Extracting Communication Networks - Borgatti 2005

[jamesallenevans]

Post questions here for:

Borgatti, Stephen P. 2005. “Centrality and Network Flow.” Social networks 27(1): 55-71.

[lkcao]

This piece is really illuminating, and urges me to think further about network theories. However, In practice, I know many people use centrality measures interchangeably not because they do not care about theory, but because there is a high correlation between multiple centrality measures and they are useful proxies. For example, if we take table 5 of this Borgatti paper and calculate the correlation table of different measures, we get the following table:

which means they are nearly the same thing in linear algebra sense. So can we say that Borgatti piece is more important in theoretical sense than practical sense? Or the difference between multiple measures will be more significant in a larger network?

[katykoenig]

This paper examines the use of different centrality measures with different flow processes. Specifically, the author tests Freeman closeness and Freeman betweenness measures on different flow processes and concludes that development of new measures is necessary to better predict non-geodesic (and more realistic) flow. Would other eigenvector centrality measures (like Katz or PageRank with their dampening parameter) better model non-geodesic flows, like money exchange or gossip, than the measures described in this paper?

[rachel-ker]

This paper was helpful to understand some of the metrics and the implicit flow characteristics it measures. I noticed that in comparing results for the simulation e.g. in Table 2, the similarity of the scores is compared by the given rank of the metrics. Is the purpose of these centrality measures to produce an idea of ranking for a particular network? Given that certain processes like infection and gossip are frequently modelled in simulations and agent-based modelling, why would not understand the networks through these simulations instead of creating new measures?

[sunying2018]

This paper demonstrates many different centrality measures. But I am quite confused by one popular measure of centrality - eigenvector centrality mentioned in this article. It seems like it has the same logic as page rank algorithm. For the eigenvalue decomposition of adjacency matrix, we need to find the the eigenvector which is corresponding to the largest eigenvalues, I am not sure what's the meaning of "eigenvector score" mentioned in this article and whether it is same as eigenvalues?

[laurenjli]

I liked how the author explained the different types of flow inherent in each measure. He states at the end that the measures should be thought of as expected values of node participation. If that's the case, how can the actual node participation be determined?

[HaoxuanXu]

This paper demonstrates a range of centrality measures for different types of flow processes, with each centrality measure corresponding to specific network scenarios such as the package delivery process. It will be interesting to know if there's a network process that fits multiple centrality measures.

[tzkli]

This paper examines the applicability of different centrality measures based on the assumptions baked into their construction. Similar to @clk16 , I'm wondering whether the similar rank orders of the simulations and Freeman measure are an artifact of the small number of actors in those simulations or a regular pattern? Also, as this paper was published a while ago, do we have something like a cookbook of centrality measures now?

[bjcliang-uchi]

In ML we have ways to combine multiple algorithms to make better predictions. I am just wondering if, since one type of centrality has to sacrifice certain types of network features, there is a way to combine different measures together.

[di-Tong]

Related to @clk16 's question, I wonder why different centrality measurements perform similarly even though they are designed for very different situations. Do some measurements perform more similar to each other than others and why?

[ckoerner648]

It is interesting to read about different types of centrality measures and the particular ways how things can spread in different networks: a $20 bill is passed on from one person to another, emails can follow an answer-response pattern, and attitudes of which fashion items are “in” may be multiplied by certain nodes in a network. It is interesting to see that certain statistical centrality measures make different assumptions of how centrality is defined–and thus calculated. I’d be interested in how to turn a text into a network and how to define the centrality of words–is their frequency more important or their position relative to other parts of the text?

[wunicoleshuhui]

I find it very informative that the type of flow processes determines the level of importance for the central nodes. However, since this paper was published in 2004, I'm wondering how much the changes in communication technologies impacted the type of flow processes in communication. How should we categorize, say, social media interactions? Should we characterize these as paths, trails, walks, or a combination of these types of flows?

[cindychu]

This article is so interesting and informative, changed the way I perceived the concept of centrality before greatly. However, since ‘centrality’ is a major concept in graph/network analysis to examine, while there are so many restrictions for calculating centrality, is there any way to more robustly apply these centrality calculations into networks of multiple different characteristics? for example, is it possible to neglect some parts of non-connected graph when calculating closeness centrality?

[skanthan95]

I don't have much prior background in social network analysis, but this reading reminded me of an article, The Relative Ineffectiveness of Criminal Network Disruption. Here, degree and betweenness centrality measures should've pointed to the most central criminals. But, when these criminals were subsequently pursued, it did not lead to their criminal network becoming disrupted (in fact, the network strengthened in some cases). The authors state that this is because powerful criminals are less traceable, and so the analyses were actually revealing the most vulnerable criminals instead. Could this problem have been solved by using other measures of centrality? Should the idea of centrality been left behind in this case?

[deblnia]

If betweenness and centrality measures are "expected values" and limited as such, how are we supposed to describe networks? In Borgatti's view, should we just make a note of flow simulations (although I don't know that that's possible in non-simulated networks) and justify the limited centrality measure that way? Or do we abandon the concept altogether, and move to MCMC network descriptors like MLE?

[heathercchen]

This paper illustrates new topological methods for evaluating centrality, which provides me with novel perspectives on social network analysis. My question is in the author's framework of analyzing network flows, is it possible to assign different weights to different nodes such that we can include another form of information that indicates relative significance?

[jsmono]

The author proposed an interesting angle to analyze social networks. When he mentioned the limitations of his theory, he mentioned the case where flows originate at each node systematically but have no particular target. What would be an example of this? For example, do social media posts such as Instagram feeds belong to this category since the targe is relatively random?

[yirouf]

It's interesting to know all kinds of flow processes, but regarding social networks, I have similar questions about the tradeoff between centrality and network features, is it possible that we might combine some of the flows to make up for the tradeoffs?

[chun-hu]

A very informative piece on different centrality measures in network theories. Like others, I'm also curious about whether we can combine the flows as we do in ensemble learning to eventually produce a stronger measurement?

[luxin-tian]

Echoing others' questions, I am interested to see if there are any "bagging" measures that can be used with some robustness?

[gracefulghost31]

The authors acknowledge that one of the limitations is its lack of applicability to obtain estimates of the expected values for the flow process in the absence of a specified target. What would be the way around to accomplish this under this framework?

[ccsuehara]

I was wondering if, when using these centrality measures, one can predict how central the nodes will be in the future. Do initial conditions reiterate what final outcomes will be? I understand that the simulations help you get a result by adding stop conditions, but this is also implying that we reached the "end of times". How does prediction analysis complement with ending the simulation?

[ziwnchen]

The findings of the paper are really inspiring! It argues that the centrality measures based on static network topology actually have certain assumptions about the dynamic network flow. Therefore, centrality, to some degree, could be viewed as "expected participation if things flow in assumed way". Unfortunately, this paper does not provide specific centrality measure formula for different kinds of flows, which makes it hard to apply it to practical research. My question is, is there any related research in this field that explores particular centrality formula for particular types of flows?

[VivianQian19]

Borgatti’s article illuminates how the importance of a node in a network is a result of both the position of the note and features in the flow process (69). His two-dimensional typology of network flow as well as his discussion of each centrality measure’s assumptions and how they match with different processes of network flow is also very informative. As such, he argues that centrality measures such as betweenness and closeness generate expected value only when assumptions of network flow is met. I’m wondering rather than thinking about flow process of information follows his ideal types (e.g. gossip, viral infection), how to account for network flows that occur in more than one network flow?

[vahuja92]

I enjoyed reading about the different types of flows, and the examples given at the beginning. I have a similar question to @wunicoleshuhui. How can we categorize flows of information on social media sites? How would we measure centrality in this case?

[arun-131293]

The main contribution of this paper seems to be to explicate the assumptions about network flows implicit in different methods of calculating network centrality. Once these assumptions are explicated, abstract terms like "node importance" and "node participation" give way to clearer features like "which nodes will receive flows (quickly, frequently, and certainly) and which are in a position to control flows"

[rkcatipon]

Like @vahuja92 @wunicoleshuhui I'm also curious about social networks! I've seen social media software products leverage eigenvector centrality through engagements on posts (e.g. likes, retweets) to generate social network graphs. The products would then assume a serial duplication path for the information flow. But for the meta-information like diffusion of influence in a social network, and not just how a post goes viral, I am also curious.

[alakira]

This paper emphasizes the assumption of dynamic flow with different centrality. I'm not familiar with network analysis. Is there a way to detect the characteristics of dynamics itself? Also, based on @skanthan95 's comments, is there a way to detect nodes that might be hidden from existing networks?

[sanittawan]

I like how the structure of the paper is very easy to follow. I wonder if anyone can give an example to illustrate how to determine what kind of data one is dealing with. For instance, how can I be sure that the data that I am working on meets the implicit assumptions of these measures? Is there a good way to determine if it's a walk, trail, path or geodesics? I can probably imagine a situation where I would be unsure between two typologies which, as the author mentions, would lead to poor or wrong results.

[kdaej]

I believe the time used in this paper is not the physical time but the incidents of nodes. However, I can imagine a case when the influence of one incident fades with time. Is it possible to align the nodes on the actual timeline and compare it with the centrality analysis to see how time plays a role in the process? For instance, the citations network indicates influences between scholars. At the same time, scholars are more likely to cite recent articles. In this case, taking into account the real-time for the distance between nodes might provide more information about the network.

[YanjieZhou]

I have got informed of the network theories and very interested in its application in content analysis. But I am also wondering whether there are other methods that can add to the robustness of bagging procedure.

[cytwill]

One thing I am interested in this paper is the classification of different types of flows mentioned. In this article, centrality metrics are used to evaluate those different type flows, which I think are of great potential to be applied to developing some measures to control the flows in those networks. Therefore, I am wondering if there are any specific real-life applications/examples of using these centrality metrics to improve/constrain those network flows. Also, sometimes the vertex in a network graph can have very diverse attributes. for example, some refer to people, while others are articles. So under this situation, could those different network flows be defined in the same way, or if the definitions of those centrality metrics remains the same for each type of vertex, especially those at the at an overlap position?

[Lizfeng]

The author conveys an important message in this paper: choice of centrality measurement should be based on the type of network. By categorizing different kinds of network through two dimensions: diffusion type and trajectory type, the author was able to point out that certain centrality measure such as Freemen Centrality is only suited for a specific type of network such as package delivery. The author pushes for considering the centrality measure in a particular network context in the future.