exogenous variables as hyperevent attributes

[Jasmine Wu]

Dear Juergen and eventnet users, 

I was wondering if there is a way to set up statistics for hyperevent attributes in eventnet? 

One example is in the co-citation/co-authorship setting. You may cite a paper because (1) the paper is relevant (based on a certain exogenous variable, e.g., paper topic) and (2) the authors of the paper. And I am interested in which effect (paper or author) is larger. In this case, the paper topic is an exogenous hyperevent attribute. 

I could create a DIR_HYPER_LEVEL attribute, but I do not see an option in DHE statistics to add a DHE_EDGE_STAT, like what we can do for node attributes with DHE_NODE_STAT. 

Thanks,

Jasmine

[juergen lerner]

Dear Jasmine,

I see several options, depending on how the papers' topics are exactly coded and what are the topic-dependent effects that you want to estimate.

I think what is given is that the hyperedges you want to explain have a format (I,J), where I={i1, i2, ...} are the authors and J={j1, j2, ...} are the references

(1) There is actually a DHE statistic dependent on values of hyperedge attributes: you simply can take DHE_SUB_REPETITION_STAT. It's a bit unfortunate that this is called "repetition" - when you want to define it based on an exogenous attribute (the topic). But note that the subrep stat is defined dependent on a hyperedge attribute (could be a directed or undirected hyperedge attribute, dependent on the settings). This hyperedge attribute is often defined via prior events (justifying the name "repetition") but it can also be specified via exogenous hyperedge attributes. The exogenous hyperedge attribute might assign values to single nodes, dyads, or general hyperedges (general subsets or pairs of subsets). Note that also single nodes or dyads can be considered as special cases of hyperedges. So you can code their values in a UHE/DHE_ATTRIBUTE (and you have to code them as hyperedge attribute to be able to use them in sub repetition, even if they are defined for single nodes or dyads).

If this approach works, note that attribute values that are set at time t are only "visible" to statistics computed at time t' > t. (In real subset repetition this is to ensure that events depend only on the past.) So you have to set the exogenous hyperedge attributes just a moment before the actual interaction that you want to explain, possibly by making the time scale more fine-grained. This comment also applied to the other cases below.

(2) I don't know how many topics you have but if the number is small (k) you may define k different node-level attributes for papers and then use DHE_NODE_STAT (k variants of it, for each of the topics). 

(3) You may also define a third mode (topics) besides authors and papers. You then attach a paper node to its topic nodes via a dyad attribute and then you may use DHE_NEIGHBOR_STAT. This can be dependent on a dyad attribute (connecting papers to topics) and (potentially) a node attribute (quantifying an aspect of the neighboring nodes, that is, the topics). This might be better than Option 2 if you have many topics and especially if a paper can have more than one topic.

(4) Yet a different variant might be to introduce a third mode "topics" as in Option 3 and then use a closure statistic or perhaps even a 4cycle statistic making use, among others of the paper-topic dyad attributes, somewhere on the intermediat two-path or three-path that is being closed by the new event. This might be used to define even more complex effects.

(5) If the event itself (that is, the publication event itself) also has a topic you may code it as an event type (if there are not too many topics). An alternative might be to "attach" the topic information of each event to the output table of eventnet and then compute effects such as the share of cited papers that have the same topic as the publishe (citing) paper.

(6) Point 5 demonstrates a current limitation of eventnet: while the number of modes is theoretically unlimited (allowing to specify node sets for authors, papers, and topics), each single event connects just two modes. This will be overcome in the next eventnet version, where "generalized hyperevents" can connect any number of node types. This is related to a previous conversation "Categories in RHEM effects". It would allow you to specify also the topics of the published/citing paper directly in the event and to define effects related with topic-nodes (such as citing papers within the same topic). 

I hope that some of it helps.

Best wishes

  Juergen

[Jasmine Wu]

Thanks so much for the thoughtful response, Juergen. I'll see which option fits best with my scenario. 

Best,

Jasmine