To increase mobile user engagement, photo sharing sites are trying to identify interesting and memorable pictures. Past proposals for identifying such pictures have relied on either metadata (e.g., likes) or visual features. In practice, techniques based on  those two inputs  do not always work: metadata is sparse (only few pictures have considerable number of likes), and extracting visual features is computationally expensive. In mobile solutions,  geo-referenced content becomes increasingly important. The premise behind this work is that pictures of a neighborhood is linked to the way the neighborhood is perceived by people: whether it is, for instance, distinctive and beautiful or not.  Since 1970s, urban theories proposed by Lynch, Milgram and Peterson aimed at systematically capturing the way people perceive neighborhoods. Here we tested whether those theories could be put to use for automatically identifying appealing city pictures.  We did so by gathering geo-referenced Flickr pictures in the city of London; selecting six urban qualities associated with those urban theories; computing proxies for those qualities from online social media data; and ranking Flickr pictures based on those proxies. We find that our proposal enjoys three main desirable properties: it is effective, scalable, and aware of contextual changes such as  time of day and weather condition. All this suggests new promising research directions for multi-modal learning approaches that automatically identify appealing city pictures.

Social media offer a real-time, unfiltered view of how disasters affect communities. Crisis response, disaster mental health, and — more broadly — public health can benefit from automated analysis of the public’s mental state as exhibited on social media. Our focus is on Twitter data from a community that lost members in a mass shooting and another community—geographically removed from the shooting — that was indirectly exposed. We show that a common approach for understanding emotional response in text: Linguistic Inquiry and Word Count (LIWC) can be substantially improved using machine learning. Starting with tweets flagged by LIWC as containing content related to the issue of death, we devise a categorization scheme for death-related tweets to induce automatic text classification of such content. This improved methodology reveals striking differences in the magnitude and duration of increases in death-related talk between these communities. It also detects subtle shifts in the nature of death-related talk. Our results offer lessons for gauging public response and for developing interventions in the wake of a tragedy.

The Olympic Games are an important sporting event with notable consequences for the general economic landscape of the host city. Traditional economic assessments focus on the aggregated impact of the event on the national income, but fail to provide micro-scale insights on why local businesses will benefit from the increased activity during the Games.In this paper we provide a novel approach to modeling the impact of the Olympic Games on local retailers by analyzing a dataset mined from a large location-based social service, Foursquare. We hypothesize that the spatial positioning of businesses as well as the mobility trends of visitors are primary indicators of whether retailers will rise their popularity during the event. To confirm this we formulate a retail winners prediction task in the context of which we evaluate a set of geographic and mobility metrics. We find that the proximity to stadiums, the diversity of activity in the neighborhood, the nearby area sociability, as well as the probability of customer flows from and to event places such as stadiums and parks are all vital factors. Through supervised learning techniques we demonstrate that the success of businesses hinges on a combination of both geographic and mobility factors. Our results suggest that location-based social networks, where crowdsourced information about the dynamic interaction of users with urban spaces becomes publicly available, present an alternative medium to assess the economic impact of large scale events in a city.

The internet has become a resource for adolescents who are distressed by social and emotional problems. Social network analysis can provide new opportunities for helping people seeking support online, but only if we understand the salient issues that are highly relevant to participants personal circumstances. In this paper, we present a stacked generalization modeling approach to analyze an online community supporting adolescents under duress. While traditional predictive supervised methods rely on robust hand-crafted feature space engineering, mixed initiative semi-supervised topic models are often better at extracting high-level themes that go beyond such feature spaces. We present a strategy that combines the strengths of both these types of models inspired by Prevention Science approaches which deals with the identification and amelioration of risk factors that predict to psychological, psychosocial, and psychiatric disorders within and across populations (in our case teenagers) rather than treat them post-facto. In this study, prevention scientists used a social science thematic analytic approach to code stories according to a fine-grained analysis of salient social, developmental or psychological themes they deemed relevant, and these are then analyzed by a society of models. We show that a stacked generalization of such an ensemble fares better than individual binary predictive models.

College students often have to team up for classprojects, and they select each other based not only onpast performance (e.g., grades) but also on whetherthey get along (e.g., whether they trust each other).There has not been any study on the relationshipbetween team formation for class projects and socialmedia. To fix that, we ask a group of university studentsto tell us with whom they wish to work, gather theironline Facebook data, and test the predictors of teamformation. We find that self-organized selection ofteam members does not strongly depend on pastgrades but rather on Facebook-derived proxies fortie strength, popularity, and homophily. These resultshave important theoretical implications for the teamformation literature and practical implications foronline educational platforms.

The use of EEG biometrics, for the purpose of automatic people recognition, has received increasing attention in the recent years. Most of current analysis rely on the extraction of features characterizing the activity of single brain regions, like power-spectrum estimates, thus neglecting possible temporal dependencies between the generated EEG signals. However, important physiological information can be extracted from the way different brain regions are functionally coupled. In this study, we propose a novel approach that fuses spectral coherencebased connectivity between different brain regions as a possibly viable biometric feature. The proposed approach is tested on a large dataset of subjects (N=108) during eyes-closed (EC) and eyes-open (EO) resting state conditions. The obtained recognition performances show that using brain connectivity leads to higher distinctiveness with respect to power-spectrum measurements, in both the experimental conditions. Notably, a 100% recognition accuracy is obtained in EC and EO when integrating functional connectivity between regions in the frontal lobe, while a lower 97.41% is obtained in EC (96.26% in EO) when fusing power spectrum information from centro-parietal regions. Taken together, these results suggest that functional connectivity patterns represent effective features for improving EEG-based biometric systems.

Deep learning has recently led to great successes in tasks such as image recognition (e.g Krizhevsky et al., 2012). However, deep networks are still outmatched by the power and versatility of the brain, perhaps in part due to the richer neuronal computations available to cortical circuits. The challenge is to identify which neuronal mechanisms are relevant, and to find suitable abstractions to model them. Here, we show how aspects of spike timing, long hypothesized to play a crucial role in cortical information processing, could be incorporated into deep networks to build richer, versatile representations. We introduce a neural network formulation based on complex-valued neuronal units that is not only biologically meaningful but also amenable to a variety of deep learning frameworks. Here, units are attributed both a firing rate and a phase, the latter indicating properties of spike timing. We show how this formulation qualitatively captures several aspects thought to be related to neuronal synchrony, including gating of information processing and dynamic binding of distributed object representations. Focusing on the latter, we demonstrate the potential of the approach in several simple experiments. Thus, neuronal synchrony could be a flexible mechanism that fulfills multiple functional roles in deep networks.

Bayesian optimization (Mockus et al., 1978) is a method for performing global optimization of unknown "black box" objectives that is particularly appropriate when objective function evaluations are expensive (in any sense, such as time or money). For example, consider a food company trying to design a low-calorie variant of a popular cookie. In this case, the design space is the space of possible recipes and might include several key parameters such as quantities of various ingredients and baking times. Each evaluation of a recipe entails computing (or perhaps actually measuring) the number of calories in the proposed cookie. Bayesian optimization can be used to propose new candidate recipes such that good results are found with few evaluations. Now suppose the company also wants to ensure the taste of the cookie is not compromised when calories are reduced. Therefore, for each proposed low-calorie recipe, they perform a taste test with sample customers. Because different people, or the same people at different times, have differing opinions about the taste of cookies, the company decides to require that at least 95% of test subjects must like the new cookie.

Geographical location is vital to geospatial applications like local search and event detection. In this paper, we investigate and improve on the task of text-based geolocation prediction of Twitter users. Previous studies on this topic have typically assumed that geographical references (e.g., gazetteer terms, dialectal words) in a text are indicative of its authors location. However, these references are often buried in informal, ungrammatical, and multilingual data, and are therefore non-trivial to identify and exploit. We present an integrated geolocation prediction framework and investigate what factors impact on prediction accuracy. First, we evaluate a range of feature selection methods to obtain location indicative words. We then evaluate the impact of non-geotagged tweets, language, and user-declared metadata on geolocation prediction. In addition, we evaluate the impact of temporal variance on model generalisation, and discuss how users differ in terms of their geolocatability. We achieve state-of-the-art results for the text-based Twitter user geolocation task, and also provide the most extensive exploration of the task to date. Our findings provide valuable insights into the design of robust, practical text-based geolocation prediction systems.

We present a supervised-learning algorithm from graph data (a set of graphs) for arbitrary twice-differentiable loss functions and sparse linear models over all possible subgraph features. To date, it has been shown that under all possible subgraph features, several types of sparse learning, such as Adaboost, LPBoost, LARS/LASSO, and sparse PLS regression, can be performed. Particularly emphasis is placed on simultaneous learning of relevant features from an infinite set of candidates. We first generalize techniques used in all these preceding studies to derive an unifying bounding technique for arbitrary separable functions. We then carefully use this bounding to make block coordinate gradient descent feasible over infinite subgraph features, resulting in a fast converging algorithm that can solve a wider class of sparse learning problems over graph data. We also empirically study the differences from the existing approaches in convergence property, selected subgraph features, and search-space sizes. We further discuss several unnoticed issues in sparse learning over all possible subgraph features.