Definition of Criteria for Impact Analysis A work

Definition of Criteria for Impact Analysis A work in progress Christoph Fuchs

Overview 1. What I won't talk about 2. Twofold 3. Classical social network analysis 4. Criteria in the online world 5. Specific definition of criteria 6. Open questions

What kind of impact? I will not talk about impact in the "real world"! Such as. . . Total revenue Offline bookings Offline sales. . . • •

What I will talk about Impact of actors and shared items on online social platforms. Looking for impact criteria which are accessible (via automatic retrieval) independent of platform specifics • •

Two kinds Impact analysis is twofold

Two kinds of impact analysis • actor level (potential impact)"Who is the most important guy? "

Two kinds of impact analysis • item level (impact after publication)"What is 'hot' right now? "

In other terms • • = page post Impressions". . . measure the number of times a post from your Page is displayed, whether the post is clicked on or not. " Reach". . . measures the number of people who received impressions of a Page post. "

"Classical" vs. "Online"

Classical Social Network Analysis aka Graph Theory

Measures from "classical" Social Network Analysis All* classical approaches are on the actor level. • • • Centrality Prestige described by Wasserman and Faust in 'Social Network Analysis' * at least all I have found so far.

Who is the "most important guy"?

Centrality measure C(actor) = [0. . 1] Normalized on the graph size. Online: Where does your graph end?

Centrality • • • Degree Centrality"Who has the most ties to other actors? " Closeness Centrality"How close is an actor to all the others? " Betweenness Centrality"Interactions between two nonadjacent actors might depend on the other actors in the set of actors, especially the actors who lie on the paths between the two. "

Degree Centrality "Who has the most ties to other actors? " Cdinbound(orange) = 4 Cdoutbound(orange) = 3 Cdoutbound(green_women) = 0. 2 [normalized: 4 / 5 [normalized: 3 / 5 = 0. 8] = 0. 6]

Closeness Centrality "How close is an actor to all the others? " Online problem: Where does your graph end? Relevant for the online world?

Betweenness Centrality "Are you between any important actors? " Online problems - Mostly direct communication (to immediates) - Distance can be easily reduced to 1

Prestige • • • Degree Prestige. Indegree of each actor. Proximity Prestige. Based on average distance of actors in an influence domain. Status or Rank Prestige"If one’s influence domain is full of prestigious actors, one’s prestige should also be high. If, however, an actor’s domain contains only peripheral, or marginally important, actors, then the rank of this actor should be low. "

Criteria in the online world

Criteria?

Criteria?

Criteria?

Platform specific criteria

Item level • • "How many people have seen it? " "How many people have clicked on it? "

Item level • "How many people have reacted on it? And how did they react? " Like (+1, thumbs up, . . . ) o Favourite o Reply or comment o • "How many people have re-shared it? "

Criteria? • All kinds of criteria. . . o o o Views and clicks Unary feedback Binary feedback Ratings Re-publication

Criteria? • . . . and more Replies (may not be text) o Comments o Sentiment of comments o

Criteria? • Platform specific examples Moods. Deviant. ART o Bumpingvarious Forums o Pull Requests, Forks, Watches github o. . . o

What to make of it? So many things to measure!

We need impact criteria which are • • accessible independent of platform specifics

Actor level criteria

Actor level criteria Criterion (#) Value Actors who are friends with you Integer Actors whom you are friends with Integer Actors who want to be friends with you [*] Integer Actors you want to be friends with [*] Integer Actors subscribed to your stream ("following" you) Integer Actors to whom you are subscribed Integer Reputation score [? ] Integer [*] Are those important? [? ] Not sure about available reputation scores - platform dependency.

Item level criteria Criterion (#) Value Views Integer Clicks Integer Positive unary feedback Integer Special unary feedback Integer Negative unary feedback Integer Re-publications Integer Replies Integer Comments Integer / Text Sentiment Hash with Integer

Open Questions. . . or "Problems"

Open questions • What about re-publications? o broadens the 'impact area' o you don't have access to data outside your immediate peers o It may be important who re-published it. Almost impossible to measure how many times an item has been viewed if it goes viral.

Open questions Is normalization important? • Normalization o Usually dependent on graph size o Online social network graphs are huge!

Open questions Is the actor's reputation important? • What about. . . o the success of previously published items? o the sentiment of previously published items? o Is any publicity good publicity?

Open questions • Weighing o • What criteria are the most important? Re-publication >? Like Combination across platforms o One platform may only have 'Like' o Online social network graphs are huge!

Open questions • Algorithm? Can existing classical approaches be used? o Existing open-source algorithms? o Is performance important? o To be investigated. . .

Thank you. Questions? Discussion!