Clustering is a vital tool when handling data making it a central part of data science. By grouping similar objects together, it helps us find what we are looking for. I don’t go to a bakery to find a book. Clustering is part of a wider idea in science as we are always faced with thousands of potential or actual measurements but we need to focus on the few which are relevant to the process we are trying to understand. I do not need to know the nuclear properties of the constituents of a gas to understand its properties, while measuring temperature, pressure and volume do throw a lot of light on that problem. In whatever branch of science we are working in, we are always trying to reduce the dimensionality of our data, to use the language of statistics and data analysis.
Many of the techniques we use will need a measure of distance and it is most natural to call upon the everyday distance as defined by any ruler – formally the Euclidean distances d where for example d2 = x2 + y2 + z2 for the distance between the origin and a point at (x,y,z) in 3-dimensions.
However, what if time is present? Time is very different from space. Mathematically it leads to new types of geometry for space-times, Lorentzian rather than Euclidean. The simplest example is the Minkowski space-time used for studying special relativity. James Clough and I have been using Minkowski space as part of our study of networks which have a sense of time built into them - Directed Acyclic Graphs (see my blog on Time Constrained Networks for instance). Essentially these networks have a time associated with each vertex and then any edges present always point in one direction in time, say from the more recent vertex to an older one. Typically the time is a real physical time but for these types of network one can always construct an effective if artificial time coordinate.
There are many types of data with a directed acyclic graph structure. Citation networks are excellent examples and we will use them to illustrate our ideas in the rest of this article. Each node in a citation network is a document. The edges represent the entries in the bibliography of one document which always reference older documents - our arrow of time. We have worked with several different types of citation network: academic paper networks based on sections of the arXiv paper repository, US Supreme court judgements, and patents. My blog on citation network modelling gives some more background and how I think about citation networks in general.
Combining these two concepts James Clough and I have adapted a well known clustering method, MDS (Multidimensional scaling), so that it works for directed acyclic graphs (Clough and Evans 2016b). Traditional MDS is usually applied to data sets where you have a matrix of distances between each object. For a network, this would usually be the length of the shortest path between each node. MDS then assumes that these objects/nodes are embedded in a Euclidean space and suggests the best set of coordinates for the objects in that space. Clustering can then be performed by looking at which points are close together in this space. We found a way to take account of the fact that two papers on exactly the same topic can be published at the same time in different places. They are clearly ‘close’ together in any common sense definition of close yet there is no direct connection through their citation network. Our method will show that these papers are similar just from the pattern of their citations. Indeed the text could be fairly different (perhaps with two documents on networks, one uses the terms node, link, network while the second uses vertex, edge, graph for the same concepts) but the way these two documents are used by others later, or the way the to documents were based on the same material, indicates they are likely to be working on the same ideas.
Once you have the coordinates of each document in the citation network there are many other standard geometric tools you can use to do other jobs. For instance to recommend similar papers to one you are reading, you just look for other documents close in a geometric sense given the coordinates we have calculated. In the figure we show the top two hundred papers from the first decade of the hep-th part of the arXiv paper repository (this is dominated by string theory). The visualisation uses coordinates found using our Lorentzian MDS technique.
Our work with Minkowski space fits into broader programme of looking at networks in terms of the geometry of different types of space, what I call Netometry (Networks + Geometry, or perhaps Neteometry is better), as exemplified by Krioukov et al 2009. For instance, a good indication that a low dimensional Minkowski space might be a good representation of many citation networks came from our measurements of dimension (Clough and Evans 2016a).
I’ve really enjoyed reading my copy of Exploring Big Historical Data: The Historian’s Macroscope (Macroscope for short here) by Shawn Graham, Ian Milligan and Scott Weingart. As the authors suggest the book will be ideal for students or researchers from humanities asking if they can use big data ideas in their work. While history is the underlying context here, most of the questions and tools are relevant whenever you have text based data, large or small. For physical scientists, many of whom are not used to text data, Macroscope prompts you to ask all the right questions. So this is a book which can really cross the disciplines. Even if some readers are like me and they find some aspects of the book very familiar, they will still find some new stimulating ideas. Failing that, will be able to draw on the simplicity of the explanations in Macroscope for their own work. I know enough about text and network analysis to see the details of the methods were skipped over but enough of a broad overview was given for someone to start using the tools. PageRank and tf-idf (term frequency–inverse document frequency) are examples where that practical approach was followed. Humanities has lot of experience of working with texts and a physical scientist like myself can learn a lot from their experience. I have heard this piecemeal in talks and articles over the last ten years or so but I enjoyed having them reinforced in a coherent way in one place. I worry a bit that that the details in Macroscope of how to use one tool or another will rapidly date but on the other hand it means a novice has a real chance to be able to try these ideas out just from this book alone. It is also where the on line resources will come into their own. So I am already planning to recommend this text to my final year physics students tackling projects involving text. My students can handle the technical aspects without the book but even there they will find this book gives them a quick way in.
I can see that this book works as I picked up some of the simpler suggestions and used it on a pet project which is to look at the way that the staff in my department are related through their research interests. I want to see if any bottom-up structure of staff research I can produce from texts written by staff matches up to existing and proposed top-down structures of faculties – departments – research groups. I started using by using python to access to the Scopus api. I’m not sure you can call Elsevier’s pages on this api documentation and even stackoverflow struggled to help me but the blog Getting data from the Scopus API helped a lot. A hand collected list of Scopus author ids enabled me to collect all the abstracts from recent papers coauthored by each staff member. I used python libraries to cluster and display the data, following a couple of useful blogs on this process, and got some very acceptable results. However I then realised that I could use the text modelling discussed in the book on the data I had produced. Sure enough a quick and easy tool was suggested in Macroscope, one I didn’t know, Voyant Tools. I just needed a few more lines to my code in order to produce text files, initially one per staff member containing all their recent abstracts in one document. With the Macroscope book in one hand, I soon had a first set of topics, something easy to look at and consider. This showed me that words like Physical and American were often keywords, the second of these being quite surprising initially. However, a quick look at the documents with a text editor (a tool that is rightly never far away in Macroscope) revealed that many abstracts start with a copyright statement such as “2015 American Physical Society”, something I might want to remove as this project progresses. I am very wary of such data clustering in general but with proper thought, with checks and balances of the sort which are a key part of Macroscope, you can extract useful information which was otherwise hidden.
So even for someone like me who has used or knows about sophisticated tools in this area and is (over) confident that they can use such tools, the technical side of Macroscope should provide a very useful short cut despite my initial uncertainty. Beyond that I found that having the basic issues and ideas behind these approaches reinforced and well laid out was really helpful for me. For someone starting out, like some of my own physical science masters and bachelors students working on some of my social science projects, they will find this book invaluable. A blog or intro document will often show you how to run a tool but they will not always emphasise the wider principles and context for such studies, something you get with Macroscope.
I should make clear that I do have some formal connections with this book, one of my contributions to the pool of academic goodwill. I suggested the general topic of digital humanities and Shawn Graham in particular as a potential author at an annual meeting of the physics and maths advisory committee for ICP (Imperial College Press). For free sandwiches we pass on ideas for topics or book projects to the publisher. I also commented on the formal proposal from all three authors to ICP, for which I often get a free book. My copy of Macroscape was obtained for reviewing a recent book proposal for ICP. Beyond this I get no remuneration from ICP. It is nice to see a topic and an author I highlighted to come together in a real book but the idea is the easy bit and hardly novel in this case. Taking up the idea and making it into a practical publishing project is down to Alice Oven and her ICP colleagues, and to the authors Shawn Graham, Ian Mulligan and Scott Weingart. That’s particularly true here as the book was produced in an unusual open source way and ICP had the guts to go along with the authors to try this different type of approach to publishing.
Exploring Big Historical Data: The Historian’s Macroscope
Shawn Graham (Carleton University, Canada),
Ian Milligan (University of Waterloo, Canada),
Scott Weingart (Indiana University, USA)
ISBN: 978-1-78326-608-1 (hardback)
ISBN: 978-1-78326-637-1 (paperback)
When one document refers to another in it’s text, this is called a citation. The pattern of these citations is most naturally represented as a network where the nodes are the documents and the links are the citations between the documents. When documents were physical documents, printed or written on paper, then these citations must (almost always) always point back in time to older documents. This arrow of time is imprinted on these citation networks and it leads to interesting mathematical properties.
One of the most interesting features of citations is that they have been carefully curated, sometimes for hundreds of years. The data I use on the US supreme court judgments goes back to the founding of the USA. So citation data is one of the oldest and continuous ‘big data’ sets to study.
The reason why records of citations have been maintained so carefully is that they record the process of innovation, be it in patents, law or in academic study. When you try to claim a patent you must by law indicate the prior art, earlier patents with relevant (but presumably less advanced) ideas. When a judge makes a judgement in a case in the USA, they draw on earlier cases which have interpreted, and so created, the law needed to come to a conclusion in the current case. And of course, academics can’t discuss the whole of existing science when explaining their own new ideas so they have to refer back to papers where previous researchers have set out a key idea needed in the current work. Citations are therefore a vital part of knowledge transfer, new ideas build on all the work done in earlier judgments, previous patents or older papers. That is why citations have been so carefully recorded. The network formed by documents and their citations show whose giant shoulders we are standing on when innovations are made, to paraphrase Newton.
From a theoretical point of view there are many interesting features in these networks. If you follow citations from one document to another to another and so on, at each step you will always reach an older paper. So you can never come back to the starting point and such paths. There are no cycles in a citation network (it is an example of a directed acyclic graph). If you look at the number of citations each document gets, how many newer documents refer back to one document, then they follow a fat-tailed distribution – a few documents have most of these citations, while most documents have very few citations each. Derek de Solla Price’s 1965 paper for an early discussion of this feature. Moreover, if you look at documents in the same year, you get roughly the same shape for the number of documents with a given number of citations (see Radicchi et al 2008, Evans et al 2012), at least for well cited documents. Since these networks are of such great interest,many other features have been noted too.
One way for a theorist to understand what is happening is to build a model from a few simple rules which captures as many of the features as possible. One of the first was that of Derek de Solla Price (1965) whose theory of “cumulative advantage” suggested that as new documents were created they would cite existing papers in proportion to the number of citations they already had, that is the richer get richer. This follows a principle used in many other models of fat-tails in other data, and indeed was later rediscovered in the context of the number of links to modern web pages – Barabási and Albert’s preferential attachment (1999). One trouble with this simple model is that the oldest documents are always the ‘rich’ ones with the most links. In reality, each year of publication there are a few documents with many citations (relative to the average number for that year) and most have very few. The Price model does not give this as all documents published in the same year will have roughly the same number of citations. To address this problem we (Sophia Goldberg, Hannah Anthony and myself,, Goldberg et al 2015) searched for a simple model which reproduced this behaviour – fat tails for the citation data of papers published in one field and one year (Goldberg et al, 2015).
The simplest model we found works as follows. At each step we add a new document, representing the evolution in time of our citation network.
A new document first looks at recently published documents, as it is well known that citations tend to favour more recent documents. What we mean by recent is set by one of our parameters, a time scale τ.
We choose these recent documents partly at random (fraction (1-p)) and partly with cumulative attachment (fraction p) in which we pick recent papers to cite in proportion to the number of their current citations. This choice of papers is not realistic since it requires the authors to be able to choose from all recent documents while in reality authors only have a limited knowledge. However this stage is meant to capture, statistically at least, the way authors learn about recent developments: a recommendation from a colleague, a talk at a conference, scanning new editions of certain journals and so forth. Sometimes it will be essentially random, sometimes this first choice will reflect the attention papers have or will receive.
Once we have chosen these primary documents to cite, our new document then looks at the references within these primary documents.
Each paper cited in the primary paper is then cited, copied, by the new document with probability q, the third and last parameter of the model.
This ‘copying’ process is known to be a way of getting cumulative attachment with only local knowledge of the network (see for instance my paper with Jari Saramäki (Evans & Saramäki 2005) and references therein). That is you only need to read the papers you have already cited, already found, to find these secondary documents to reference. There is no need for the model to know about the whole network at this point, reflecting the limited knowledge of actual authors.
It was only when we added the last copying process that we found our model reproduced the fat-tails seen within the citations to documents published in the same year. Nothing else we tried gave a few well cited papers in every year. Comparing with one data set, taken from the hep-th section of the arXiv repository, we found that the best values for our parameters led to a typical paper in our model of hep-th made up as follows:
- Two primary papers chosen at random from recent papers.
- Two primary papers chosen in proportion to the number of their citations from recent papers.
- Eight secondary papers chosen by copying a reference from one of the first four primary papers.
This may seem like a very high level of papers being copied from the primary ones – on average we found 70% of papers were secondary citations, papers already cited in other papers being cited. One has to ask if the more recent paper, the primary one, contained all the information from the earlier ones as well as innovations being built on in the current paper. Did the new documents really derive useful information from the secondary papers cited? Often you see old ‘classic papers’ gathering citations as they are name checked in the introduction to a new paper. Not clear if the classic paper was even read while performing the current research. This feeling that some papers gain attention and acquire citations that does not reflect any direct influence on the current work is supported by at least two other studies. One was a study by Simkin and Roychowdhury (2005) of the way errors in the bibliographies of papers are copied in later papers. They suggest that this meant 80% of citations came from such copying of references. In another approach, James Clough, Tamar Loach, Jamie Gollings and myself (Clough et al, 2015) exploited the special properties of citation networks and this also suggested that 70%-80% of links were unnecessary for the logical structure of academic citation networks.
Of course constructing simple models will never capture the whole story. Models are, though, a good way to see if we have understood the key principles underlying a system.
- Barabási, A.-L., Albert, R. (1999). Emergence of scaling in random networks. Science, 286, 173.
- Clough, J.R., Gollings, J., Loach, T.V., Evans, T.S. (2015). Transitive reduction of citation networks. J. Complex Networks, 3, 189-203 [doi: 10.1093/comnet/cnu039, arXiv:1310.8224].
- Clough, J.R., Evans, T.S. (2014). What is the dimension of citation space? arXiv:1408.1274.
- Evans, T.S., Saramäki, J.P. (2005). Scale Free Networks from Self-Organisation. Phys.Rev.E, 72, 026138
[doi: 10.1103/PhysRevE.72.026138 , arXiv:1408.2970]
- Goldberg, S., Anthony, H., Evans, T.S. (2015). Modelling Citation Networks. Scientometrics, 105, 1577-1604
[doi: 10.1007/s11192-015-1737-9 , arXiv:1408.2970 ]
- Simkin, M.V., Roychowdhury, V.P. (2005). Stochastic modeling of citation slips. Scientometrics, 62, 367-384
- Price, D.J.d.S. (1965). The scientific foundations of science policy. Nature, 206, 233-238.
- Radicchi, F., Fortunato, S., Castellano, C. (2008). Universality of citation distributions: Toward an objective measure of scientific impact. PNAS 105, 17268-17272.
One of the key features of complex networks is that they capture interactions which have no limitations. In most electronic systems, be they Facebook, emails or web pages, we can make connections across the world with little if any cost.
However what if there are constraints on the links made in a network? Surely we should change the way we study networks if space, time or some other constraint is having a significant effect on the formation or use of the network. This has been a major interest of mine over the last few years. Space is one obvious limitation as in some cases long distance are less likely to be made. There has been a lot of work in this area over many decades but I will leave this constraint for another blog.
It is only more recently that the role of time in networks has began to receive more attention. A lot of this recent interest in how to deal with networks where the connections, are made at one time. That is because most communication networks, emails, phone calls and so forth, are of this type. The recent review by Holmes and Saramäki (2012) is such temporal edge networks.
Yet networks are made of two parts: vertices and edges. My recent work has focussed on the case where it is the vertices, not the edges, which are created at a definite time. In such temporal vertex networks, causality forces the interactions between nodes to always point in one direction. For example consider a citation network formed by academic papers. The nodes in our network are the academic papers and the links are formed by their bibliographies. So if paper A refers to another paper B then we can be (almost) sure that A was written after B. Information can therefore flow only from B to A. In fact any set of documents can only refer to older ones such networks are common. In law, judges refer to previous judgments to support their arguments. When registering a patent, prior art needs to be cited, that is other previously granted work which may have some relevance to the current claim.
The same types of structure occur in several other situations. Any situation where there is a logical flow has the same causal structure. If we have a project where the nodes represent individual tasks then an edge from task S to task T could represent the fact that task T requires task S to have been completed before task T is started. This has been the focus of work on temporal vertex networks in computer science. The logical flow of a mathematical argument or an excel spreadsheet show the same properties. These networks define what is called a partially ordered set or poset and it is under this title that you find relevant work coming from mathematicians. A final example comes from the Causal Sets approach to quantum gravity (see Dowker 2006 for a review). Here space-time is discrete not continuous, and these discrete points are the nodes of the network. The nodes are connected by edges only if they are causally connected and causality again gives these a common direction.
All of these temporal vertex networks have a key property. That is they contain no loops if you always follow the direction on the edges. You can not go backwards in time. Thus the traditional name for such a network is a directed acyclic networks or DAG for short.
So the question is how can we adapt traditional network measures to deal with the fact that these networks, DAGs, are constrained by causality? Are there new measures we should employ which give more insights to such networks?
I’ve been looking at these problems with several students (undergraduates in their final year projects and some MSc students), one of whom, James Clough, is now working for his PhD on this topic.
Paths in networks are always important. However one feature of a DAG we have been exploiting is that if we always follow the direction of the arrows, the direction of time, then not all nodes are connected. If we like we could add edges whenever there is such a path connected a late node to an earlier one, a process known as transitive completion. On the other hand we could remove as many edges as we can while leaving the causal relationships intact, a process known as transitive reduction. That is, if there is a path between two nodes in the network before transitive reduction, then there will still be a link afterwards.
What we have done (in Transitive Reduction of Citation Networks) is look at how real data from citation networks behaves after transitive reduction. What we find is that different types of citation network behave very differently. The citation network formed from academic papers taken from the arXiv repositoryand the network of US Supreme Court judgments both show that about 80% of the edges are not needed to retain all the causal relationships. On the other hand the patents network shows the opposite behaviour with all but 15% of edges being essential. The edges removed tend to be the citation to older papers. One interpretation is that academics and and judges may be citing well known early papers and judgments though their current work is only directly related to more recent documents. Perhaps some of these citations do not indicate the early work was needed but reflect other motivations, such as simple copying of popular papers or review in the field which at best only have general relevance. For academic papers this interpretation is supported by the work of Simkins and Roychowdhury In this sense unnecessarily.
The number of citations to a document after transitive reduction certainly gives us a different view of the importance of different documents. For instance paper hep-th/9802109 on the arXiv (Gauge Theory Correlators from Non-Critical String Theory by Gubsner et al.) was cited by 1641 papers in the network, but only three citations remained after TR! On the other hand, paper hep-th/9905111 (Large N Field Theories, String Theory and Gravity by Aharony et al.) has also large number of citations in the raw data, 806, yet after transitive reduction it has 77, so retaining far more of its original citations. Perhaps information in the second paper was used more diversely.
We can find similar examples in the US Supreme Court citation network. The case Schneider vs. New Jersey (1939)’ has 144 citations in the original data but this drops to just just one after transitive reduction. Stromberg vs. California (1931) also falls from 132 citations to just one. Conversely, the case Heller vs. New York (1973) only shows a slight fall after transitive reduction, falling from from 68 to 48 citations and has the most citations in our reduced network. The second most cited case after transitive reduction is Hamling vs. United States, which drops from 68 to 38 citations. Wikipedia lists hundreds of Supreme Court cases but the last two are not famous enough to make the Wikipedia list. Our analysis suggests they may have more importance than a simple citation count would suggest. At the very least it might be be worth checking out documents that are highly cited in the essential.
Another way to look at citation networks is to see if we can define a dimension to the network. That is we can try to quantify how much variation there is in the citation process. A low dimension means that there are few directions , few distinct themes relevant for citation in a document. A high dimension indicates that there is a wide range of relevant but distinct directions from which a document will draw on for inspiration. What James Clough and I found (in What is the dimension of citation space?) is that we were often able to assign an interesting value for the dimension of our citation data. For academic papers, we found that different fields of research have different dimensions. For papers in the hep-th arXiv section (largely string theory) we found a low dimension of around 2 while for theoretical papers closely linked to particle physics experiments (hep-ph section) we found more variation as indicated by a higher dimension of 3. The quant-ph also around 3 while the astro-ph section had a slightly higher dimension of around 3.5. So clearly despite similarities in the main data using standard measures, our time-aware dimension measures show clear differences in the citation behaviour of different areas. String theory in particular seems to be a tightly knit collection of work with each work largely dependent on all the other work, few independent directions can be pursued. The US Supreme Court judgments were more complicated. Small samples (usually from modern judgments) showed a dimension of around 2.5 to 3 but larger samples, typically ranging from modern to the earliest judgments, had lower dimensions, closer to 2. We interpreted this as reflecting the way that there were few early judgments compared to the number produced to day. So that the further back we traced in time to find the influence of judgments on recent ones, the smaller the variation. Perhaps that is not so surprising and we might expect a similar shape if we could follow scientific papers back to the 18th century! patents on the other hand showed a much higher dimension though again these were involved.
It is clear from just the few studies we have made that time makes a crucial difference to the structure of a network. We have tried a few new measures adapted to take account of time and in doing so we have thrown up some intriguing features in real data. There is surely much more to find when networks are embedded in time.
Dowker, F. Causal sets as discrete spacetime, 2006. Contemporary Physics, 47, 1-9
Holme, P. & Saramäki, J. 2012. Temporal Networks Physics Reports, 519, 97-125
Simkin M.V. and Roychowdhury V.P., 2003. Read before you cite! Complex Systems 14 269-274
You do not need to know the detailed properties of every small part making up a gas, it turns out the bulk properties of a gas can be derived from very general principles. In the same way when looking at Facebook data, we might be able to identify groups of people who behave in a similar way. Searching for these groups or clusters in data is central in many areas of physical and social science. It is often easier to understand the behaviour of a large system by looking at these clusters, which are much fewer in number.
In terms of networks, the clustering is based on the structure (topology) of the network and the groups found are called Communities. In this case we might expect a coherent group to be one which has more links between members of the group than it ha to nodes outside the group in other clusters. I have done some work on what is called Community Detection, particularly in methods which assign nodes to the membership of several clusters (e.g. my line graph and clique graph papers referenced below). After all, my social connections are likely to show that I am part of several groups: work colleagues, family relationships, connections made through hobbies or sports.
For some time I have been very wary about the meaning of the clusters found with such methods and particular about claims of one method being able to find “better” communities than another. A recent paper prompted me to think about this again. In Community detection in networks: structural clusters versus ground truth, Hric, Darst, and Fortunato from Aalto University in Finland (a big centre for networks research) asked if the network methods were finding different sorts of clusters from those found using other aspects of the data. Typically when testing a community detection method, one sets up artificial networks in which each node is assigned to one community. The edges between nodes are then assigned at random but with a preference for edges to be between nodes from the same community. I can do all the tests I like on artificial data but I am always worried that this approach has introduced some hidden bias. Perhaps we end up choosing the methods that ‘work’ on artificial data but which are perhaps not so good on real messy data? It all comes down to the fact that we have mathematical ways to quantify the difference between community assignments but defining what we mean by “the best” clustering is impossible. Even with artificial networks, the “ground truth” is not generally an absolute truth. Typically the “truth” are input parameters and the actual network generated is partly random. So while the resulting artificial network is correlated with the ground truth it is not designed to be a perfect match. So in this case the “actual truth” will, in almost most cases, be different from the ground truth. [Note added 5th January 2016. Another more recent paper which includes some expert evaluation of communities found as well as a comparison of many different methods is Šubelj, van Eck and Waltman 2015 Clustering scientific publications based on citation relations: A systematic comparison of different methods.]
I also worry about what we do when we run network community detection methods on large real data sets where there is no simple ground truth. When I have done this, I can find a variety of possible answers for communities in the data. Many look reasonable but none correlate perfectly with each other or with what I know from other sources. This leaves me wondering if the automatic methods are finding one truth and my other information gives another. Alternatively the automatic methods might be rubbish, good on artificial cases, not so good in reality. There is no simple way of telling.
In any case do real networks have a “ground truth”? Quite often people have data from other sources about real networks and they use this to construct a “ground truth”. The test is then to see if automatic methods can find this ground truth. However what if the other data is wrong? People don’t always tell the truth, they can deliberately mislead or they can misunderstand the problem. Children surveyed about their friendships may tell you who they’d like to be friends with (the most popular person in the class) and not who they actually spend time with.
Take the famous Zachary karate club data set used by many (including myself) as a simple test. This is a network of members of a karate club that split in two during the sociologist’s study. Let us accept that the professionalism of Zachary has produced data that is a true reflection of the situation despite the difficulty of measuring associations in social science. If you look at the published paper it actually gives two truths. One is based on which of two factions the members actually joined, and one based on an automatic community detection method. I suspect most people are using the latter as the ground truth (unwittingly) when testing their work. Perhaps this is a further example supporting the claim that academics only read 20% of their references. Worse the data given in the published karate club paper is not consistent – the unweighted adjacency matrix is not symmetric. So which truth was used for all those papers using the Karate club network?
Another example comes from some work I did on overlapping community methods. Like many other people I downloaded a standard data set from Mark Newman’s web site, an extremely useful resource. The American College Football data was created by Girvan and Newman (in Community structure in social and biological networks) and represents the games played between American College Football teams in one season. Also provided are the conference membership of each team. Teams play more games against teams from the their own conference than from any one other conference. In fact this data is so well clustered that surely no method should get anything wrong beyond a few independent teams as my visualisations here illustrate (taken from my clique based clustering paper). So I looked at the “mistakes” made by my method. After about two afternoons of wading through interminable web sites of stats on American College football and Wikipedia pages on the College Conference system, I realised that in fact most of the “mistakes” were not from the automatic community detection but lay in “the ground truth”, that is in the conferences assigned to teams in the data file. It turns out that the assignments in the original football.gml file are for the 2001 season while the file records information about the games played for the 2000 season. For instance the Big West conference existed for football till 2000 while the Sun Belt conference was only started in 2001. There were 11 conferences and 5 independents in 2001 but 10 conferences and 8 independents in 2000. Care is needed as American College athletic conferences cover many sports, with some sports joining or dropped from any one conference time to time. Teams can also switch conferences too. In fact around 10% of the college teams playing American football at the top level changed conferences around 2000-2001. (Note added 5th January 2016. These errors were also noted in Gschwind et al 2015, Social Network Analysis and Community Detection by Decomposing a Graph into Relaxed Cliques, only the second paper I’ve seen which does this independent from my discussion).
So often the “ground truth” is just another truth not some absolute truth! The errors in the Zachary Karate club and American College Football data do not matter in one sense as they still provide valid and valuable tests for methods. The conclusions in the hundreds of papers using these data sets and which use these questionable ground truths would not change. Indeed it highlights one role for automatic methods. You can see that where Girvan and Newman’s methods get the “wrong” answer in their original paper (Community structure in social and biological networks) they are in fact highlighting problems with their conference data. Validation of data is a very useful if boring job. A final question will always be if there is a single truth. For instance I am in the theoretical physics group of the physics department of the Faculty of Natural Sciences at Imperial College London. That top-down hierarchical truth is important when allocating desks and teaching. However another truth would emerge if you studied my research relationships. Those are with staff and students based in other physics research groups and with colleagues from other departments and even other faculties.
So I was really pleased to see that Community detection in networks: structural clusters versus ground truth were questioning the meaning of truth in community detection from a quantitative point of view. Clustering of data, finding communities in data is of tremendous value commercially and for research, but there is still a lot more work to do before we understand these different truths.
M. Girvan, M.E.J. Newman, Community structure in social and biological networks, PNAS (2002) 99, 7821-782
W. Zachary, Information-Flow Model For Conflict And Fission In Small-Groups Journal Of Anthropological Research, (1977) 33 452–473
D. Hric, R.K. Darst,and S.Fortunato, Community detection in networks: structural clusters versus ground truth, arXiv:1406.0146
T.S. Evans, and R. Lambiotte, Line Graphs, Link Partitions and Overlapping Communities Phys.Rev.E, 2009, 80, 016105 [arXiv:0903.2181].
T.S. Evans, Clique Graphs and Overlapping Communities J. Stat. Mech. (2010) P12037 http://dx.doi.org/10.1088/1742-5468/2010/12/P12037
L. Šubelj, N.J. van Eck, L. Waltman, Clustering scientific publications based on citation relations: A systematic comparison of different methods, arXiv:1512.09023 .