arXiv subjects

Rebekka Burkholz

Publications and source records attributed to Rebekka Burkholz.

47 records · Page 3Linked to original sources

arXiv2021

Scaling up Continuous-Time Markov Chains Helps Resolve Underspecification

Modeling the time evolution of discrete sets of items (e.g., genetic mutations) is a fundamental problem in many biomedical applications. We approach this problem through the lens of continuous-time Markov chains, and show that the resulting learning task is generally underspecified in the usual setting of cross-sectional data. We explore a perhaps surprising remedy: including a number of additional independent items can help determine time order, and hence resolve underspecification. This is in sharp contrast to the common practice of limiting the analysis to a small subset of relevant items, which is followed largely due to poor scaling of existing methods. To put our theoretical insight into practice, we develop an approximate likelihood maximization method for learning continuous-time Markov chains, which can scale to hundreds of items and is orders of magnitude faster than previous methods. We demonstrate the effectiveness of our approach on synthetic and real cancer data.

cs.LG↗

arXiv2020

Cascade Size Distributions: Why They Matter and How to Compute Them Efficiently

Cascade models are central to understanding, predicting, and controlling epidemic spreading and information propagation. Related optimization, including influence maximization, model parameter inference, or the development of vaccination strategies, relies heavily on sampling from a model. This is either inefficient or inaccurate. As alternative, we present an efficient message passing algorithm that computes the probability distribution of the cascade size for the Independent Cascade Model on weighted directed networks and generalizations. Our approach is exact on trees but can be applied to any network topology. It approximates locally tree-like networks well, scales to large networks, and can lead to surprisingly good performance on more dense networks, as we also exemplify on real world data.

cs.AI↗

arXiv2019

Initialization of ReLUs for Dynamical Isometry

Deep learning relies on good initialization schemes and hyperparameter choices prior to training a neural network. Random weight initializations induce random network ensembles, which give rise to the trainability, training speed, and sometimes also generalization ability of an instance. In addition, such ensembles provide theoretical insights into the space of candidate models of which one is selected during training. The results obtained so far rely on mean field approximations that assume infinite layer width and that study average squared signals. We derive the joint signal output distribution exactly, without mean field assumptions, for fully-connected networks with Gaussian weights and biases, and analyze deviations from the mean field results. For rectified linear units, we further discuss limitations of the standard initialization scheme, such as its lack of dynamical isometry, and propose a simple alternative that overcomes these by initial parameter sharing.

stat.ML↗

arXiv2019

International crop trade networks: The impact of shocks and cascades

Analyzing available FAO data from 176 countries over 21 years, we observe an increase of complexity in the international trade of maize, rice, soy, and wheat. A larger number of countries play a role as producers or intermediaries, either for trade or food processing. In consequence, we find that the trade networks become more prone to failure cascades caused by exogenous shocks. In our model, countries compensate for demand deficits by imposing export restrictions. To capture these, we construct higher-order trade dependency networks for the different crops and years. These networks reveal hidden dependencies between countries and allow to discuss policy implications.

physics.soc-ph↗

arXiv2018

Efficient message passing for cascade size distributions on finite trees

How big is the risk that a few initial failures of networked nodes amplify to large cascades that endanger the functioning of the system? Common answers refer to the average final cascade size. Two analytic approaches allow its computation: a) (heterogeneous) mean field approximation and b) belief propagation. The former applies to (infinitely) large locally tree-like networks, while the latter is exact on finite trees. Yet, cascade sizes can have broad and multi-modal distributions that are not well represented by their average. Full distribution information is essential to identify likely events and to estimate the tail risk, i.e. the probability of extreme events. Here, we lay the basis for a general theory to calculate the cascade size distribution in finite networks. We present an efficient message passing algorithm that is exact on finite trees and a large class of cascade processes. An approximation version performs well on locally tree-like networks.

physics.soc-ph↗

arXiv2018

Explicit size distributions of failure cascades redefine systemic risk on finite networks

How big is the risk that a few initial failures of nodes in a network amplify to large cascades that span a substantial share of all nodes? Predicting the final cascade size is critical to ensure the functioning of a system as a whole. Yet, this task is hampered by uncertain or changing parameters and missing information. In infinitely large networks, the average cascade size can often be well estimated by established approaches building on local tree approximations and mean field approximations. Yet, as we demonstrate, in finite networks, this average does not even need to be a likely outcome. Instead, we find broad and even bimodal cascade size distributions. This phenomenon persists for system sizes up to $10^{7}$ and different cascade models, i.e. it is relevant for most real systems. To show this, we derive explicit closed-form solutions for the full probability distribution of the final cascade size. We focus on two topological limit cases, the complete network representing a dense network with a very narrow degree distribution, and the star network representing a sparse network with a inhomogeneous degree distribution. Those topologies are of great interest, as they either minimize or maximize the average cascade size and are common motifs in many real world networks.

physics.soc-ph↗

arXiv2017

Modeling the formation of R\&D alliances: An agent-based model with empirical validation

We develop an agent-based model to reproduce the size distribution of R\&D alliances of firms. Agents are uniformly selected to initiate an alliance and to invite collaboration partners. These decide about acceptance based on an individual threshold that is compared with the utility expected from joining the current alliance. The benefit of alliances results from the fitness of the agents involved. Fitness is obtained from an empirical distribution of agent's activities. The cost of an alliance reflects its coordination effort. Two free parameters $a_{c}$ and $a_{l}$ scale the costs and the individual threshold. If initiators receive $R$ rejections of invitations, the alliance formation stops and another initiator is selected. The three free parameters $(a_{c},a_{l},R)$ are calibrated against a large scale data set of about 15,000 firms engaging in about 15,000 R\&D alliances over 26 years. For the validation of the model we compare the empirical size distribution with the theoretical one, using confidence bands, to find a very good agreement. As an asset of our agent-based model, we provide an analytical solution that allows to reduce the simulation effort considerably. The analytical solution applies to general forms of the utility of alliances. Hence, the model can be extended to other cases of alliance formation. While no information about the initiators of an alliance is available, our results indicate that mostly firms with high fitness are able to attract newcomers and to establish larger alliances.

physics.soc-ph↗

arXiv2017

Correlations between thresholds and degrees: An analytic approach to model attacks and failure cascades

Two node variables determine the evolution of cascades in random networks: a node's degree and threshold. Correlations between both fundamentally change the robustness of a network, yet, they are disregarded in standard analytic methods as local tree or heterogeneous mean field approximations because of the bad tractability of order statistics. We show how they become tractable in the thermodynamic limit of infinite network size. This enables the analytic description of node attacks that are characterized by threshold allocations based on node degree. Using two examples, we discuss possible implications of irregular phase transitions and different speeds of cascade evolution for the control of cascades.

nlin.AO↗

arXiv2017

A framework for cascade size calculations on random networks

We present a framework to calculate the cascade size evolution for a large class of cascade models on random network ensembles in the limit of infinite network size. Our method is exact and applies to network ensembles with almost arbitrary degree distribution, degree-degree correlations and, in case of threshold models, with arbitrary threshold distribution. With our approach, we shift the perspective from the known branching process approximations to the iterative update of suitable probability distributions. Such distributions are key to capture cascade dynamics that involve possibly continuous quantities and that depend on the cascade history, e.g. if load is accumulated over time. These distributions respect the Markovian nature of the studied random processes. Random variables capture the impact of nodes that have failed at any point in the past on their neighborhood. As a proof of concept, we provide two examples: (a) Constant load models that cover many of the analytically tractable cascade models, and, as a highlight, (b) a fiber bundle model that was not tractable by branching process approximations before. Our derivations cover the whole cascade dynamics, not only their steady state. This allows to include interventions in time or further model complexity in the analysis.

physics.soc-ph↗

arXiv2015

Systemic risk in multiplex networks with asymmetric coupling and threshold feedback

We study cascades on a two-layer multiplex network, with asymmetric feedback that depends on the coupling strength between the layers. Based on an analytical branching process approximation, we calculate the systemic risk measured by the final fraction of failed nodes on a reference layer. The results are compared with the case of a single layer network that is an aggregated representation of the two layers. We find that systemic risk in the two-layer network is smaller than in the aggregated one only if the coupling strength between the two layers is small. Above a critical coupling strength, systemic risk is increased because of the mutual amplification of cascades in the two layers. We even observe sharp phase transitions in the cascade size that are less pronounced on the aggregated layer. Our insights can be applied to a scenario where firms decide whether they want to split their business into a less risky core business and a more risky subsidiary business. In most cases, this may lead to a drastic increase of systemic risk, which is underestimated in an aggregated approach.

physics.soc-ph↗

arXiv2015

How Damage Diversification Can Reduce Systemic Risk

We consider the problem of risk diversification in complex networks. Nodes represent e.g. financial actors, whereas weighted links represent e.g. financial obligations (credits/debts). Each node has a risk to fail because of losses resulting from defaulting neighbors, which may lead to large failure cascades. Classical risk diversification strategies usually neglect network effects and therefore suggest that risk can be reduced if possible losses (i.e., exposures) are split among many neighbors (exposure diversification, ED). But from a complex networks perspective diversification implies higher connectivity of the system as a whole which can also lead to increasing failure risk of a node. To cope with this, we propose a different strategy (damage diversification, DD), i.e. the diversification of losses that are imposed on neighboring nodes as opposed to losses incurred by the node itself. Here, we quantify the potential of DD to reduce systemic risk in comparison to ED. For this, we develop a branching process approximation that we generalize to weighted networks with (almost) arbitrary degree and weight distributions. This allows us to identify systemically relevant nodes in a network even if their directed weights differ strongly. On the macro level, we provide an analytical expression for the average cascade size, to quantify systemic risk. Furthermore, on the meso level we calculate failure probabilities of nodes conditional on their system relevance.

physics.soc-ph↗