We propose an algorithm to find the community structure in complex networks based on the combination of spectral analysis and modularity optimization. The clustering produced by our algorithm is as accurate as the best algorithms on the literature of modularity optimization; however, the main asset of the algorithm is its efficiency. The best match for our algorithm is Newman’s fast algorithm, which is the reference algorithm for clustering in large networks due to its efficiency. When both algorithms are compared, our algorithm outperforms the fast algorithm both in efficiency and accuracy of the clustering, in terms of modularity. Thus, the results suggest that the proposed algorithm is a good choice to analyze the community structure of medium and large networks in the range of tens and hundreds of thousand vertices.

To balance electricity production and demand, it is required to use different prediction techniques extensively. Renewable energy, due to its intermittency, increases the complexity and uncertainty of forecasting, and the resulting accuracy impacts all the different players acting around the electricity systems around the world like generators, distributors, retailers, or consumers. Wind forecasting can be done under two major approaches, using meteorological numerical prediction models or based on pure time series input. Deep learning is appearing as a new method that can be used for wind energy prediction. This work develops several deep learning architectures and shows their performance when applied to wind time series. The models have been tested with the most extensive wind dataset available, the National Renewable Laboratory Wind Toolkit, a dataset with 126,692 wind points in North America. The architectures designed are based on different approaches, Multi-Layer Perceptron Networks (MLP), Convolutional Networks (CNN), and Recurrent Networks (RNN). These deep learning architectures have been tested to obtain predictions in a 12-h ahead horizon, and the accuracy is measured with the coefficient of determination, the R² method. The application of the models to wind sites evenly distributed in the North America geography allows us to infer several conclusions on the relationships between methods, terrain, and forecasting complexity. The results show differences between the models and confirm the superior capabilities on the use of deep learning techniques for wind speed forecasting from wind time series data.

Location-based social networks (LBSNs) such as Twitter or Instagram are a good source for user spatio-temporal behaviour. These networks collect data from users in such a way that they can be seen as a set of collective and distributed sensors of a geographical area. A low rate sampling of user's location information can be obtained during large intervals of time that can be used to discover complex patterns, including mobility profiles, points of interest or unusual events. These patterns can be used as the elements of a knowledge base for different applications in different domains such as mobility route planning, touristic recommendation systems or city planning. The aim of this paper is twofold, first to analyse the frequent spatio-temporal patterns that users share when living and visiting a city. This behaviour is studied by means of frequent itemsets algorithms in order to establish some associations among visits that can be interpreted as interesting routes or spatio-temporal connections. Second, to analyse how the spatio-temporal behaviour of a large number of users can be segmented in different profiles. These behavioural profiles are obtained by means of clustering algorithms that show the different patterns of behaviour of visitors and citizens. The data analysed were obtained from the public data feeds of Twitter and Instagram within an area surrounding the cities of Barcelona and Milan for a period of several months. The analysis of these data shows that these kinds of algorithms can be successfully applied to data from any city (or general area) to discover useful patterns that can be interpreted on terms of singular places and areas and their temporal relationships.

Vector-space word representations obtained from neural network models have been shown to enable semantic operations based on vector arithmetic. In this paper, we explore the existence of similar information on vector representations of images. For that purpose we define a methodology to obtain large, sparse vector representations of image classes, and generate vectors through the state-of-the-art deep learning architecture GoogLeNet for 20 K images obtained from ImageNet. We first evaluate the resultant vector-space semantics through its correlation with WordNet distances, and find vector distances to be strongly correlated with linguistic semantics. We then explore the location of images within the vector space, finding elements close in WordNet to be clustered together, regardless of significant visual variances (e.g., 118 dog types). More surprisingly, we find that the space unsupervisedly separates complex classes without prior knowledge (e.g., living things). Afterwards, we consider vector arithmetics. Although we are unable to obtain meaningful results on this regard, we discuss the various problem we encountered, and how we consider to solve them. Finally, we discuss the impact of our research for cognitive systems, focusing on the role of the architecture being used.This work is partially supported by the Joint Study Agreement no. W156463 under the IBM/BSC Deep Learning Center agreement, by the Spanish Government through Programa Severo Ochoa (SEV-2015-0493), by the Spanish Ministry of Science and Technology through TIN2015-65316-P project and by the Generalitat de Catalunya (contracts 2014-SGR-1051), and by the Core Research for Evolutional Science and Technology (CREST) program of Japan Science and Technology Agency (JST).Peer ReviewedPostprint (published version