Source-linked AI summary
Measuring user influence on Twitter: A survey
Fabián Riquelme, Pablo González-Cantergiani
TL;DR
The paper addresses how to identify influential Twitter users amid competing definitions and a large, diverse literature on centrality measures. It surveys and classifies activity, popularity, and influence measures, along with their computational properties. The survey finds especially broad use of PageRank and topical-sensitive approaches, while predictive measures remain few.
Problem
Influential-user measurement lacks an agreed definition, while useful measures must classify relevance and remain computationally feasible.
Method
The paper comprehensively surveys and classifies Twitter measures by activity, popularity, influence, relationships, PageRank, timelines, topical analysis, and computational complexity.
Results
Almost one half of existing influence measures are based in one way or another on PageRank, while predictive influence measures are few.
Takeaways & Limitations
Twitter influence research spans diverse measures, with retweets prominent for influence, follows for popularity, and replies for activity.
Takeaways & Limitations
The survey leaves comprehensive correlation of all included measures as an open problem, and existing checks do not establish correlation.
Abstract
from arXiv · showhide
Centrality is one of the most studied concepts in social network analysis. There is a huge literature regarding centrality measures, as ways to identify the most relevant users in a social network. The challenge is to find measures that can be computed efficiently, and that can be able to classify the users according to relevance criteria as close as possible to reality. We address this problem in the context of the Twitter network, an online social networking service with millions of users and an impressive flow of messages that are published and spread daily by interactions between users. Twitter has different types of users, but the greatest utility lies in finding the most influential ones. The purpose of this article is to collect and classify the different Twitter influence measures that exist so far in literature. These measures are very diverse. Some are based on simple metrics provided by the Twitter API, while others are based on complex mathematical models. Several measures are based on the PageRank algorithm, traditionally used to rank the websites on the Internet. Some others consider the timeline of publication, others the content of the messages, some are focused on specific topics, and others try to make predictions. We consider all these aspects, and some additional ones. Furthermore, we include measures of activity and popularity, the traditional mechanisms to correlate measures, and some important aspects of computational complexity for this particular context.
1 Introduction
Social-network centrality ranks users by relative importance, but influence is difficult to define consistently. This survey organizes Twitter measures and examines their relationships, scope, and computational demands.
- Centrality and influence: Centrality measures rank network actors using different relevance criteria, such as adjacent-edge counts or shortest-path distances.Degree counts adjacent edges, whereas closeness uses the minimum sum of shortest paths to all other nodes.
- Centrality and influence: Influence can spread through cascades from initially active nodes to distant users, motivating influence measures based on propagation.A node may be reached indirectly through a sequence of intermediary nodes rather than only by a direct tie.
- Twitter motivation: Twitter influence is important to study because many users are inactive or provide little additional information.The survey places influential-user identification in the broader context of Twitter’s extensive information spread and applications.
- Survey scope: The survey classifies measures by activity, popularity, and influence, while also tracking Twitter relationships, PageRank, timelines, topical analysis, and complexity.Its classification table records whether measures use follows, retweets, mentions, replies, likes, PageRank, timelines, or content analysis.
- Survey scope: The paper claims to be the first attempt to bring together all centrality measures used in the Twitter network.It contrasts this comprehensive scope with earlier partial surveys covering fewer measures or broader online-social-network settings.
2 A brief description of the Twitter network
Twitter can be modeled as a directed graph using follows, dynamic interactions, or users and tweets together. The survey’s measures combine network relationships with metrics, topical analysis, and other signals, though influence indicators remain conceptually contested.
- Twitter relationships: Twitter relationships include user-to-user, user-to-tweet, tweet-to-tweet, and tweet-to-user actions, with likes treated as analogous to favorites.The relationship types and valid actions are summarized in Table 2.
- Network representation: Twitter is modeled as a directed graph whose nodes and edges represent users and their relationships.The traditional model uses follow-up ties; alternative models represent dynamic interactions or users and tweets together.
- Network representation: G1 models follows, G2 models mentions, replies, and retweets, and G3 represents all possible relationships.The more complete representation partitions nodes into users and tweets, with edges encoding four relationship types.
- Metrics and measures: Metrics provide numerical network information, while measures combine metrics into formulas or algorithms that rank users.The survey describes metric systems involving original tweets, replies, retweets, mentions, and graph characteristics.
- Influence signals: Retweet metrics are the best quantitative indicators for choosing which tweet to read, but qualitative factors such as friendship may matter more.This distinction contributes to the difficulty of defining influential users.
3 Twitter API
Twitter data can be collected through REST, Streaming, and commercial APIs, but access limits and costs constrain network-scale influence analysis.
- API data sources: The REST API provides Twitter data through requests, including user, tweet, and content attributes useful for influence metrics.User attributes include followers, followees, and statuses; tweet data includes creation time, retweets, favorites, hashtags, URLs, and mentions.
- API data sources: REST API requests consume limited tokens within 15-minute windows, delaying large-scale data collection.Token availability varies with authentication and request type, and depleted tokens require waiting for the window to reset.
- API data sources: 15765 tokens, or 262.75 hours in the best case, are required to retrieve the followers of an account with 78,822,185 followers.The estimate uses the followers/ids request, which returns at most 5000 identifiers per token.
- API data sources: The Streaming API maintains a persistent real-time connection and supports filters for tweets and other events.Unlike REST, it provides a continuous stream rather than request-by-request access.
- API data sources: Since 2012, the Streaming API has provided access to about 1% of Twitter’s total data, while full historical access through Gnip is costly.The paper reports an annual paid-tweet subscription quote reaching $39,000 at the end of 2015.
4 What is an influential user?
Influence lacks a standard definition, so Twitter studies use heterogeneous criteria and distinguish influence from activity and popularity.
- Conceptualizing influence: There is no agreement on what constitutes an influential user, and new measures therefore use different relevance criteria.The paper notes that influential users may also be described as opinion leaders, innovators, prestigious, authoritative, or topical experts.
- Conceptualizing influence: Influential users are not necessarily users whose individual tweets are influential.Tweet influence may instead be assessed through extension, intensity, content, or sender popularity.
- Conceptualizing influence: Influence behavior can be examined through topic similarity, latency between same-topic tweets, and whether users create or spread topics.These dimensions are used in a framework for observing influence between users.
- Conceptualizing influence: The paper classifies measures into activity, popularity, and influence because these properties have no direct correlation.For example, RT2 is treated as an influence measure, whereas F1 and F3 are treated as popularity measures.
5 Activity measures
Twitter activity measures quantify observable participation, ranging from simple tweet counts to timeline-, graph-, topic-, and interaction-based measures.
- Activity measures: Observable participation includes tweets, retweets, mentions, and replies, while reading without leaving network traces cannot be measured.The paper defines activity as constant and frequent participation during a period, regardless of attention received.
- Activity measures: TweetRank counts a user’s tweets, while Tweet count score combines original tweets and retweets.The paper presents General Activity as a broader sum of visible user actions.
- Activity measures: Signal strength indicates the strength of an author’s topical signal, while authorship measures originality with higher values closer to 1.The topical-signal framework can restrict analysis to tweets about a specific topic and may omit likes.
- Activity measures: Timeline-based measures can require offline analysis because they cannot be calculated immediately.Effective readers are followers who have not yet tweeted on a trending topic when the user posts, measuring speed in addressing new topics.
- Activity measures: ActivityScore uses followers, followees, and tweets over time, while DiscussRank identifies users initiating conversations around topics.DiscussRank uses multigraphs with edges based on followers, retweets, and mentions.
- Activity measures: IP Influence also measures passivity, defined as difficulty being influenced by another user during a period.The measure uses retweets, followers, and followees; the paper reports criticism that it omits followers and does not correlate with URL clicks.
6 Popularity measures
Twitter popularity measures range from simple follower-based counts to topic-sensitive and interaction-based scores. They differ in computational cost, real-time availability, and whether they capture recognition, activity, or celebrity identification.
- Popularity reflects recognition by many users, but celebrities may have popular accounts without being active or influential.Clint Eastwood’s account is given as an example of many followers with no followees or tweets.
- FollowerRank normalizes traditional in-degree, while the TFF measure uses the follower-to-followee ratio F1/F3.These measures rely on follow-up relationships between users.
- Follower and followee counts can differ greatly, motivating adjusted popularity measures that mitigate extreme scale differences.F1 and F3 are available through the Twitter API, so these popularity measures can be computed in real time.
- Reciprocal-relationship adjustments punish spammers but substantially increase computational costs and can restrict computation to offline environments.The added reciprocal(i) metric counts followers who are also followees.
- The Acquaintance family combines how well-known users are with replies, mentions, and retweets directed toward a user.AAI measures how identifiable a user is by considering the users who identify them and is correlated with F1.
- Action-Reaction combines fan loyalty and attention, identifying Twitter celebrities while avoiding the complexity limitations of AAI.It uses conditional probabilities based on replies, mentions, and retweets.
- Starrank applies PageRank to daily mention graphs and considers acceleration of mentions over time rather than explicitly using follow-up relationships.
7 Influence measures
Twitter influence measures define and estimate users’ ability to affect others or spread information using diverse network, interaction, content, temporal, and predictive criteria.
- Influential users are defined as users whose network actions can affect many others or spread information.
- Twitter measures commonly use retweets, mentions, and followers, while some apply closeness or betweenness to passive network topology.
- PageRank-based measures favor users connected to other well-connected users and are applied to graphs of retweets, replies, mentions, or followers.
- The variable-mass influence model represents user score as velocity driven by mentions and moderated by follower count, with positive correlation to URL clicks.
- TrueTop applies weighted eigenvector centrality to retweet, reply, and mention graphs to increase scores for users outside detected sybil regions.
- Topical and predictive frameworks combine network structure, interactions, content, temporal analysis, or supervised learning to identify influential users.
8 Some additional aspects
The survey examines computational trade-offs, ranking correlations, and an illustrative Twitter case study showing that centrality criteria can produce different results.
- Offline algorithms use more network and user information, whereas real-time algorithms trade completeness for speed and current rankings.
- Most surveyed measures run offline; PageRank and HITS require a network graph, while Twitter’s scale and API limits constrain graph construction.
- Twitter measures often lack complexity analyses, although PageRank, simple metrics, and topical verification have distinct computational bounds.
- Spearman’s ρ and Kendall’s τ compare rankings, ranging from −1 for inverse rankings to 1 for equal rankings.
- The Higgs dataset contains 456,626 users, and the experiment formed a 2,842-node, 62,637-edge subgraph from the top-ranked users.
- FollowerRank appeared similar to A Score, while TFF appeared similar to Paradoxical discounted; formal correlation analysis remained future work.
9 Open research problems
Open problems include developing comprehensive, efficient classifications and correlations of Twitter measures, alongside broader studies of influential tweets.
- An open problem is defining an efficient measure that covers all types of Twitter metrics and can generalize to additional actions such as likes.
- The survey argues that classifying existing measures may be more useful than creating new ones without new ranking criteria.
- A comprehensive correlation of the surveyed measures is needed to identify similarities, differences, redundancies, and independencies.
- Influence detection can also target minimal influencer sets whose removal causes network collapse through optimal percolation.
- Future surveys could examine influential tweets using tweet content quality and influence over readers.
10 Conclusions
The survey classifies Twitter activity, popularity, and influence measures, revealing substantial diversity and no consensus on identifying influential actors. It also highlights PageRank’s prevalence, limited reproducibility for some predictive measures, and computational trade-offs.
- The survey classifies Twitter measures into activity, popularity, and influence categories, with especially diverse criteria across the literature.
- Popularity measures mostly use follow relationships, activity measures mostly use replies, and influence measures prominently use retweets.Favorites or likes are the least-used metric type in the surveyed measures.
- Metrics F2, F4, F5, and F6 considerably increase response time because they require precomputing implicit follow-up relationships.They may suit complete, fixed event databases, but the survey recommends avoiding them when data must be updated.
- Almost one half of existing influence measures use the PageRank algorithm, while predictive influence measures remain few.Some predictive algorithms are not explicitly specified and therefore cannot be reused by other researchers.
- The survey’s criteria diversity indicates that influence in social networks has not yet reached a consensus.