Michele Coscia

When you’re studying complex systems, one of the most important questions you might have is: how will this system evolve in the future? If you’re modeling your system as a network — as I like to do in my spare time — this boils down to predicting the arrival of new nodes and links. This is the realm of link prediction. In this post, I’ll describe one advancement in the field that I developed with fellow NERD Michael Szell in the paper “Multiplex Graph Association Rules for Link Prediction“, to appear next year at the ICWSM conference.

There are many ways to predict new links in a network, but most of these methods have a disadvantage: they can only give you a score for potential future connections between two nodes that are already in the network when you observe it. In other words, they cannot predict new incoming nodes. But with a technique called “graph association rules”, used by the GERM algorithm published in 2009, we can predict new nodes. How is that possible?

In simple terms, a “graph association rule” is a rule that tells you: every time you see in your network a pattern A, it will turn into pattern B, with a certain degree of confidence. The rule is extracted by counting how many times patterns A and B appear. For instance, in the image below, if pattern A (the triangle) appears 9 times and pattern B (the triangle with a dangling node) appears 6 times, the confidence of the rule is 2/3. 66% of the time, a triangle has attracted a dangling node. Note that pattern B must include pattern A, otherwise it’s difficult to hypothesize that A evolved into B.

GERM has a problem of its own, which Michael and I set out to solve: it can predict incoming nodes and links, but it cannot distinguish between different link types. In other words, every predicted link is the same to GERM. However, many real world networks have link types: nodes can connect in different ways. For instance, on social media, you connect to the people you know in different ways: via Facebook, Twitter, Linkedin, etc.

You’d model such system with a multiplex network, which allows for link types. If you have a multiplex network, you need multiplex graph association rules for link prediction. Which is exactly the title of our paper! What a crazy coincidence!

In the paper we re-purpose Moss, a graph pattern miner that can extract multiplex patterns, to build such rules. We created a pre- and post-processor of Moss that can construct the rules based on the patterns it finds. Now we can give colors to the links that are featured in our rules, as the figure below shows. This is a generalization of the signed link predictor I already wrote about a long time ago (the second ever post on this blog. I feel old).

Doing so isn’t painless though. We made sacrifices. For instance, our rule extractor doesn’t really understand the passage of time. It knows that the input network is in the past and spits out the rules to predict its evolution, but it doesn’t know how long a rule will take to complete. Unlike GERM, which can tell you that a rule will take n timesteps to complete.

This downside is minor though. Our link predictor performs well, as witnessed by the ROC curves below (our method in red). The comparisons are other multiplex link predictors. Not only are they worse at predicting links, but they have the added disadvantage of being unable to predict the arrival of new nodes. They also have issues with memory consumption, because they generate a score for each pair of nodes that is not connected in the training data — which, for sparse networks, is a lot of scores. Our predictor, instead, only gives scores to the links that are valid consequences of the rules that we found, usually way fewer than all unconnected node pairs.

If you want to play with our link predictor, you can do so by downloading the code I made public for the replication of the paper’s results. The code is very academic — meaning: badly written, unreasonably fragile, and horribly inefficient. I have in the works an extension with more efficient and robust code, and a generalization from multiplex to fully multilayer networks. Stay tuned!

Continue Reading

If there is a thing that I love more than looking at silly pictures on the Interwebz for work is to watch rugby for work. I love rugby: in my opinion it is the most beautiful team sport out there. It tingles my network senses: 15 men on the field have to coordinate like a single organism to achieve their goal — crossing the goal line with the ball by passing it backwards instead of forward. When Optasports made available some data collected during 18 rugby matches I felt I could not miss the opportunity for some hardcore network nerding on them. The way teams weave their collaboration networks during a match must have some relationship with their performance, and I was going to find out what this relationship might be.

For my quest I teamed up with Luca Pappalardo and Paolo Cintia, two friends of mine who are making an impact on network and big data sports analytics, both in soccer and in cycling. The result was “The Haka Network: Evaluating Rugby Team Performance with Dynamic Graph Analysis“, a paper recently presented at the DyNo workshop in San Francisco. Our questions were:

1. Is there a relationship between the topology of the network of passes and the success of the team?
2. Is there a relationship between disruptions made by tackles and territorial gains?
3. If we want to predict a team’s success, is it better to build networks of passes and disruptions for each action separately or for the entire match?
4. Can we use these relationships to “predict” the outcome of the match?

A passage network is simply a network whose nodes are the players of a team and the directed connections go from the player originating a pass to the player receiving the ball. We consider only completed passages: the ones that did not result in an error or lost possession. In the above picture, those are the green edges and they are always established between players belonging to the same team. In rugby, players are allowed to tackle the current ball carrier of the opponent team. When that happens, we create another directed edge, this time in what we call “disruption network”. The aim of a tackle is to prevent the opponent team from gaining meters. These are the red edges in the above picture and can only be established between players belonging to opposite teams. The picture you see is the collection of all passes and tackles which happened in the Italy vs New Zealand match in 2012. It is a multilayer network as it contains edges of two different types: passes and tackles.

Once we have pass and disruption networks we can calculate a collection of network measures. I’ll give a brief idea here, but if you are looking for more formal definitions you’ll have to search for them in the paper:

• Connectivity: how many pass connections you have to remove to isolate players;
• Assortativity: the tendency of players to pass the ball to players with a similar number of connections — in high assortativity central players pass to other central players and marginal players to other marginal players;
• Components: how many “sinks” there are, in that the ball never goes back to the bulk of the team when it is passed to a player in a sink;
• Clustering: how many triangles there are, meaning that the team can be decomposed in many different smaller sub-teams of three players.

These are the features we calculated for the pass networks. The disruption case is slightly different. We calculated the same features for the team when removing the tackled player, weighted on the relative number of tackles. If 50% of the tackles hit player number 11, then 50% of the disrupted connectivity is the connectivity value of the pass network when removing player 11. The reason is that the tackled player is temporarily removed from the game, so we need to know how the team performs without him, weighted on the number of times this occurrence happens.

So, it is time to give some answers. Shall we?

1. Is there a relationship between the topology of the network of passes and the success of the team?

Yes, there is. We calculate “success” as the number of meters gained, ball in hand, by the team. The objective of rugby is to cross the goal line carrying the ball, so meters made is a pretty good indicator. We control for two things. First, the total number of passes: it simply means the team was able to hold onto the ball longer, so it is trivially expected to result in more meters. Second, the home advantage, which is a huge factor in rugby: Italy won only 12 out of 85 matches in the European “Six Nations” tournament, and 11 of them were in Italy. After these controls, we find that two features have good correlations with meters made: connectivity and components. The more edges are needed to isolate a player, the more meters a team is expected to make (p < .01, R2 = 47%). More sinks in a team is associated with lower gains in meters.

2. Is there a relationship between disruptions made by tackles and territorial gains?

Again: yes. In this case it seems that all calculated features matter to predict meters made. The strongest factor is again leftover connectivity. It means that if the connectivity of the pass network increases after the tackled player is removed from it then the team is able to advance more. Simplifying: if you are able to tackle only low connectivity players, then your opponent is able to gain more territory (p < .01, R2 = 48%).

3. Is it better to build networks of passes and disruptions for each action separately or for the entire match?

The answer to the previous two questions were made by calculating the features on the global match networks. The global network uses all the data from a match, exactly like the pass and disruption edges depicted in the above figure. In principle, one could calculate these features as the match unfolds: sequence by sequence. In fact, networks features at the action level work very well in soccer, as Luca and Paolo already proved. Does that work also in rugby?

Surprisingly, the answer is no. We recalculated the features for each passage of play. A passage of play is the part of a match from when a team gets into possession of the ball until it loses it, scores, or the game flow stops for an infraction. When we calculate features at this level, we find very weak correlations: almost nothing is significant and, when it is, the predictive power is very low. We think that this is because in rugby our definition of sequence is too strict. While soccer is a tactical game — where each sequence counts for itself — rugby is a grand strategy game: sequences build cumulative advantage which pays off after a series of them — or only in the match as a whole.

4. Can we use these relationships to “predict” the outcome of the match?

This is the real queen question of the post, and we do not fully answer it, unfortunately. However, we have a very good reason to think that the answer could be positive. We created a predictor which trains on 17 matches and then, given the global multi-layer network, will pick the winner. You can see the problem of the approach here: we use the network of the match as it happened to “predict” the outcome. However, we did that only because we did not have enough matches for each individual team: we believe we can first predict how pass and disruption networks will shape in a new match using historic data and then use that to predict the outcome. That will be future works, maybe if some team is intrigued by networks and wants to contact us for a collaboration… (wink wink).

The reason I still like to report on our predictor is that it has a very promising property. Its accuracy was 83%. We compared with a prediction made with official rugby rankings, whose performance is worse: 76% accuracy. We also tested against bookmakers, who are better than us with their 86% accuracy. However, historic data on bets only cover more important matches — only 14 out of 18 — and matches between minor teams are usually less predictable. The fact that we are on par on a more difficult task is remarkable. More importantly, bookies tend to just “choose the best team”. For instance, they always predict a New Zealand win. The Haka, however, is not always enough and our networks caught that. New Zealand lost to England in a big upset on December 1st 2012. The bookmakers didn’t see that coming, but our network approach could have.

Continue Reading

As a kid from the 80s*, I remember how revolutionary the cellphone era was. It happened so fast. It seemed that, overnight, you could carry in your pocket a device connecting you to everybody you knew, no matter how far. To me, it changed everything. But did it? Yes, over-apprehensive parents can check their babies at the swipe of a finger, and whoever does not carry their cellphone with themselves at all times is labeled as a weirdo — I’m guilty of that. But the telecommunication revolution promised something more: the elimination of distance in communication. Did it deliver? This question was the motivation engine for the paper “Evidence That Calls-Based and Mobility Networks Are Isomorphic” which I wrote with my boss Ricardo Hausmann and which recently appeared in PLoS One.

The question is rather daring, so we decided to take it step by step. The simplest thing we came up was: let’s draw a map of cellphone calls and see if it looks like a geographical map. If it does, we might be onto something. To do so, we obtained data from telecommunication operators in Colombia. They provided us call detail records, where identifiers were encrypted to preserve the anonymity of the people making and receiving the calls. We also aggregated the data to make even the slightest re-identification impossible: every ID was associated to the municipality in which it spent most of its time and so all data was lumped together at the municipality level. At this point, we could draw a map of which municipalities had a significant call traffic with one another. This we called the “Call-based” network:

Click to enlarge

Before jumping to conclusions with this picture, we built a sister network. Since we just said we knew the location of a phone when making a call, we can keep a record of the different municipalities where we spotted the phone. Again, we joined together all data at the municipality level. This sister network is then a “Mobility” network of Colombia:

Click to enlarge

It seems there’s something here. The two networks appear to be similar: Bogotá seems to be a prominent center and the connections have a geographical component embedded into them. To make this more evident, we drew the networks on a Colombian map. The color of the municipalities is the same color of the nodes in the pictures above: nodes with the same color are very related in the network — network clusters.

Click to enlarge

The call-based network is on the left, the mobility is on the right. Blocks of the same color on the left are a clear indication of the call connections being influenced by geography. If there was no relation, the map would look like the Harlequin shirt, with colors scattered evenly across the territory. Mobility clusters are also short-range, although the pattern is harder to see because I had to use many more colors: the clusters are smaller. But the two networks are closely related: in fact, the larger call-based clusters contain the smaller mobility ones, as we show in the paper. We can say that there is a strong relationship between calls and mobility.

This is nice, because it fits with many works in computer science that actually use social relationships to predict human mobility… and vice versa. On the other hand, it is not nice because the existence of these papers also tells us ours is not a new result. Moreover, my starting point was to hint that the call-based and mobility networks are obeying the same laws, not that they are merely correlated. We need to go a step further.

Our step was to consider the difference that distance makes in the two networks. When looking at mobility, the distance between an origin and a destination is an important cost. In the call-based networks, things are a bit trickier. If modern telecommunication really delivered what it promised, distance should be a really low cost, and probably non-linear. To start a social relationship it is not needed to be in the same place at any given time, and even if we move to opposite ends of the world, we can still call each other. As a consequence, there shouldn’t be a way to scale the cost of distance in the call-based network to look like the one in the mobility network.

When we attempted to perform such scaling, we discovered it was actually possible. We checked, at any given distance, the ratio between commuters and callers. If two municipalities are at 50km distance, and there are twice as many commuters than callers, we have a dot on coordinates (50, 2). If we take two municipalities at 100km distance, and the commuters are just a third of the number of callers, the data point is at coordinates (100, .33). Once we consider all data points, we can fit our green line, AKA the scaling function from calls to mobility:

When we used this adjustment to calculate new call-based clusters using the distance cost “as if” it was the mobility network, we obtained the mobility clusters. We detail in the paper the reasons why this is not as circular as it seems.  In practice, our green line is a transformation function that morphs the call-based network into the mobility network. If modern telecommunication really killed distance, that green line shouldn’t exist, or at least it should be so wobbly to be practically useless.

There are many ways in which you could interpret this result. One that Ricardo and I like focuses on the relationship between face-to-face and electronic mediated meetings. It’s not like the people you call are the ones you really would rather meet but you cannot. It’s more like you call AND you meet, whenever it is possible. Face-to-face and electronic mediated meetings are not really substitutes in this world, they are more like complements. To come back to my opening, I’d say new technologies didn’t eliminate distance from the communication equation. Alleviate, yes. But ultimately, it’s more like an increased bandwidth than a revolution. At least so far.

* Shut up, I’m still in my twenties. Everybody knows 1996 was only 10 years ago.

Continue Reading

You need to forgive me for the infamous click-bait title I gave to the post. You literally need to, because you have to save your hate for the actual topic of the post, which is Big Data. Or whatever you want to call the scenario in which scientists are flooded with so much data that traditional approaches break, for one reason or another. I like to use the Big Data label just because it saves time. One of the advantages of Big Data is that it’s useful. Once you can manage it, simple analysis will yield great profits. Take Google Translate: it does not need very sophisticated language models because millions of native speakers will contribute better translations, and simple Bayesian updates make it works nicely.

Of course there are pros and cons. I am personally very serious about the pros. I like Big Data. Exactly because of that love, honesty pushes me to find the limits and scrutinize the cons of Big Data. And that’s today’s topic: “yet another person telling you why Big Data is not such a great thing (even if it is, sometimes)” (another very good candidate for a click-bait title). The occasion for such a shameful post is the recent journal version of my work on human mobility borders (click for the blog post where I presented it). In that work we analysed the impact of geographic resolution on mobility data to locate the real borders of human mobility. In this updated version, we also throw temporal resolution in the mix. The new paper is “Spatial and Temporal Evaluation of Network-Based Analysis of Human Mobility“. So what does the prediction of human mobility have to do with my blabbering about Big Data?

Big Data is founded on the idea that more data will increase the quality of results. After all, why would you gather so much data at the point of not knowing how to manage them if it was not for the potential returns? However, sometimes adding data will actually decrease the research quality. Take again the Google Translate example: a non native speaker could add noise, providing incorrect translations. In this case the example does not really hold, because it’s likely that the vast majority of contributions comes from people who are native speakers in one of the two languages involved. But in my research question about human mobility it still holds. Remember what was the technique in the paper: we have geographical areas and we consider them nodes in a network. We connect nodes if people travel from an area to another.

Let’s start from a trivial observation. Weekends are different from weekdays. There’s sun, there’s leisure time, there are all those activities you dream about when you are stuck behind your desk Monday to Friday. We expect to find large differences in the networks of weekdays and in the networks of weekends. Above you see three examples (click for larger resolution). The number of nodes and edges tells us how many areas are active and connected: there are much fewer of them during weekends. The number of connected components tells us how many “islands” there are, areas that have no flow of people between them. During weekends, there are twice as much. The average path length tells us how many connected areas you have to hop through on average to get from any area to any other area in the network: higher during weekdays. So far, no surprises.

If you recall, our objective was to define the real borders of the macro areas. In practice, this is done by grouping together highly connected nodes and say that they form a macro area. This grouping has the practical scope of helping us predict within which border an area will be classified: it’s likely that it won’t change much from a day to another. The theory is that during weekends, for all the reasons listed before (sun’n’stuff), there will be many more trips outside of a person’s normal routine. By definition, these trips are harder to predict, therefore we expect to see lower prediction scores when using weekend data.

The first part of our theory is proven right: there are indeed much less routine trips during weekends. Above we show the % of routine trips over all trips per day. The consequences for border prediction hold true too. If you use the whole week data for predicting the borders of the next week you get poorer prediction scores. Poorer than using weekday data for predicting weekday borders. Weekend borders are in fact much more volatile, as you see below (the closer the dots to the upper right corner, the better the prediction, click for higher resolution):

In fact we see that the borders are much crazier during weekends and this has a heavy influence on the whole week borders (see maps below, click for enjoying its andywarholesque larger resolution). Weekends have a larger effect on our data (2/7), much more than our example in Google Translate.

The conclusion is therefore a word of caution about Big Data. More is not necessarily better: you still need theoretical grounds when you add data, to be sure that you are not introducing noise. Piling on more data, in my human mobility study, actually hides results: the high predictability of weekday movements. It also hides the potential interest of more focused studies about the mobility during different types of weekends or festivities. For example, our data involves the month of May, and May 1st is a special holiday in Italy. To re-ignite my Google Translate example: correct translations in some linguistic scenarios are incorrect otherwise. Think about slang. A naive Big Data algorithm could be caught in between a slang war, with each faction claiming a different correct translation. A smarter, theory-driven, algorithm will realize that there are slangs, so it will reduce its data intake to solve the two tasks separately. Much better, isn’t it?

Continue Reading

The literature about community discovery, which deals with the problem of finding related groups of nodes in a network, is vast, interesting and full of potential practical applications. However, if I would have to give one critique of it, it would be about its self referential character. Most community discovery papers I read in computer science and physics journals are mainly about finding communities. Not much time is spent thinking about what to do with them, or what they mean. My first post in this blog was about a community discovery algorithm. Recently, an extended version of that paper has been accepted in a computer science journal. Since that first post, I (mainly) added some crucial modifications and features to the algorithm. I don’t want to talk about those here: they are boring. I also didn’t bring up this paper to boast about it. Okay, maybe a little. I did it because the paper touches upon the issue I am talking about here: it tries to do something with communities, it tries to explain something about them. Namely, it asks: why do communities overlap?

First of all: communities do overlap. When trying to detect them, many researchers realized that hard partitions, where each node can belong to one and only one community, are not always a good idea. Most of them found this a problem. Others were actually very happy: the problem gets harder! Nice! (Researchers are weird). Blinded by their enthusiasm, they started developing algorithms to deal with this overlap. Not many asked the question I am trying to answer here: why do communities overlap? As a result, some of these algorithms detect this overlap, but using approaches that do not really mean anything in real life, it’s just a mathematical trick. Others, instead, build the algorithm around a core hypothesis.

This hypothesis is nothing unheard of. Communities overlap because people have complex lives. Some of your college mates also attend your yoga class. And you know your significant other’s colleagues, which puts you in their community. All these communities have you as common member, and probably some more people too. The beauty of this is that it is not only intuitive: it works well in finding communities in real world social networks. So well that it is the assumption of my approach and of many others outstanding algorithms (this and this are the first two that pop into my mind, but there are probably many more). Another beautiful thing about it is that it is almost obvious, and so it is probably true. But here we hit a wall.

The fact that it is simple, reasonable and it works well in practice proves nothing about its property of being true. There are things that are not simple nor reasonable, but nevertheless true (hello quantum physics!). And there is practical knowledge that does not quite correspond to how things work (in my opinion, most computer science is a patch and nobody really knows why it works). Unless we test it, we cannot say that this nice practical principle actually corresponds to something happening in reality. So how do we go on and prove it? In the paper I proposed a first step.

This brings me back to another old love of mine. Multidimensional networks. They are networks in which we put multiple relations in a cage together in mating season and see what happens (research is fun). The idea behind the paper is that multidimensional networks give us the perfect tool to test the hypothesis. In monodimensional networks you have no clue why two people are connecting besides the obvious “they know each other”. In a multidimensional network, you know why they know each other, it’s information embedded in the type of the relation. So, the hypothesis is that different types of relations are the cause of the community overlap, and with multidimensional networks we can look at how communities distribute over relations. First, let us take a look at what two overlapping communities look like in a multidimensional network.

We collected a multidimensional social network putting together relationships between users in Facebook, Twitter and Foursquare. We then used DEMON to extract overlapping communities from each dimension. We then took two communities with extensive overlap in the Facebook dimension (picture below).

We then looked at the very same set of nodes, but now in the Foursquare network. In the picture below, we kept the edges, and the node positioning, of the Facebook network to make the comparison easier, but keep in mind that the edges in the Foursquare dimension are different, and they are the ones that decide to what community the nodes belong.

Very interesting. The communities look a lot alike, although the shared (and non shared) nodes are slightly different. Now node 7369 is shared (it wasn’t in Facebook) while node 8062 isn’t (whereas it was before). Let’s put another nail on the coffin and see the communities these nodes belong in Twitter (same disclaimer applies):