Here I will list my ten favorite papers of the 2010s related to my research. It’s not an ordered list and it will not be too serious, don’t hate me if your paper is not on the list. Here we go:
This will be about marmot (cute!) social networks, but first a rant: Network science* has primarily been about method development. I get the impression that purely applied netsci papers are looked down upon. If science is about understand the world around us, we just need methods that are good enough, but applied to relevant data in a good way. Incremental method development can never be more interesting than a paper advancing the theory using new data. So here is a great applied netsci (or SNA if you wish) paper. The authors study a great data set of the social interactions of marmots. They have both positive and negative ties, which probably rings your social-balance bell. Indeed, the authors find a weak pattern of social balance, meaning that marmot friends share enemies. Transitivity (an overabundance of positive triangles) is an even stronger signal. Marmots are thus not only cute, they are friendly as well (at least to each other).
* Science by people you regularly come across at conferences like NetSci or Complenet.
This work builds on a combination of three separate but equally deep connections between physics and computer science: 1) A mapping between the stochastic blockmodel and the Potts model of statistical physics, made possible by the recent developments in inverse problems in statistical physics. 2) The insight that problems in computer science can exhibit phase transitions between regimes of hard and easy instances. 3) The belief-propagation algorithm and its connection to the cavity method of statistical physics.
It turns out that, in stochastic blockmodels, the hard instances of the inference problem (i.e. when using stochastic blockmodels for community detection), are exactly the cases when communities are ill-defined. More interestingly, however, is that the transition from well- to ill-defined communities is not gradual (in the large-N limit), but happens as a phase transition. It is thus meaningful to talk about a detectability limit. The authors calculate this limit exactly for the case of communities of equal average degree and study it numerically in other cases.
This paper really set the tone of a new decade. It is usually said that the advantage of publish on a new topic early is that one does not need to be too careful, but this is a paper both with groundbreaking ideas and rigor. Another hallmark of greatness is that it took a long time for me to appreciate its greatness. For a while I thought it was just a contrived way of modeling networks. I think I got my epiphany by a talk by Ginestra Bianconi. By now most of you know the idea of the paper: Embedding a network in space is a great way of summarizing the network structure. Everyone that ever plotted a graph by Pajek or Gephi knows that. It is only that Euclidean geometry is not the best for empirical networks, much because of the heterogeneous nature of real-world networks made famous by Barabási & al. a decade earlier.
The ‘10s is arguably the decade of data science. Remember when it started “data science” and “big data” were not even in the dictionary. In this era, papers presenting data sets should thus be in the limelight, and this is probably my favorite such. It features a catalogue of animal social networks that is very useful for my small-NetSci line of research. (Some great science of mine based on it coming up in the next decade.) Some networks are actually the smallest possible, i.e. empty, which is a bit profound—if you measure a social network, but there is no interaction, is it useless? Of course not (since it says something about the system), but it’s also the last example that’d come to my mind when I’m teaching network science. (This also brings Harary & Read’s tongue-in-cheek “Is the null-graph a pointless concept?” to mind.)
If you liked the book (TED talk, etc.), you’ll love the paper! This is kinda the AJS version of Everything is Obvious, but being an academic paper argued is a more precise and academic way (thus an even better read).
I think the fallacy of building scientific arguments on common sense is even more general than what this paper discusses. Social and behavioral science is the human endeavor to understand the world of people. Of course the primary ideas in building theories must be allowed to come from one’s own experience. But whether they generalize to a publishable result depends on if they could be supported by data. (I also note that this is mistake people coming from formal or natural science (like myself) to study social topics (somewhat ironically) tend to make even more than those with a formal training in social science.)
If you wish, network science is all about indirect connections. Everything of interest could be seen as a question of how nodes influence other nodes through intermediate nodes. These connections form paths, so it is natural to see paths as the key to understand networks. This is the gist of Schoch and Brandes’s paper that goes on to show that many centrality measures can be unified by so called path algebras. The path algebras can break down centrality ranking into classes such that no vertex in one class is ranked above a vertex in the other class by any of these centrality measures. S & B go on to show that previous axiomatic approaches to centrality measures can be simplified much by path algebras.