Search

Transfer Learning (TL) encompasses a number of Machine Learning Techniques that take a pre-trained model aimed at solving a task in a Source Domain and try to reuse it to improve the performance of a related task in a Target Domain An important issue in TL is that the effectiveness of those techniques is strongly dataset-dependent. In this work, we investigate the possible structural causes of the varying performance of Heterogeneous Transfer Learning (HTL) across domains characterized by different, but overlapping feature sets (this naturally determine a partition of the features into Source Domain specific subset, Target Domain specific subset, and shared subset). To this purpose, we use the Partial Information Decomposition (PID) framework, which breaks down the multivariate information that input variables hold about an output variable into three kinds of components: Unique, Synergistic, and Redundant. We consider that each domain can hold the PID components in implicit form: this restricts the information directly accessible to each domain. Based on the relative PID structure of the above mentioned feature subsets, the framework is able to tell, in principle: 1) which kind of information components are lost in passing from one domain to the other, 2) which kind of information components are at least implicitly available to a domain, and 3) what kind information components could be recovered through the bridge of the shared features. We show an example of a bridging scenario based on synthetic data.

Transfer Learning (TL) encompasses a number of Machine Learning Techniques that take a pre-trained model aimed at solving a task in a Source Domain and try to reuse it to improve the performance of a related task in a Target Domain An important issue in TL is that the effectiveness of those techniques is strongly dataset-dependent. In this work, we investigate the possible structural causes of the varying performance of Heterogeneous Transfer Learning (HTL) across domains characterized by different, but overlapping feature sets (this naturally determine a partition of the features into Source Domain specific subset, Target Domain specific subset, and shared subset). To this purpose, we use the Partial Information Decomposition (PID) framework, which breaks down the multivariate information that input variables hold about an output variable into three kinds of components: Unique, Synergistic, and Redundant. We consider that each domain can hold the PID components in implicit form: this restricts the information directly accessible to each domain. Based on the relative PID structure of the above mentioned feature subsets, the framework is able to tell, in principle: 1) which kind of information components are lost in passing from one domain to the other, 2) which kind of information components are at least implicitly available to a domain, and 3) what kind information components could be recovered through the bridge of the shared features. We show an example of a bridging scenario based on synthetic data.

Graphs, with their innate ability to encapsulate both topological and semantic information, have become a first-class tool in the landscape of pattern recognition and machine learning to represent weakly structured and heterogeneous data across countless domains. The field has evolved significantly since its early stages, with researchers pushing the boundaries of what graph-based learning algorithms can do. Inspired by the success of other deep learning models, various methods were introduced to redefine graph convolution, offering a suitable representation for spatio-temporal data as well. Graph-based representations find applications in diverse domains like computer vision, natural language processing, traffic forecasting, and molecular graph structure analysis in chemistry. This year's edition of the Graph Models for Learning and Recognition track is a testament to the relentless pursuit in understanding and harnessing the potential of these complex structures, thus reflecting the ever-expanding use of graph-based representations and models. The conference proceedings include a diverse range of papers, each addressing unique facets of graph-based representation, prediction, and model design, as well as related applications. We received high quality papers from all over the world, with a total of 15 submissions. The review process was very competitive, with each paper going through at least three reviews. As a result, 4 full papers were selected for presentation in the track. Three of them deal with enhancements in graph structure representation, in particular improving metapath-based processing on heterogeneous graphs and node classification, detecting change points in evolving graphs based on martingale theory, and exploring the impact of multiplicative integration-based layers for improved neighbourhood aggregation in GCNs. The fourth paper applies GCNs on 3D facial models for effective gender classification, showcasing results on the BP4D+ dataset. We would l

We study the local limit distribution of the number of occurrences of a symbol in words of length n generated at random in a regular language according to a rational stochastic model. We present an analysis of the main local limits when the finite state automaton defining the stochastic model consists of two primitive components. The limit distributions depend on several parameters and conditions, such as the main constants of mean value and variance of our statistics associated with the two components, and the existence of communications from the first to the second component. The convergence rate of these results is always of order $O(n^-1/2)$. For the same statistics we also prove an analogous $O(n^-1/2)$ convergence rate of the Gaussian local limit law whenever the stochastic model consists of one primitive component.

Arguably one of the most notable forms of the principle of parsimony was formulated by the philosopher and theologian William of Ockham in the 14th century, and later became well known as Ockham’s Razor principle, which can be phrased as: “Entities should not be multiplied without necessity.” This principle is undoubtedly one of the most fundamental ideas that pervade many branches of knowledge, from philosophy to art and science, from ancient times to modern age, then summarized in the expression “Make everything as simple as possible, but not simpler” as likewise asserted by Albert Einstein. The sparse modeling is an evident manifestation capturing the parsimony principle just described, and sparse models are widespread in statistics, physics, information sciences, neuroscience, computational mathematics, and so on. In statistics the many applications of sparse modeling span regression, classification tasks, graphical model selection, sparse M-estimators, and sparse dimensionality reduction. It is also particularly effective in many statistical and machine learning areas where the primary goal is to discover predictive patterns from data, which would enhance our understanding and control of underlying physical, biological, and other natural processes, beyond just building accurate outcome black-box predictors. Common examples include selecting biomarkers in biological procedures, finding relevant brain activity locations, which are predictive about brain states and processes based on fMRI data, and identifying network bottlenecks best explaining end-to-end performance. Moreover, the research and applications of efficient recovery of high-dimensional sparse signals from a relatively small number of observations, which is the main focus of compressed sensing or compressive sensing, have rapidly grown and became an extremely intense area of study beyond classical signal processing. Likewise interestingly, sparse modeling is directly related to various artificial visi

The sparse modeling is an evident manifestation capturing the parsimony principle just described, and sparse models are widespread in statistics, physics, information sciences, neuroscience, computational mathematics, and so on. In statistics the many applications of sparse modeling span regression, classification tasks, graphical model selection, sparse M-estimators and sparse dimensionality reduction. It is also particularly effective in many statistical and machine learning areas where the primary goal is to discover predictive patterns from data which would enhance our understanding and control of underlying physical, biological, and other natural processes, beyond just building accurate outcome black-box predictors. Common examples include selecting biomarkers in biological procedures, finding relevant brain activity locations which are predictive about brain states and processes based on fMRI data, and identifying network bottlenecks best explaining end-to-end performance. Moreover, the research and applications of efficient recovery of high-dimensional sparse signals from a relatively small number of observations, which is the main focus of compressed sensing or compressive sensing, have rapidly grown and became an extremely intense area of study beyond classical signal processing. Likewise interestingly, sparse modeling is directly related to various artificial vision tasks, such as image denoising, segmentation, restoration and superresolution, object or face detection and recognition in visual scenes, and action recognition. In this manuscript, we provide a brief introduction of the basic theory underlying sparse representation and compressive sensing, and then discuss some methods for recovering sparse solutions to optimization problems in effective way, together with some applications of sparse recovery in a machine learning problem known as sparse dictionary learning., Comment: 42 pages

We propose the analysis of a scalable parallel MCMC algorithm for graph coloring aimed at balancing the color class sizes, provided that a suitable number of colors is made available. Firstly, it is shown that the Markov chain converges to the target distribution by repeatedly sampling from suitable proposed distributions over the neighboring colors of each node, independently and hence in parallel manner. We prove that the number of conflicts in the improper colorings genereted thoughout the iterations of the algorithm rapidly converges in probability to 0. As for the balancing, given to the complexity of the distributions involved, we propose a qualitative analysis about the balancing level achieved. Based on a collection of multinoulli distributions arising from the color occurrences within every node neighborhood, we provide some evidence about the character of the final color balancing, which results to be nearly uniform over the color classes. Some numerical simulations on big social graphs confirm the fast convergence and the balancing trend, which is validated through a statistical hypothesis test eventually.

Optimal well-being is a new multi-dimensional construct, which incorporates the well-known preexisting notions of subjective and psychological well-being. Classical models for describing the predictors of optimal well-being binary response variable are generalized linear models (GLM), such as logistic regression model. Since the number of predictors might be relatively large in these models, we devise a sparse optimization method for the regression problem based on subsequent iterations of a suitable sparse quadratic approximant problem, so that the resulting parameter vector estimate is sparse and indicates few significant predictors. We conduct empirical assessments using data of the European Social Survey, in order to identify the set of determinants which better predict optimal well-being by means of the proposed sparse regression method. ESS data analysis confirms that few selected predictors provide good data interpretation and no loss of information in the frequency of correct classification for people meeting the criteria of optimal well-being. Moreover, simulations with different structural parameter values indicate that sparse logistic model performs better in terms of the estimation of the true vector of parameters in a more parsimonious setting compared to classical logistic regression. The benefits increase as the structural sparsity of the optimization problem becomes stronger.

Given a class label y assigned by a classifier to a point x in feature space, the counterfactual generation task, in its simplest form, consists of finding the minimal edit that moves the feature vector to a new point x′, which the classifier maps to a pre-specified target class y′≠ y. Counterfactuals provide a local explanation to a classifier model, by answering the questions “Why did the model choose y instead of y′ : what changes to x would make the difference?". An important aspect in classification is ambiguity: typically, the description of an instance is compatible with more than one class. When ambiguity is too high, a suitably designed classifier can map an instance x to a class set Y of alternatives, rather than to a single class, so as to reduce the likelihood of wrong decisions. In this context, known as set-based classification, one can discuss set-based counterfactuals. In this work, we extend the counterfactual generation problem – normally expressed as a constrained optimization problem – to set-based counterfactuals. Using non-singleton counterfactuals, rather than singletons, makes the problem richer under several aspects, related to the fact that non-singleton sets allow for a wider spectrum of relationships among them: (1) the specification of the target set-based class Y′ is more varied (2) the target solution x′ that ought to be mapped to Y′ is not granted to exist, and, in that case, (3) since one might end up with the availability of a number of feasible alternatives to Y′, one has to include the degree of partial fulfillment of the solution into the loss function of the optimization problem.

La tipologia Editorial message non possiede abstract.

N/A

We study the local limit distribution of the number of occurrences of a symbol in words of length $n$ generated at random in a regular language according to a rational stochastic model. We present an analysis of the main local limits when the finite state automaton defining the stochastic model consists of two primitive components. The limit distributions depend on several parameters and conditions, such as the main constants of mean value and variance of our statistics associated with the two components, and the existence of communications from the first to the second component. The convergence rate of these results is always of order $O(n^{-1/2})$. We also prove an analogous $O(n^{-1/2})$ convergence rate to a Gaussian density of the same statistic whenever the stochastic models only consists of one (primitive) component., Comment: 26 pages. It includes works presented at the following conferences: 20th DCFS (LNCS 10952, 114-125), 23rd DLT (LNCS 11647, 306-318), 20th ICTCS (CEUR-WS.ORG 2504, 78-90)

Classification learning is a preeminent human ability within the animal kingdom but the key mechanisms of brain networks regulating learning remain mostly elusive. Recent neuroimaging advancements have depicted human brain as a complex graph machinery where brain regions are nodes and coherent activities among them represent the functional connections. While long-term motor memories have been found to alter functional connectivity in the resting human brain, a graph topological investigation of the short-time effects of learning are still not widely investigated. For instance, classification learning is known to orchestrate rapid modulation of diverse memory systems like short-term and visual working memories but how the brain functional connectome accommodates such modulations is unclear. We used publicly available repositories (openfmri.org) selecting three experiments, two focused on short-term classification learning along two consecutive runs where learning was promoted by trial-by-trial feedback errors, while a further experiment was used as supplementary control. We analyzed the functional connectivity extracted from BOLD fMRI signals, and estimated the graph information processing in the cerebral networks. The information processing capability, characterized by complex network statistics, significantly improved over runs, together with the subject classification accuracy. Instead, null-learning experiments, where feedbacks came with poor consistency, did not provoke any significant change in the functional connectivity over runs. We propose that learning induces fast modifications in the overall brain network dynamics, definitely ameliorating the short-term potential of the brain to process and integrate information, a dynamic consistently orchestrated by modulations of the functional connections among specific brain regions.

Within the paradigm of Mobile Cloud Computing (MCC) mobile devices such as mobile phones and tablets can unload computation to a local Cloud consisting of both static and mobile devices. Due to the proximity, such Cloud can better fulfill service and latency requirements for QoS sensitive applications. However, in order to function properly, such mechanics requires a collaborative approach that not all the devices might follow: some can contribute much less than others and deviate from the Service Level Agreement. This creates a free-riding problem in MCC and a corresponding QoS issue. An approach towards this problem consists of providing incentives to nodes so that they act as inspectors and occasionally audit the recent behavior of the nodes with which they interact. This double role of potential inspectee and inspectors can be modeled within Game Theory (GT) to predict the behavior of the agents. In this work, we show that an appropriate model for this game is a symmetric four-strategies social dilemma. The solution (a.k.a. equilibrium) of the game provides a unique way to set the incentives so as to drive the system toward the desired behavior. However, even the hybrid inspector-inspectee agent approach is in principle open to a potential flaw: the possibility of inspector-inspectee collusion. The main contribution of this work is that the collusion behavior is not an equilibrium of the game, thus hybrid agent rational players should never collude.

The massive size and complexity of big datasets such as those coming from social, natural and sensor environments raise utmost challenges to unsupervised cluster analysis methods in terms of performance scalability in designing algorithms, also considering parallel and distributed networking context. To cope with these hindrances, the parallelization of clustering techniques, also benefiting from GPU-centered computation, can contribute to fill the gap in applicative areas such as optimization of network routing or management of large-scale IoT networks, thus enabling the extraction, processing and policy making relying on rich network information that are typically represented in the form of graphs. One established approach to clustering graphs is through the coloring techniques, and indeed, graph clustering and graph coloring can be viewed as tied. We devise a graph clustering technique based on a Markov Chain method aimed at b-coloring the data points, that works in efficient vertex-centric parallel manner and produces a valid clustering with reduced number of color classes. We assess our algorithm against synthetic data encapsulating group structure characteristics and present a brief convergence analysis of the method.

In this work we address the problem of critical source selection in social sensing. We propose an approach to the ranking of information streams, which is aware of the interdependence among streams (redundancy and synergies), of the cost of individual streams, and of the cost related to the integration of multiple streams. The method is based on the use of the Coalitional Game Theory concept of Power Index, and relies on the polynomial-time estimate of the stream sets characteristics. With respect to other works using a power index, the method takes into account that the problem has a non-trivial cost structure.

Face recognition using a single reference image per subject is challenging, above all when referring to a large gallery of subjects. Furthermore, the problem hardness seriously increases when the images are acquired in unconstrained conditions. In this paper we address the challenging Single Sample Per Person (SSPP) problem considering large datasets of images acquired in the wild, thus possibly featuring illumination, pose, face expression, partial occlusions, and low-resolution hurdles. The proposed technique alternates a sparse dictionary learning technique based on the method of optimal direction and the iterative l 0 -norm minimization algorithm called k-LIMAPS. It works on robust deep-learned features, provided that the image variability is extended by standard augmentation techniques. Experiments show the effectiveness of our method against the hardness introduced above: first, we report extensive experiments on the unconstrained LFW dataset when referring to large galleries up to 1680 subjects; second, we present experiments on very low-resolution test images up to 8 × 8 pixels; third, tests on the AR dataset are analyzed against specific disguises such as partial occlusions, facial expressions, and illumination problems. In all the three scenarios our method outperforms the state-of-the-art approaches adopting similar configurations.

In a recent work we have determined the local limit distribution of pattern statistics representing the number of occurrences of a symbol in words of length n in a regular language generated at random according to a suitable stochastic model. Such a model is defined by a finite automaton with weights in R+, consisting of two primitive components, having some transition from the first to the second component. In the present work we extend those results to the case when there is no communication among the components, and hence the associated formal series is the sum of two rational series recognized by finite state automata with primitive transition matrix. We obtain local limit laws of Gaussian type when there is a dominant component or when, in equipotent case, the main terms of mean value and variance are equal. On the contrary, if these terms are not the same then the local limit distribution is a convex combination of Gaussian laws. All convergence rates of our limits are of the order O(n−1/2). This completes the analysis of local limit laws of symbol statistics under a bicomponent stochastic model1.

We study the local limit distribution of sequences of random variables representing the number of occurrences of a symbol in words of length n in a regular language, generated at random according to a rational stochastic model. We present an analysis of the main local limits when the finite state automaton defining the stochastic model consists of two primitive components. Our results include an evaluation of the convergence rate, which in the various cases is of an order slightly slower than O(n−1/2).

Complete knowledge of all direct and indirect interactions between proteins in a given cell would represent an important milestone towards a comprehensive description of cellular mechanisms and functions. Although this goal is still elusive, considerable progress has been made—particularly for certain model organisms and functional systems. Currently, protein interactions and associations are annotated at various levels of detail in online resources, ranging from raw data repositories to highly formalized pathway databases. For many applications, a global view of all the available interaction data is desirable, including lower-quality data and/or computational predictions. The STRING database (http://string-db.org/) aims to provide such a global perspective for as many organisms as feasible. Known and predicted associations are scored and integrated, resulting in comprehensive protein networks covering >1100 organisms. Here, we describe the update to version 9.1 of STRING, introducing several improvements: (i) we extend the automated mining of scientific texts for interaction information, to now also include full-text articles; (ii) we entirely re-designed the algorithm for transferring interactions from one model organism to the other; and (iii) we provide users with statistical information on any functional enrichment observed in their networks

A successful approach for predicting functional associations between non-homologous genes is to compare their phylogenetic distributions. We have devised a phylogenetic profiling algorithm, SVD-Phy, which uses truncated singular value decomposition to address the problem of uninformative profiles giving rise to false positive predictions. Benchmarking the algorithm against the KEGG pathway database, we found that it has substantially improved performance over existing phylogenetic profiling methods.AVAILABILITY AND IMPLEMENTATION: The software is available under the open-source BSD license at https://bitbucket.org/andrea/svd-phy CONTACT: lars.juhl.jensen@cpr.ku.dk.

A successful approach for predicting functional associations between non-homologous genes is to compare their phylogenetic distributions. We have devised a phylogenetic profiling algorithm, SVD-Phy, which uses truncated singular value decomposition to address the problem of uninformative profiles giving rise to false positive predictions. Benchmarking the algorithm against the KEGG pathway database, we found that it has substantially improved performance over existing phylogenetic profiling methods. AVAILABILITY AND IMPLEMENTATION: The software is available under the open-source BSD license at https://bitbucket.org/andrea/svd-phy CONTACT: lars.juhl.jensen@cpr.ku.dk Supplementary information: Supplementary data are available at Bioinformatics online.

Pseudocapacitive materials, such as binary metal sulfides, show great promise as electrode candidates for energy storage devices due to their higher specific capacitance than that of mono-metal sulfides and binary metal oxides, but generally suffer from low energy densities when assembled in supercapacitor devices. To push the energy density limit of pseudocapacitive materials in devices, a new class of electrode materials with favorable architectures is strongly needed. Here, rationally designed and coaxially grown Ni-Co-S and polypyrrole on vertically aligned carbon nanotube (VA-CNT) arrays/3D carbon fiber paper (CFP) is presented as a novel freestanding electrode for energy storage devices. Our study has revealed that the catalyst preparation is the key step and the presence of an Al2O3 buffer layer is essential for the growth of VA-CNTs. The 3D hierarchical VA-CNTs/CFP allows high areal loading and high utilization efficiency of the active materials. The binder-free asymmetric energy storage devices with Ni-Co-S/VA-CNTs/CFP as the positive electrode and polypyrrole/VA-CNTs/CFP as the negative electrode, respectively, demonstrate a high energy density of 82 W h kg-1at 200 W kg-1.

Pseudocapacitive materials, such as binary metal sulfides, show great promise as electrode candidates for energy storage devices due to their higher specific capacitance than that of mono-metal sulfides and binary metal oxides, but generally suffer from low energy densities when assembled in supercapacitor devices. To push the energy density limit of pseudocapacitive materials in devices, a new class of electrode materials with favorable architectures is strongly needed. Here, rationally designed and coaxially grown Ni-Co-S and polypyrrole on vertically aligned carbon nanotube (VA-CNT) arrays/3D carbon fiber paper (CFP) is presented as a novel freestanding electrode for energy storage devices. Our study has revealed that the catalyst preparation is the key step and the presence of an Al2O3 buffer layer is essential for the growth of VA-CNTs. The 3D hierarchical VA-CNTs/CFP allows high areal loading and high utilization efficiency of the active materials. The binder-free asymmetric energy storage devices with Ni-Co-S/VA-CNTs/CFP as the positive electrode and polypyrrole/VA-CNTs/CFP as the negative electrode, respectively, demonstrate a high energy density of 82 W h kg-1at 200 W kg-1.

Pseudocapacitive materials, such as binary metal sulfides, show great promise as electrode candidates for energy storage devices due to their higher specific capacitance than that of mono-metal sulfides and binary metal oxides, but generally suffer from low energy densities when assembled in supercapacitor devices. To push the energy density limit of pseudocapacitive materials in devices, a new class of electrode materials with favorable architectures is strongly needed. Here, rationally designed and coaxially grown Ni-Co-S and polypyrrole on vertically aligned carbon nanotube (VA-CNT) arrays/3D carbon fiber paper (CFP) is presented as a novel freestanding electrode for energy storage devices. Our study has revealed that the catalyst preparation is the key step and the presence of an Al2O3 buffer layer is essential for the growth of VA-CNTs. The 3D hierarchical VA-CNTs/CFP allows high areal loading and high utilization efficiency of the active materials. The binder-free asymmetric energy storage devices with Ni-Co-S/VA-CNTs/CFP as the positive electrode and polypyrrole/VA-CNTs/CFP as the negative electrode, respectively, demonstrate a high energy density of 82 W h kg-1at 200 W kg-1.

Complete knowledge of all direct and indirect interactions between proteins in a given cell would represent an important milestone towards a comprehensive description of cellular mechanisms and functions. Although this goal is still elusive, considerable progress has been made-particularly for certain model organisms and functional systems. Currently, protein interactions and associations are annotated at various levels of detail in online resources, ranging from raw data repositories to highly formalized pathway databases. For many applications, a global view of all the available interaction data is desirable, including lower-quality data and/or computational predictions. The STRING database (http://string-db.org/) aims to provide such a global perspective for as many organisms as feasible. Known and predicted associations are scored and integrated, resulting in comprehensive protein networks covering >1100 organisms. Here, we describe the update to version 9.1 of STRING, introducing several improvements: (i) we extend the automated mining of scientific texts for interaction information, to now also include full-text articles; (ii) we entirely re-designed the algorithm for transferring interactions from one model organism to the other; and (iii) we provide users with statistical information on any functional enrichment observed in their networks.

To assess the effectiveness of urban energy conservation and GHG mitigation measures, a detailed Long- range Energy Alternatives Planning (LEAP) model is developed and applied to analyze the future trends of energy demand and GHG emissions in Xiamen city. Two scenarios have been designed to describe the future energy strategies in relation to the development of Xiamen city. The 'Business as Usual' scenario assumes that the government will do nothing to influence the long-term trends of urban energy demand. An 'Integrated' scenario, on the other hand, is generated to assess the cumulative impact of a series of available reduction measures: clean energy substitution, industrial energy conservation, combined heat and power generation, energy conservation in building, motor vehicle control, and new and renewable energy development and utilization. The reduction potentials in energy consumption and GHG emissions are estimated for a time span of 2007-2020 under these different scenarios. The calculation results in Xiamen show that the clean energy substitution measure is the most effective in terms of energy saving and GHG emissions mitigation, while the industrial sector has the largest abatement potential.

The study of nano-cylinder structure has attracted much attention due to the application of multi-wall carbon nanotubes (MWCNTs). while some TEM observations indicate that they are formed by seamless concentric cylinders, other TEM and high pressure X-ray diffraction studies suggest that they look like scrolls of graphite sheets. Although many people now accept the concentric cylinder model, there has been no confirmation reported. On the other hand, this structural difference of MWCNTs plays a crucial role in determining the properties and suitability for future applications. For example, the periodical boundary condition can only be imposed for cylinders, but not for scrolls. To resolve this issue, we employed high-resolution X-ray diffraction to measure detailed profiles of the Bragg peaks for high-purity MWCNTs. We then identified some unusual observations unique to the nano-cylinder structure, followed by the analysis of the structural difference in the Fourier transform between nanotubes formed by scrolls and concentric cylinders. The simulation results are then compared with the experimental data to reveal the structural details., Pub in: Materials Research Society Symposium Proceedings, Volume 703. Prepared in collaboration with Axcel Photonics Inc., 45 Bartlett St., Marlborough, MA and Xerox Research Center of Canada, Mississauga, Ontario, Canada. This article is from ADA401575 Nanophase and Nanocomposite Materials IV held in Boston, Massachusetts on November 26-29, 2001

Complete knowledge of all direct and indirect interactions between proteins in a given cell would represent an important milestone towards a comprehensive description of cellular mechanisms and functions. Although this goal is still elusive, considerable progress has been made—particularly for certain model organisms and functional systems. Currently, protein interactions and associations are annotated at various levels of detail in online resources, ranging from raw data repositories to highly formalized pathway databases. For many applications, a global view of all the available interaction data is desirable, including lower-quality data and/or computational predictions. The STRING database (http://string-db.org/) aims to provide such a global perspective for as many organisms as feasible. Known and predicted associations are scored and integrated, resulting in comprehensive protein networks covering >1100 organisms. Here, we describe the update to version 9.1 of STRING, introducing several improvements: (i) we extend the automated mining of scientific texts for interaction information, to now also include full-text articles; (ii) we entirely re-designed the algorithm for transferring interactions from one model organism to the other; and (iii) we provide users with statistical information on any functional enrichment observed in their networks