Record edition for the 13th conference on Network Science: almost 800 attendees and a long waiting list. Vittoria Colizza and Alain Barrat did a great job at organizing the event, that turned out to be a great success. Besides many excellent keynotes and two days of satellite meetings, the conference featured five parallel tracks with contributions on theory of complex systems, epidemics & spreading, biology, economics, and socio-technical systems. I have attended the Quantifying Success and the Machine Learning in Network Science satellites and, during the main event, I stuck to the Social Systems track. Here’s the summary of the week’s highlights, from the perspective of a Computer Scientist.
The brain’s wanderings
Network Wanderings (Danielle S. Bassett, Penn Engineering) In 2017, Danielle Bassett received the Lagrange prize (a top prize in complexity science) for her pioneering work in bringing together network science and neuroscience. This year, she is the recipient of the The Erdős-Rényi Prize, a prestigious award that is given every year by the Network Science Society to a scientist under-40 to recognize special achievements in the field of network science. In her keynote, she provided an overview of the remarkable work that her team has done in recent years to understand how the brain works (or fails to work properly under degenerative conditions) and how network science can contribute to crack longstanding questions about how to draw and interpret functional maps of the brain. But the most fascinating concept in her presentation is the one of Network Wanderings. Simply put, given a graph of concepts, the question is what is the most effective way to navigate this graph for a person to effectively learn those concepts. In her recent work, Prof. Bassett has shown that some network topologies can ease the learning process. She conducted experiments with participants exposed to sequences of visual stimula. For each stimulus, subjects were asked to respond by pressing a combination of one or two buttons on a keyboard. The sequences were produced based on a pre-determined graph of transitions between stimula. They found that pairs of sequences connected by highly embedded edges were reproduced more effectively than pairs across bridging edges, even when the node degree in the network was constant. This is a very powerful idea that applies not only to the brain but might apply to online learning, recommendations, and consumption of information. Congratulations, Danielle!
Misinformation, biases, inequality
Top 5 reasons why misinformation spreads — #1 will (not) shock you! (Fil Menczer Indiana University). Fil Menczer has studied the phenomenon of (mis)information spreading online for more than a decade. The research work he has led, and the tools his team has developed to build a more trustworthy Web are collected under the OSoMe project page. In his very engaging keynote, Fil has shown that the interplay between cognitive, social, and algorithmic biases makes us vulnerable to misinformation. Segregation is a key ingredient; he showed that, adding a low probability of homophily-driven unfriending to a social influence model, people segregate in groups and ideas stop spreading. He also showed that high-quality content emerges in scenarios where i) information load is low ii) the cost of accessing to new information is high, and iii) the content is consumed proportionally to its popularity. On top of those factors, the #1 reason why social networks effectively spread misinformation is that Bots can exploit very efficiently all these factors at once, thus being able to amplify the echo of fake news in the online ecosystem. When asked about the generality of Twitter-based studies and the actual impact of Twitter on the real-word societal dynamics, he provided two quite interesting points: many social platforms provide unmonitored, Twitter-like methods to spread misinformation (see Facebook pages) and Twitter is very often used by journalists to write news that are then broadcast to a much wider audience.
The power of networked counterpublics (Brooke Foucalt Welles, Northeastern University). If on one hand fake news pollute the social and political discourse on social media, on the other hand there are communities of online activism that provide critical mechanisms to overcome mainstream news narratives and empower marginalized groups. In her plenary talk, freshly tenured professor Brooke Foucalt Welles builds on Warner’s theory of counterpublics and talks about networked counterpublics in online communities. Starting from the tragic story of 18-year-old Michael Brown, shot and killed by a police officer in Ferguson, Missouri, Prof. Welles tells the tale of how people started tweeting about it. Contributing to hashtags of social awareness is often labeled as “lazy activism”, but she showed that networked counterpublics, organized by hashtag, can actually change drive social change by providing a collective perspective on facts that would not emerge otherwise. Her final recommendation is to “Keep tweeting!” because it makes the difference. Her book “#HashtagActivism: Race and Gender in America’s Networked Counterpublics” will be out in 2019.
Minorities in social networks (Claudia Wagner, GESIS). Studies about inequality and biases are trending the computational social science community, not least because their importance in explaining (and alleviating) much of the difficulties that minorities and underrepresented categories experience everyday. Claudia Wagner and her team at GESIS are working to measure how algorithmic biases and limited access to social capital impact the online visibility of women and minority groups. They are doing it looking at multiple settings and datasets. A new network model they developed proves how homophily and preferential attachment confines minorities in a condition of low visibility. An analysis of Wikipedia biographies reveals an impressive gap between men and women. Last, women have weaker and less embedded ties in networks of scientific collaborations. The road to gender equality is uphill…
Machine learning for (visual) knowledge equity (Miriam Redi, Wikimedia Foundation). Similar to the real word, the online world is full of barriers that are hard to overcome. The barriers arising from the differences in culture and language are those that Miriam Redi, together with her fellow teammates at the Wikimedia Foundation, are trying to overcome. She gave us an overview of a number of projects currently under development at Wikimedia. Two of them struck me the most. The first is about automatically learning when a statement needs a citation; this apparently narrow task hides a number of fundamental questions, for example understanding what can be considered “common knowledge” in a culture. The latter, still in its infancy, is to learn how to use new forms of visual language to enrich and complement knowledge and overcome language barriers. Given Dr. Redi’s expertise in measuring intangible properties of images and training cross-cultural models for visual sentiment concept ontologies, my bet is that the project has very good chances to turn into a success.
Structure and evolution of social networks
Attributed networks: Social Circles, summarization, comparison (Leman Akoglu, Carnegie Mellon). Given an attributed subgraph representing a community (a set of nodes annotated with labels), how do we evaluate its quality? Leman Akoglu introduces the concept of Normality for network communities. Normality takes into account the network structure and attributes together to quantify both internal consistency and external separability. This approach allows many boundary-edges as long as they come from external hubs (which are connected to many nodes, so they can be “discounted”).
Role discovery in networks (Tina Eliassi-Rad, Northeastern University). Community detection on graphs focuses on finding groups of highly connected nodes. Role discovery on graphs goes beyond grouping nodes based on their connectivity and embeddedness and finds groups of nodes that share similar graph topological structure. The idea makes a lot of sense but the first practical methods to implement it have been proposed only very recently. Tina Eliassi-Rad and her collaborators propose a new method to perform this task in a supervised fashion. The results look quite compelling.
Tie strength precedes social embeddedness (Esteban Moro, Carlos III University). Considering the humongous amount of work done inthe past to predict missing/future links in social networks, it is surprising that comparatively little work has been done on predicting ties strength. Esteban Moro took on this challenge starting from its underlying theoretical foundations. A few decades ago, Mark Granovetter theorized that tie strength and network embeddedness (i.e., the number of common neighbors between the two linked individuals) are deeply intertwined concepts. Prof. Moro analyzed traces from call data records to determine if tie strength comes before or after embeddedness. It turns out that, usually, strong links are created first and common neighbors grow around them in a later stage. He then introduced some results about link strength prediction, showing that classic features used in link prediction are way less effective for this task.
Network dynamics of innovation processes (Iacopo Iacopini, Queen Mary University London). Processes that operate on semantic spaces oscillate between states of exploration and exploitation. For example, when authors write books, they use a recurrent core set of words and, occasionally, they use new ones. This type of processes originate the so-called Heap’s Law, which describes the scaling of the number of distinct elements as a function of the total number of elements used in the process. Empirically, in most scenarios this is an exponential law with an exponent lower than 1. Iacopo Iacopini introduced a model for the emergence of innovations, in which cognitive processes are described as edge-reinforced random walks on the network of links among concepts, and an innovation corresponds to the first visit of a node. The model is simple and elegant and it is able to reproduce Heap’s laws for the emergence of new topics in scientific papers.
The human perception of social relationships (Luca Maria Aiello, Nokia Bell Labs). Graphs are convenient abstractions to model social interactions but every time we draw a social graph we make an implicit assumption about what links between nodes mean. The meaning of social links can be represented at different levels of abstraction: they can encode tie strength, polarity, sentiment, topics of conversation, and so on. In previous work, stemming from Peter Blau’s theory of social exchange, we have have proposed a method that pushes the representation of social links to a new level. Given a conversation graph (a social network with text messages on the edges) the method could estimate whether the relationship is about social support, knowledge exchange, or status exchange. Those three aspects are not exhaustive of all types of social interactions though. Recently, we have conducted a literature review and a crowdsourcing study to discover what are the fundamental relationship types that people and that sociologists has studied. It turns out that there are 10 and that they can help us to cast better predictions on how social networks evolve. Stay tuned for the upcoming paper.
Overseeding in social networks (Shankar Iyer, Facebook). Cracking the problem of influence maximization in real social systems remains the Holy Grail of all provides of social Web services. Shankar Iyer and Lada Adamic propose a more nuanced approach to the problem that makes an attempt to go beyond some of the oversimplified network models used in the past. Factors for the adoption of a new product or idea cannot be exhaustively captured by the network topology only. The need of supporting ties, the cold-start process of adjustment to the novelty, and the risk of permanent churn are contextual factors that might play a major role too. If those elements are not taken into account, massively seeding a new product in the social network might actually reduce its long-term adoption.
Social contagion of political extremism (Meysam Alizadeh, Indiana University). There are several models of network contagion. Complex contagion, structural diversity, and embeddedness are among the most established ones. They are partially conflicting as they rely on different structural indicators to model contagion. Meysam Alizadeh (together with Santo Fortunato, Ingmar Weber and Michael Macy) compared them in the setting of online political discussion, with a special attention to supporters of extremist parties. The result is that there’s no single model that is best suited for all settings. This finding suggests that contagion dynamics might be driven by different mechanisms, depending on the context.
The Metric Backbone of Contact Networks in Epidemic Spread Models (Luis M Rocha, Indiana University). A backbone graph is a subgraph that preserves the main structural properties of the original graph but number of edges is much smaller than in the original one. The disparity filter and the noise-corrected backbone are some of the backboning algorithms that have been proposed in the past. Luis Rocha and collaborators proposed a new approach that brings network backboning for weighted networks to its extreme. They have tested it on several dataset ranging from networks of axonal connections to graphs of social interactions from the sociopatterns project and they shown amazing results: the resulting reduced graphs preserve shortest paths but erase up to ~90% of the nodes. Details and code will be available here.
Understanding markovian network dynamics (Markus Strohmaier, Aachen University). In his opening keynote at the Machine Learning satellite, Prof. Markus Strohmaier gave an overview of an approach to discover classes of users that navigate graphs. The approach uses the concept of exceptional model mining and it applies to any type of navigation, from web browsing, business reviews and online music played. The most compelling use study presented is on human mobility from temporal Flickr data. The technique not only extracts user classes but it also uses the attributes of the nodes to provide meaningful explanations of those output classes. A great approach for everyone who’s working with mobility data.
Field theory for recurrent mobility (José Ramasco, IFISC). Two among the most established frameworks to model mobility are the gravity model and the radiation model. José Ramasco introduced a new perspective on the problem by using field theory formalism. In a nutshell, he models each agent that moves into space with a vector that expresses its average mobility direction and intensity. The approach naturally lends itself to find centers on mobility, also in polycentric cities. A very nice theoretical take on the theme of urban mobility.
Economical Segregation of Encounter Networks in cities (Esteban Moro, Carlos III University). There are many things you can do when you have fine grained mobility data and accurate estimates of socio-economic indicators at individual level. Esteban Moro and other folks at MIT use that type of data to estimate how much places of gathering mix people who have different economic conditions. The Gini index of hotspots vary greatly across the city. The paper is not yet public but will be coming out soon, looking forward to it.
Networks and Culture
Network analysis of perfumes (Vaiva Vasiliauskaite, Imperial College London). Remember the flavor network of food pairings? Vaiva Vasiliauskaite and Tim Evans worked on a similar concept but with perfumes. Using data of perfume descriptions, they built and analyze the network of similarities between perfumes that share common descriptors. Using additional data on perfume composition, they study which components are those that best enhance perfumes. They also produced a nice smell wheel that reminded me the categorization of urban smells from the Smelly Maps paper I co-authored. Well done, Vaiva. Very nice work!
Evolution of cultural elites: quantitative evidence from fashion (Peter Klimek, Robert Kreuzbauer, Stefan Thurner). The nature of social dynamics that underlies changes in art and fashion has been debated for long time. Using a large dataset of 8 million musical albums release over 60 years, Peter Klimek and colleagues found a strong evidence that fashion cycles in musical styles are driven by an evolutionary process that can be explained using signaling theory: individuals with high status who can afford sending strong signals to the majority of low-status individuals are trend-setters. Very fun examples from logos of metal bands. The paper will be out soon on the arXiv.
Social Influence in Music Listenership (Taha Yasseri, Oxford Internet Institute). Can live music events generate complex contagion in music streaming? Taha Yasseri found positive evidence by studying a large dataset from Last.fm. Using regression discontinuity analysis to estimate causal patterns, he observed that concert attendance has an effect on music listenership not only among attendees but also in their network of friends. One of the most entertaining talks of the conference, with a legendary selfie with the audience and a spicy remark to the infamous Reviewer#2 who rejected his submission.
Political coalitions and divisions on Twitter in 18 countries (Livia Teernstra, Justus Uitermark). The polarization and fragmentation of political discourse between political party members varies quite a lot across countries. Livia Teernstra and Justus Uitermark have found four types of network configuations. At the two extremes, Greece (with no dialogue between parties) and Finland (a virtuous example of dialogue and cooperation between parties).
Quantifying quality and success
Quantifying the effect of performance, success and social structures on user engagement (Rossano Schifanella, University of Turin). Web users have an intuitive understanding about how popularity of content is not always associated with quality. Rossano Schifanella takes a systematic approach to measure the distance between the two by using algorithmic tools that can measure visual beauty of pictures. He showed that modern deep learning pipelines can estimate very reliably the visual appeal of a picture as perceived by human observers. By applying the algorithm at scale on the Flickr photo sharing site, he found that popularity of pictures is broadly distributed whereas picture beauty is normally distributed. This originates a strong mismatch: the vast majority of high-quality pictures enjoy little or no attention.
The impact of social recommender systems (Luca Maria Aiello, Nokia Bell Labs). The struggle between popularity and quality can be the bane of recommendation systems too. In my invited talk at the Machine Learning satellite, I described how it might be convenient for companies to reinforce the rich-get-richer phenomenon of popular content as an easy strategy for monetization. But I also show that better alternatives are possible.
Simple contextualization strategies or content recommendations based on quality, rather than popularity, are effective methods that can trigger also positive side effects. In a recently published paper we used causal inference from longitudinal data to observe how social recommendations that suggest individuals with a high-quality production on Flickr might be effective in increasing the future content production. There’s a catch though: being recommended only people who produce exceptionally good content increases our fear of inadequacy and, as a consequence, boosts the probability of disengagement and churn.
Faculty hiring and the spread of scientific ideas (Aaron Clauset, Univeristy of Colorado at Boulder). Aaron Clauset took a deep dive into the question of whether science is a real meritocracy. He did that by investigating the role of the faculty hiring process in spreading of ideas across communities. His analysis provides strong evidence that the faculty hiring network is the main factor that determines which ideas survive and prosper. He also shown that research from prestigious universities spreads more quickly and broadly than that from less prestigious ones. Compared to prestige, quality plays a little role in spreading: his models indicate that renowned institutions are more effective in diffusing low-quality ideas than minor institutions are in spreading high-quality ones.
Quantifying success in science and arts (Roberta Sinatra, Central European University). Roberta Sinatra and her team have been studying the dynamics of success for quite some time and this year they have presented several studies that look at success in different disciplines and from multiple angles. Prof. Sinatra (whose newborn daugther is perhaps the youngest attendee in the history of NetSci) presented her latest work about interdisciplinarity in science. Looking at how many citations a paper receives from different disciplines, she was able to place papers in an “interdisciplinarity space”. Most of the papers and all those which won Nobel prizes are placed at the border of this space (i.e., they’re cited by only one or two disciplines). Future will tell if this trend is bound to change anytime soon.
Information is beautiful
The process of designing effective and beautiful visualizations (Miriah Meyer, Univeristy of Utah, and Paolo Ciuccarelli, Density Design Lab). Visualization tools have become essential for extracting meaning from the immense amount of data we are faced with today. Miriah Meyer and Paolo Ciuccarelli talked about their experience of interacting with stakeholders whose work got great benefit from the use of visualization. In her keynote, Miriah gave the example of MizBee, a tool she built to study the co-localization of genes on the same chromosome within/across individuals or species. Before that tool, domain experts used to look at their data by drawing co-occurrence matrices that turned out to be largely ineffective, if not misleading. Using the tool, the project PI was able to get to new scientific discoveries. When asked: “How long it would have taken to get there with no viz” he replied: “Honestly… I don’t think I would ever have gotten here.”
Paolo, in his inspiring dinner keynote, described a similar experience of working with network scientist stakeholders who badly wanted to produce an effective visualization a network structure only to realize after multiple iterations that the temporal processed over the network — and not the graph itself — was what they were interested in. The two main takeaways: 1) take a peek outside the set of tools you’re used to and don’t be afraid to shift visualization paradigm 2) The work of visualization experts is not only important to get appealing plots and interfaces but it is an integral part of the scientific method.
A lot of stuff! The event was huge and there was so much going on, including the great social events and celebrations. Here I’ve reported a selection of the presentation I saw but check out the program for more pointers.
Until next time…
The NetSci saga will continue next year in Chile and Vermont.
In the meantime, if you happen to be in UK in December, you should consider dropping by the Complex Network conference in Cambridge. Hope to see you there!