to reliably determine links in a network. These techniques rely on statistical methods, typically controlling the number of false positive links, but not considering false negative links. In this thesis new methodologies to improve network inference are suggested. Initial analyses demonstrate the impact of falsepositive and false negative conclusions about the presence or absence of links on the resulting inferred network. Consequently, revealing the importance of making well-considered choices leads to suggest new approaches to enhance existing network reconstruction methods. A simulation study, presented in Chapter 3, shows that different values to balance false positive and false negative conclusions about links should be used in order to reliably estimate network characteristics. The existence of type I and type II errors in the reconstructed network, also called biased network, is accepted. Consequently, an analytic method that describes the influence of these two errors on the network structure is explored. As a result of this analysis, an analytic formula of the density of the biased vertex degree distribution is found (Chapter 4). In the inverse problem, the vertex degree distribution of the true underlying network is analytically reconstructed, assuming the probabilities of type I and type II errors. Chapters 4-5 show that the method is robust to incorrect estimates of α and β within reasonable limits. In Chapter 6, an iterative procedure to enhance this method is presented in the case of large errors on the estimates of α and β. The investigations presented so far focus on the influence of false positive and false negative links on the network characteristics. In Chapter 7, the analysis is reversed - the study focuses on the influence of network characteristics on the probability of type I and type II errors, in the case of networks of coupled oscillators. The probabilities of α and β are influenced by the shortest path length and the detour degree, respectively. These results have been used to improve the network reconstruction, when the true underlying network is not known a priori, introducing a novel and advanced concept of threshold.}, language = {en} } @article{KralemannPikovskijRosenblum2014, author = {Kralemann, Bjoern and Pikovskij, Arkadij and Rosenblum, Michael}, title = {Reconstructing effective phase connectivity of oscillator networks from observations}, series = {New journal of physics : the open-access journal for physics}, volume = {16}, journal = {New journal of physics : the open-access journal for physics}, publisher = {IOP Publ. Ltd.}, address = {Bristol}, issn = {1367-2630}, doi = {10.1088/1367-2630/16/8/085013}, pages = {21}, year = {2014}, abstract = {We present a novel approach for recovery of the directional connectivity of a small oscillator network by means of the phase dynamics reconstruction from multivariate time series data. The main idea is to use a triplet analysis instead of the traditional pairwise one. Our technique reveals an effective phase connectivity which is generally not equivalent to a structural one. We demonstrate that by comparing the coupling functions from all possible triplets of oscillators, we are able to achieve in the reconstruction a good separation between existing and non-existing connections, and thus reliably reproduce the network structure.}, language = {en} }