arXiv subjects

David Garcia

Publications and source records attributed to David Garcia.

At least 55 records · Page 3Linked to original sources

arXiv2020

Dashboard of sentiment in Austrian social media during COVID-19

To track online emotional expressions of the Austrian population close to real-time during the COVID-19 pandemic, we build a self-updating monitor of emotion dynamics using digital traces from three different data sources. This enables decision makers and the interested public to assess issues such as the attitude towards counter-measures taken during the pandemic and the possible emergence of a (mental) health crisis early on. We use web scraping and API access to retrieve data from the news platform derstandard.at, Twitter and a chat platform for students. We document the technical details of our workflow in order to provide materials for other researchers interested in building a similar tool for different contexts. Automated text analysis allows us to highlight changes of language use during COVID-19 in comparison to a neutral baseline. We use special word clouds to visualize that overall difference. Longitudinally, our time series show spikes in anxiety that can be linked to several events and media reporting. Additionally, we find a marked decrease in anger. The changes last for remarkably long periods of time (up to 12 weeks). We discuss these and more patterns and connect them to the emergence of collective emotions. The interactive dashboard showcasing our data is available online under http://www.mpellert.at/covid19_monitor_austria/. Our work has attracted media attention and is part of an web archive of resources on COVID-19 collected by the Austrian National Library.

cs.CY↗

arXiv2020

Improving accuracy and speeding up Document Image Classification through parallel systems

This paper presents a study showing the benefits of the EfficientNet models compared with heavier Convolutional Neural Networks (CNNs) in the Document Classification task, essential problem in the digitalization process of institutions. We show in the RVL-CDIP dataset that we can improve previous results with a much lighter model and present its transfer learning capabilities on a smaller in-domain dataset such as Tobacco3482. Moreover, we present an ensemble pipeline which is able to boost solely image input by combining image model predictions with the ones generated by BERT model on extracted text by OCR. We also show that the batch size can be effectively increased without hindering its accuracy so that the training process can be sped up by parallelizing throughout multiple GPUs, decreasing the computational time needed. Lastly, we expose the training performance differences between PyTorch and Tensorflow Deep Learning frameworks.

cs.CV↗

arXiv2020

A Streaming On-Device End-to-End Model Surpassing Server-Side Conventional Model Quality and Latency

Thus far, end-to-end (E2E) models have not been shown to outperform state-of-the-art conventional models with respect to both quality, i.e., word error rate (WER), and latency, i.e., the time the hypothesis is finalized after the user stops speaking. In this paper, we develop a first-pass Recurrent Neural Network Transducer (RNN-T) model and a second-pass Listen, Attend, Spell (LAS) rescorer that surpasses a conventional model in both quality and latency. On the quality side, we incorporate a large number of utterances across varied domains to increase acoustic diversity and the vocabulary seen by the model. We also train with accented English speech to make the model more robust to different pronunciations. In addition, given the increased amount of training data, we explore a varied learning rate schedule. On the latency front, we explore using the end-of-sentence decision emitted by the RNN-T model to close the microphone, and also introduce various optimizations to improve the speed of LAS rescoring. Overall, we find that RNN-T+LAS offers a better WER and latency tradeoff compared to a conventional model. For example, for the same latency, RNN-T+LAS obtains a 8% relative improvement in WER, while being more than 400-times smaller in model size.

cs.CL↗

arXiv2020

An agent-based model of multi-dimensional opinion dynamics and opinion alignment

It is known that individual opinions on different policy issues often align to a dominant ideological dimension (e.g. "left" vs. "right") and become increasingly polarized. We provide an agent-based model that reproduces these two stylized facts as emergent properties of an opinion dynamics in a multi-dimensional space of continuous opinions. The mechanisms for the change of agents' opinions in this multi-dimensional space are derived from cognitive dissonance theory and structural balance theory. We test assumptions from proximity voting and from directional voting regarding their ability to reproduce the expected emerging properties. We further study how the emotional involvement of agents, i.e. their individual resistance to change opinions, impacts the dynamics. We identify two regimes for the global and the individual alignment of opinions. If the affective involvement is high and shows a large variance across agents, this fosters the emergence of a dominant ideological dimension. Agents align their opinions along this dimension in opposite directions, i.e. create a state of polarization.

physics.soc-ph↗

arXiv2019

Modeling User Reputation in Online Social Networks: The Role of Costs, Benefits, and Reciprocity

We analyze an agent-based model to estimate how the costs and benefits of users in an online social network (OSN) impact the robustness of the OSN. Benefits are measured in terms of relative reputation that users receives from their followers. They can be increased by direct and indirect reciprocity in following each other, which leads to a core-periphery structure of the OSN. Costs relate to the effort to login, to maintain the profile, etc. and are assumed as constant for all users. The robustness of the OSN depends on the entry and exit of users over time. Intuitively, one would expect that higher costs lead to more users leaving and hence to a less robust OSN. We demonstrate that an optimal cost level exists, which maximizes both the performance of the OSN, measured by means of the long-term average benefit of its users, and the robustness of the OSN, measured by means of the life-time of the core of the OSN. Our mathematical and computational analyses unfold how changes in the cost level impact reciprocity and subsequently the core-periphery structure of the OSN, to explain the optimal cost level.

cs.SI↗

arXiv2019

How emotions drive opinion polarization: An agent-based model

We provide an agent-based model to explain the emergence of collective opinions not based on feedback between different opinions, but based on emotional interactions between agents. The driving variable is the emotional state of agents, characterized by their valence and their arousal. Both determine their emotional expression, from which collective emotional information is generated. This information feeds back on the dynamics of emotional states and of individual opinions in a non-linear manner. We derive the critical conditions for emotional interactions to obtain either consensus or polarization of opinions. Stochastic agent-based simulations and formal analyses of the model explain our results. Possible ways to validate the model are discussed.

physics.soc-ph↗

arXiv2019

Analyzing gender inequality through large-scale Facebook advertising data

Online social media are information resources that can have a transformative power in society. While the Web was envisioned as an equalizing force that allows everyone to access information, the digital divide prevents large amounts of people from being present online. Online social media in particular are prone to gender inequality, an important issue given the link between social media use and employment. Understanding gender inequality in social media is a challenging task due to the necessity of data sources that can provide large-scale measurements across multiple countries. Here we show how the Facebook Gender Divide (FGD), a metric based on aggregated statistics of more than 1.4 Billion users in 217 countries, explains various aspects of worldwide gender inequality. Our analysis shows that the FGD encodes gender equality indices in education, health, and economic opportunity. We find gender differences in network externalities that suggest that using social media has an added value for women. Furthermore, we find that low values of the FGD are associated with increases in economic gender equality. Our results suggest that online social networks, while suffering evident gender imbalance, may lower the barriers that women have to access informational resources and help to narrow the economic gender gap.

cs.CY↗

arXiv2018

Mixed Precision Training

Deep neural networks have enabled progress in a wide variety of applications. Growing the size of the neural network typically results in improved accuracy. As model sizes grow, the memory and compute requirements for training these models also increases. We introduce a technique to train deep neural networks using half precision floating point numbers. In our technique, weights, activations and gradients are stored in IEEE half-precision format. Half-precision floating numbers have limited numerical range compared to single-precision numbers. We propose two techniques to handle this loss of information. Firstly, we recommend maintaining a single-precision copy of the weights that accumulates the gradients after each optimizer step. This single-precision copy is rounded to half-precision format during training. Secondly, we propose scaling the loss appropriately to handle the loss of information with half-precision gradients. We demonstrate that this approach works for a wide variety of models including convolution neural networks, recurrent neural networks and generative adversarial networks. This technique works for large scale models with more than 100 million parameters trained on large datasets. Using this approach, we can reduce the memory consumption of deep learning models by nearly 2x. In future processors, we can also expect a significant computation speedup using half-precision hardware units.

cs.AI↗

arXiv2016

The Dynamics of Emotions in Online Interaction

We study the changes in emotional states induced by reading and participating in online discussions, empirically testing a computational model of online emotional interaction. Using principles of dynamical systems, we quantify changes in valence and arousal through subjective reports, as recorded in three independent studies including 207 participants (110 female). In the context of online discussions, the dynamics of valence and arousal are composed of two forces: an internal relaxation towards baseline values independent of the emotional charge of the discussion, and a driving force of emotional states that depends on the content of the discussion. The dynamics of valence show the existence of positive and negative tendencies, while arousal increases when reading emotional content regardless of its polarity. The tendency of participants to take part in the discussion increases with positive arousal. When participating in an online discussion, the content of participants' expression depends on their valence, and their arousal significantly decreases afterwards as a regulation mechanism. We illustrate how these results allow the design of agent-based models to reproduce and analyze emotions in online communities. Our work empirically validates the microdynamics of a model of online collective emotions, bridging online data analysis with research in the laboratory.

cs.HC↗

arXiv2016

The QWERTY effect on the web: How typing shapes the meaning of words in online human-computer interaction

The QWERTY effect postulates that the keyboard layout influences word meanings by linking positivity to the use of the right hand and negativity to the use of the left hand. For example, previous research has established that words with more right hand letters are rated more positively than words with more left hand letters by human subjects in small scale experiments. In this paper, we perform large scale investigations of the QWERTY effect on the web. Using data from eleven web platforms related to products, movies, books, and videos, we conduct observational tests whether a hand-meaning relationship can be found in decoding text on the web. Furthermore, we investigate whether encoding text on the web exhibits the QWERTY effect as well, by analyzing the relationship between the text of online reviews and their star ratings in four additional datasets. Overall, we find robust evidence for the QWERTY effect both at the point of text interpretation (decoding) and at the point of text creation (encoding). We also find under which conditions the effect might not hold. Our findings have implications for any algorithmic method aiming to evaluate the meaning of words on the web, including for example semantic or sentiment analysis, and show the existence of "dactilar onomatopoeias" that shape the dynamics of word-meaning associations. To the best of our knowledge, this is the first work to reveal the extent to which the QWERTY effect exists in large scale human-computer interaction on the web.

cs.HC↗

arXiv2016

Emotions, Demographics and Sociability in Twitter Interactions

The social connections people form online affect the quality of information they receive and their online experience. Although a host of socioeconomic and cognitive factors were implicated in the formation of offline social ties, few of them have been empirically validated, particularly in an online setting. In this study, we analyze a large corpus of geo-referenced messages, or tweets, posted by social media users from a major US metropolitan area. We linked these tweets to US Census data through their locations. This allowed us to measure emotions expressed in the tweets posted from an area, the structure of social connections, and also use that area's socioeconomic characteristics in analysis. %We extracted the structure of online social interactions from the people mentioned in tweets from that area. We find that at an aggregate level, places where social media users engage more deeply with less diverse social contacts are those where they express more negative emotions, like sadness and anger. Demographics also has an impact: these places have residents with lower household income and education levels. Conversely, places where people engage less frequently but with diverse contacts have happier, more positive messages posted from them and also have better educated, younger, more affluent residents. Results suggest that cognitive factors and offline characteristics affect the quality of online interactions. Our work highlights the value of linking social media data to traditional data sources, such as US Census, to drive novel analysis of online behavior.

cs.CY↗

arXiv2016

Women Through the Glass Ceiling: Gender Asymmetries in Wikipedia

Contributing to the writing of history has never been as easy as it is today thanks to Wikipedia, a community-created encyclopedia that aims to document the world's knowledge from a neutral point of view. Though everyone can participate it is well known that the editor community has a narrow diversity, with a majority of white male editors. While this participatory \emph{gender gap} has been studied extensively in the literature, this work sets out to \emph{assess potential gender inequalities in Wikipedia articles} along different dimensions: notability, topical focus, linguistic bias, structural properties, and meta-data presentation. We find that (i) women in Wikipedia are more notable than men, which we interpret as the outcome of a subtle glass ceiling effect; (ii) family-, gender-, and relationship-related topics are more present in biographies about women; (iii) linguistic bias manifests in Wikipedia since abstract terms tend to be used to describe positive aspects in the biographies of men and negative aspects in the biographies of women; and (iv) there are structural differences in terms of meta-data and hyperlinks, which have consequences for information-seeking activities. While some differences are expected, due to historical and social contexts, other differences are attributable to Wikipedia editors. The implications of such differences are discussed having Wikipedia contribution policies in mind. We hope that the present work will contribute to increased awareness about, first, gender issues in the content of Wikipedia, and second, the different levels on which gender biases can manifest on the Web.

cs.SI↗

arXiv2016

Who Watches (and Shares) What on YouTube? And When? Using Twitter to Understand YouTube Viewership

We combine user-centric Twitter data with video-centric YouTube data to analyze who watches and shares what on YouTube. Combination of two data sets, with 87k Twitter users, 5.6mln YouTube videos and 15mln video sharing events, allows rich analysis going beyond what could be obtained with either of the two data sets individually. For Twitter, we generate user features relating to activity, interests and demographics. For YouTube, we obtain video features for topic, popularity and polarization. These two feature sets are combined through sharing events for YouTube URLs on Twitter. This combination is done both in a user-, a video- and a sharing-event-centric manner. For the user-centric analysis, we show how Twitter user features correlate both with YouTube features and with sharing-related features. As two examples, we show urban users are quicker to share than rural users and for some notions of "influence" influential users on Twitter share videos with a higher number of views. For the video-centric analysis, we find a superlinear relation between initial Twitter shares and the final amounts of views, showing the correlated behavior of Twitter. On user impact, we find the total amount of followers of users that shared the video in the first week does not affect its final popularity. However, aggregated user retweet rates serve as a better predictor for YouTube video popularity. For the sharing-centric analysis, we reveal existence of correlated behavior concerning the time between video creation and sharing within certain timescales, showing the time onset for a coherent response, and the time limit after which collective responses are extremely unlikely. We show that response times depend on video category, revealing that Twitter sharing of a video is highly dependent on its content. To the best of our knowledge this is the first large-scale study combining YouTube and Twitter data.

cs.SI↗

arXiv2016

When the Filter Bubble Bursts: Collective Evaluation Dynamics in Online Communities

We analyze online collective evaluation processes through positive and negative votes in various social media. We find two modes of collective evaluations that stem from the existence of filter bubbles. Above a threshold of collective attention, negativity grows faster with positivity, as a sign of the burst of a filter bubble when information reaches beyond the local social context of a user. We analyze how collectively evaluated content can reach large social contexts and create polarization, showing that emotions expressed through text play a key role in collective evaluation processes.

cs.SI↗

arXiv2015

Social signals and algorithmic trading of Bitcoin

The availability of data on digital traces is growing to unprecedented sizes, but inferring actionable knowledge from large-scale data is far from being trivial. This is especially important for computational finance, where digital traces of human behavior offer a great potential to drive trading strategies. We contribute to this by providing a consistent approach that integrates various datasources in the design of algorithmic traders. This allows us to derive insights into the principles behind the profitability of our trading strategies. We illustrate our approach through the analysis of Bitcoin, a cryptocurrency known for its large price fluctuations. In our analysis, we include economic signals of volume and price of exchange for USD, adoption of the Bitcoin technology, and transaction volume of Bitcoin. We add social signals related to information search, word of mouth volume, emotional valence, and opinion polarization as expressed in tweets related to Bitcoin for more than 3 years. Our analysis reveals that increases in opinion polarization and exchange volume precede rising Bitcoin prices, and that emotional valence precedes opinion polarization and rising exchange volumes. We apply these insights to design algorithmic trading strategies for Bitcoin, reaching very high profits in less than a year. We verify this high profitability with robust statistical methods that take into account risk and trading costs, confirming the long-standing hypothesis that trading based social media sentiment has the potential to yield positive returns on investment.

cs.SI↗

arXiv2015

Geography of Emotion: Where in a City are People Happier?

Location-sharing services were built upon people's desire to share their activities and locations with others. By "checking-in" to a place, such as a restaurant, a park, gym, or train station, people disclose where they are, thereby providing valuable information about land use and utilization of services in urban areas. This information may, in turn, be used to design smarter, happier, more equitable cities. We use data from Foursquare location-sharing service to identify areas within a major US metropolitan area with many check-ins, i.e., areas that people like to use. We then use data from the Twitter microblogging platform to analyze the properties of these areas. Specifically, we have extracted a large corpus of geo-tagged messages, called tweets, from a major metropolitan area and linked them US Census data through their locations. This allows us to measure the sentiment expressed in tweets that are posted from a specific area, and also use that area's demographic properties in analysis. Our results reveal that areas with many check-ins are different from other areas within the metropolitan region. In particular, these areas have happier tweets, which also encourage people from other areas to commute longer distances to these places. These findings shed light on human mobility patterns, as well as how physical environment influences human emotions.

cs.CY↗

arXiv2015

Sentiment cascades in the 15M movement

Recent grassroots movements have suggested that online social networks might play a key role in their organization, as adherents have a fast, many-to-many, communication channel to help coordinate their mobilization. The structure and dynamics of the networks constructed from the digital traces of protesters have been analyzed to some extent recently. However, less effort has been devoted to the analysis of the semantic content of messages exchanged during the protest. Using the data obtained from a microblogging service during the brewing and active phases of the 15M movement in Spain, we perform the first large scale test of theories on collective emotions and social interaction in collective actions. Our findings show that activity and information cascades in the movement are larger in the presence of negative collective emotions and when users express themselves in terms related to social content. At the level of individual participants, our results show that their social integration in the movement, as measured through social network metrics, increases with their level of engagement and of expression of negativity. Our findings show that non-rational factors play a role in the formation and activity of social movements through online media, having important consequences for viral spreading.

cs.SI↗

arXiv2015

Ideological and Temporal Components of Network Polarization in Online Political Participatory Media

Political polarization is traditionally analyzed through the ideological stances of groups and parties, but it also has a behavioral component that manifests in the interactions between individuals. We present an empirical analysis of the digital traces of politicians in politnetz.ch, a Swiss online platform focused on political activity, in which politicians interact by creating support links, comments, and likes. We analyze network polarization as the level of intra- party cohesion with respect to inter-party connectivity, finding that supports show a very strongly polarized structure with respect to party alignment. The analysis of this multiplex network shows that each layer of interaction contains relevant information, where comment groups follow topics related to Swiss politics. Our analysis reveals that polarization in the layer of likes evolves in time, increasing close to the federal elections of 2011. Furthermore, we analyze the internal social network of each party through metrics related to hierarchical structures, information efficiency, and social resilience. Our results suggest that the online social structure of a party is related to its ideology, and reveal that the degree of connectivity across two parties increases when they are close in the ideological space of a multi-party system.

stat.AP↗