ArticlePDF Available

Online Human-Bot Interactions: Detection, Estimation, and Characterization


Abstract and Figures

Increasing evidence suggests that a growing amount of social media content is generated by autonomous entities known as social bots. In this work we present a framework to detect such entities on Twitter. We leverage more than a thousand features extracted from public data and meta-data about users: friends, tweet content and sentiment, network patterns, and activity time series. We benchmark the classification framework by using a publicly available dataset of Twitter bots. This training data is enriched by a manually annotated collection of active Twitter users that include both humans and bots of varying sophistication. Our models yield high accuracy and agreement with each other and can detect bots of different nature. Our estimates suggest that between 9% and 15% of active Twitter accounts are bots. Characterizing ties among accounts, we observe that simple bots tend to interact with bots that exhibit more human-like behaviors. Analysis of content flows reveals retweet and mention strategies adopted by bots to interact with different target groups. Using clustering analysis, we characterize several subclasses of accounts, including spammers, self promoters, and accounts that post content from connected applications.
Content may be subject to copyright.
Online Human-Bot Interactions: Detection, Estimation, and Characterization
Onur Varol,1,* Emilio Ferrara,2Clayton A. Davis,1Filippo Menczer,1Alessandro Flammini1
1Center for Complex Networks and Systems Research, Indiana University, Bloomington, US
2Information Sciences Institute, University of Southern California, Marina del Rey, CA, US
Increasing evidence suggests that a growing amount of social
media content is generated by autonomous entities known
as social bots. In this work we present a framework to de-
tect such entities on Twitter. We leverage more than a thou-
sand features extracted from public data and meta-data about
users: friends, tweet content and sentiment, network patterns,
and activity time series. We benchmark the classification
framework by using a publicly available dataset of Twitter
bots. This training data is enriched by a manually annotated
collection of active Twitter users that include both humans
and bots of varying sophistication. Our models yield high ac-
curacy and agreement with each other and can detect bots of
different nature. Our estimates suggest that between 9% and
15% of active Twitter accounts are bots. Characterizing ties
among accounts, we observe that simple bots tend to interact
with bots that exhibit more human-like behaviors. Analysis of
content flows reveals retweet and mention strategies adopted
by bots to interact with different target groups. Using cluster-
ing analysis, we characterize several subclasses of accounts,
including spammers, self promoters, and accounts that post
content from connected applications.
Social media are powerful tools connecting millions of peo-
ple across the globe. These connections form the substrate
that supports information dissemination, which ultimately
affects the ideas, news, and opinions to which we are ex-
posed. There exist entities with both strong motivation and
technical means to abuse online social networks — from in-
dividuals aiming to artificially boost their popularity, to or-
ganizations with an agenda to influence public opinion. It
is not difficult to automatically target particular user groups
and promote specific content or views (Ferrara et al. 2016a;
Bessi and Ferrara 2016). Reliance on social media may
therefore make us vulnerable to manipulation.
Social bots are accounts controlled by software, algo-
rithmically generating content and establishing interactions.
Many social bots perform useful functions, such as dis-
semination of news and publications (Lokot and Diakopou-
los 2016; Haustein et al. 2016) and coordination of vol-
unteer activities (Savage, Monroy-Hernandez, and H ¨
2016). However, there is a growing record of malicious ap-
plications of social bots. Some emulate human behavior to
manufacture fake grassroots political support (Ratkiewicz
et al. 2011), promote terrorist propaganda and recruit-
ment (Berger and Morgan 2015; Abokhodair, Yoo, and Mc-
Donald 2015; Ferrara et al. 2016c), manipulate the stock
market (Ferrara et al. 2016a), and disseminate rumors and
conspiracy theories (Bessi et al. 2015).
A growing body of research is addressing social bot ac-
tivity, its implications on the social network, and the de-
tection of these accounts (Lee, Eoff, and Caverlee 2011;
Boshmaf et al. 2011; Beutel et al. 2013; Yang et al. 2014;
Ferrara et al. 2016a; Chavoshi, Hamooni, and Mueen 2016).
The magnitude of the problem was underscored by a Twit-
ter bot detection challenge recently organized by DARPA to
study information dissemination mediated by automated ac-
counts and to detect malicious activities carried out by these
bots (Subrahmanian et al. 2016).
Contributions and Outline
Here we demonstrate that accounts controlled by soft-
ware exhibit behaviors that reflects their intents and modus
operandi (Bakshy et al. 2011; Das et al. 2016), and that such
behaviors can be detected by supervised machine learning
techniques. This paper makes the following contributions:
We propose a framework to extract a large collection
of features from data and meta-data about social media
users, including friends, tweet content and sentiment, net-
work patterns, and activity time series. We use these fea-
tures to train highly-accurate models to identify bots. For
a generic user, we produce a [0,1] score representing the
likelihood that the user is a bot.
The performance of our detection system is evaluated
against both an existing public dataset and an additional
sample of manually-annotated Twitter accounts collected
with a different strategy. We enrich the previously-trained
models using the new annotations, and investigate the ef-
fects of different datasets and classification models.
We classify a sample of millions of English-speaking ac-
tive users. We use different models to infer thresholds in
the bot score that best discriminate between humans and
bots. We estimate that the percentage of Twitter accounts
exhibiting social bot behaviors is between 9% and 15%.
We characterize friendship ties and information flow be-
tween users that show behaviors of different nature: hu-
man and bot-like. Humans tend to interact with more
arXiv:1703.03107v1 [cs.SI] 9 Mar 2017
human-like accounts than bot-like ones, on average. Reci-
procity of friendship ties is higher for humans. Some bots
target users more or less randomly, others can choose tar-
gets based on their intentions.
Clustering analysis reveals certain specific behavioral
groups of accounts. Manual investigation of samples ex-
tracted from each cluster points to three distinct bot
groups: spammers, self promoters, and accounts that post
content from connected applications.
Bot Detection Framework
In the next section, we introduce a Twitter bot detection
framework ( that
is freely available online. This system leverages more than
one thousand features to evaluate the extent to which a Twit-
ter account exhibits similarity to the known characteristics
of social bots (Davis et al. 2016).
Feature Extraction
Data collected using the Twitter API are distilled in 1,150
features in six different classes. The classes and types of fea-
tures are reported in Table 1 and discussed next.
User-based features. Features extracted from user meta-
data have been used to classify users and patterns be-
fore (Mislove et al. 2011; Ferrara et al. 2016a). We ex-
tract user-based features from meta-data available through
the Twitter API. Such features include the number of friends
and followers, the number of tweets produced by the users,
profile description and settings.
Friends features. Twitter actively fosters inter-
connectivity. Users are linked by follower-friend (fol-
lowee) relations. Content travels from person to person via
retweets. Also, tweets can be addressed to specific users
via mentions. We consider four types of links: retweeting,
mentioning, being retweeted, and being mentioned. For
each group separately, we extract features about language
use, local time, popularity, etc. Note that, due to Twitter’s
API limits, we do not use follower/followee information
beyond these aggregate statistics.
Network features. The network structure carries crucial
information for the characterization of different types of
communication. In fact, the usage of network features sig-
nificantly helps in tasks like political astroturf detection
(Ratkiewicz et al. 2011). Our system reconstructs three types
of networks: retweet, mention, and hashtag co-occurrence
networks. Retweet and mention networks have users as
nodes, with a directed link between a pair of users that fol-
lows the direction of information spreading: toward the user
retweeting or being mentioned. Hashtag co-occurrence net-
works have undirected links between hashtag nodes when
two hashtags occur together in a tweet. All networks are
weighted according to the frequency of interactions or co-
occurrences. For each network, we compute a set of fea-
tures, including in- and out-strength (weighted degree) dis-
tributions, density, and clustering. Note that out-degree and
out-strength are measures of popularity.
Temporal features. Prior research suggests that the tem-
poral signature of content production and consumption may
reveal important information about online campaigns and
their evolution (Ghosh, Surachawala, and Lerman 2011;
Ferrara et al. 2016b; Chavoshi, Hamooni, and Mueen 2016).
To extract this signal we measure several temporal features
related to user activity, including average rates of tweet pro-
duction over various time periods and distributions of time
intervals between events.
Content and language features. Many recent papers have
demonstrated the importance of content and language fea-
tures in revealing the nature of social media conversa-
tions (Danescu-Niculescu-Mizil et al. 2013; McAuley and
Leskovec 2013; Mocanu et al. 2013; Botta, Moat, and Preis
2015; Letchford, Moat, and Preis 2015; Das et al. 2016).
For example, deceiving messages generally exhibit informal
language and short sentences (Briscoe, Appling, and Hayes
2014). Our system does not employ features capturing the
quality of tweets, but collects statistics about length and en-
tropy of tweet text. Additionally, we extract language fea-
tures by applying the Part-of-Speech (POS) tagging tech-
nique, which identifies different types of natural language
components, or POS tags. Tweets are therefore analyzed to
study how POS tags are distributed.
Sentiment features. Sentiment analysis is a powerful tool
to describe the emotions conveyed by a piece of text, and
more broadly the attitude or mood of an entire conversa-
tion. Sentiment extracted from social media conversations
has been used to forecast offline events including financial
market fluctuations (Bollen, Mao, and Zeng 2011), and is
known to affect information spreading (Mitchell et al. 2013;
Ferrara and Yang 2015). Our framework leverages sev-
eral sentiment extraction techniques to generate various
sentiment features, including arousal,valence and domi-
nance scores (Warriner, Kuperman, and Brysbaert 2013),
happiness score (Kloumann et al. 2012), polarization and
strength (Wilson, Wiebe, and Hoffmann 2005), and emoti-
con score (Agarwal et al. 2011).
Model Evaluation
To train our system we initially used a publicly available
dataset consisting of 15K manually verified Twitter bots
identified via a honeypot approach (Lee, Eoff, and Caver-
lee 2011) and 16K verified human accounts. We collected
the most recent tweets produced by those accounts using the
Twitter Search API. We limited our collection to 200 public
tweets from a user timeline and up to 100 of the most recent
public tweets mentioning that user. This procedure yielded a
dataset of 2.6 million tweets produced by manually verified
bots and 3 million tweets produced by human users.
We benchmarked our system using several off-the-shelf
algorithms provided in the scikit-learn library (Pedregosa et
al. 2011). In a generic evaluation experiment, the classifier
under examination is provided with numerical vectors, each
describing the features of an account. The classifier returns a
numerical score in the unit interval. A higher score indicates
a stronger belief that the account is a bot. A model’s accu-
racy is evaluated by measuring the Area Under the receiver
Table 1: List of 1150 features extracted by our framework.
User meta-data
Screen name length
(***) Happiness scores of aggregated tweets
Number of digits in screen name (***) Valence scores of aggregated tweets
User name length (***) Arousal scores of aggregated tweets
Time offset (sec.) (***) Dominance scores of single tweets
Default profile (binary) (*) Happiness score of single tweets
Default picture (binary) (*) Valence score of single tweets
Account age (days) (*) Arousal score of single tweets
Number of unique profile descriptions (*) Dominance score of single tweets
(*) Profile description lengths (*) Polarization score of single tweets
(*) Number of friends distribution (*) Entropy of polarization scores of single tweets
(*) Number of followers distribution (*) Positive emoticons entropy of single tweets
(*) Number of favorites distribution (*) Negative emoticons entropy of single tweets
Number of friends (signal-noise ratio and rel. change) (*) Emoticons entropy of single tweets
Number of followers (signal-noise ratio and rel. change) (*) Positive and negative score ratio of single tweets
Number of favorites (signal-noise ratio and rel. change) (*) Number of positive emoticons in single tweets
Number of tweets (per hour and total) (*) Number of negative emoticons in single tweets
Number of retweets (per hour and total) (*) Total number of emoticons in single tweets
Number of mentions (per hour and total) Ratio of tweets that contain emoticons
Number of replies (per hour and total)
Number of retweeted (per hour and total)
Friends ()
Number of distinct languages
Network ()
Number of nodes
Entropy of language use Number of edges (also for reciprocal)
(*) Account age distribution (*) Strength distribution
(*) Time offset distribution (*) In-strength distribution
(*) Number of friends distribution (*) Out-strength distribution
(*) Number of followers distribution Network density (also for reciprocal)
(*) Number of tweets distribution (*) Clustering coeff. (also for reciprocal)
(*) Description length distribution
Fraction of users with default profile and default picture
(*,**) Frequency of POS tags in a tweet
(*) Time between two consecutive tweets
(*,**) Proportion of POS tags in a tweet (*) Time between two consecutive retweets
(*) Number of words in a tweet (*) Time between two consecutive mentions
(*) Entropy of words in a tweet
We consider four types of connected users: retweeting, mentioning, retweeted, and mentioned.
We consider three types of network: retweet, mention, and hashtag co-occurrence networks.
*Distribution types. For each distribution, the following eight statistics are computed and used as individual fea-
tures: min, max, median, mean, std. deviation, skewness, kurtosis, and entropy.
** Part-Of-Speech (POS) tag. There are nine POS tags: verbs, nuns, adjectives, modal auxiliaries, pre-determiners,
interjections, adverbs, wh-, and pronouns.
*** For each feature, we compute mean and std. deviation of the weighted average across words in the lexicon.
operating characteristic Curve (AUC) with 5-fold cross val-
idation, and computing the average AUC score across the
folds using Random Forests, AdaBoost, Logistic Regression
and Decision Tree classifiers. The best classification perfor-
mance of 0.95 AUC was obtained by the Random Forest al-
gorithm. In the rest of the paper we use the Random Forest
model trained using 100 estimators and the Gini coefficient
to measure the quality of splits.
Large-Scale Evaluation
We realistically expect that the nature and sophistication of
bots evolves over time and changes in specific conversa-
tional domains. It is therefore important to determine how
reliable and consistent are the predictions produced by a
system trained on a dataset but tested on different data (in
the wild). Also, the continuously-evolving nature of bots
dictates the need to constantly update the models based on
newly available training data.
To obtain an updated evaluation of the accuracy of our
model, we constructed an additional, manually-annotated
collection of Twitter user accounts. We hypothesize that this
recent collection includes some bots that are more sophisti-
cated than the ones obtained years earlier with the honeypot
method. We leveraged these manual annotations to evalu-
ate the model trained using the honeypot dataset and then
to update the classifier’s training data, producing a merged
dataset to train a new model that ensures better generaliza-
tion to more sophisticated accounts. User IDs and annota-
tion labels in our extended dataset are publicly available
Data Collection
Our data collection focused on active users producing con-
tent in English, as inferred from profile meta-data. We iden-
tified active users by monitoring a large Twitter stream,
accounting for approximately 10% of public tweets, for 3
months starting in October 2015. Using this large stream
avoids known biases introduced by the public Twitter
API (Morstatter et al. 2013). Sampling from the public
stream allows us to focus on active users while avoiding the
biases of other methods such as snowball and breadth-first
sampling (Gjoka et al. 2010), which rely on the selection of
an initial group of users. Number of accounts in our collec-
tion, which consists of almost 53M English-speaking users.
This is roughly consistent with Twitter’s 2015 estimate of
67M active users in the US. The Twitter estimate is based on
a different criterion, namely users who log into the service at
least once in a month (,
and may include non-English speakers; about 13% of the
US population is Spanish-speaking according to census data.
We focus on English speaking users as they represent the
largest group on Twitter (Mocanu et al. 2013).
To restrict our sample to recently active users, we intro-
duce the further criteria that they must have produced at least
200 tweets in total and 90 tweets during the three-month ob-
servation window (one per day on average). From our orig-
inal sample, 14 million user accounts meet both criteria.For
each of these accounts, we collected their tweets through
the Twitter Search API. We restricted the collection to the
most recent 200 tweets and 100 mentions of each user, as
described earlier. Owing to Twitter API limits, this greatly
improved our data collection speed. This choice also reduces
the response time of our service and API. However the lim-
itation adds noise to the features, due to the scarcity of data
available to compute them.
Manual Annotations
We computed classification scores for each of the active ac-
counts using our initial classifier trained on the honeypot
dataset. We then grouped accounts by their bot scores, allow-
ing us to evaluate our system across the spectrum of human
and bot accounts without being biased by the distribution of
bot scores. We randomly sampled 300 accounts from each
bot-score decile. The resulting balanced set of 3000 accounts
were manually annotated by inspecting their public Twitter
profiles. Some accounts have obvious flags, such as using a
stock profile image or retweeting every message of another
account within seconds. In general, however, there is no sim-
ple set of rules to assess whether an account is human or bot.
With the help of four volunteers, we analyzed profile appear-
ance, content produced and retweeted, and interactions with
other users in terms of retweets and mentions. Annotators
were not given a precise set of instructions to perform the
classification task, but rather shown a consistent number of
both positive and negative examples. The final decisions re-
flect each annotator’s opinion and are restricted to: human,
bot, or undecided. Accounts labeled as undecided were elim-
inated from further analysis.
We annotated all 3000 accounts. We will refer to this set
of accounts as the manually annotated data set. Each anno-
tator was assigned a random sample of accounts from each
decile. We enforced a minimum 10% overlap between an-
notations to assess the reliability of each annotator. This
yielded an average pairwise agreement of 75% and moder-
ate inter-annotator agreement (Cohen’s κ= 0.41). We also
computed the agreement between annotators and classifier
outcomes, assuming that a classification score above 0.5 is
interpreted as a bot. This resulted in an average pairwise
agreement of 79% and a moderately high Cohen’s κ= 0.5.
These results suggest high confidence in the annotation pro-
cess, as well as in the agreement between annotations and
model predictions.
Evaluating Models Using Annotated Data
To evaluate our classification system trained on the honeypot
dataset, we examined the classification accuracy separately
for each bot-score decile of the manually annonated dataset.
We achieved classification accuracy greater than 90% for
the accounts in the (0.0,0.4) range, which includes mostly
human accounts. We also observe accuracy above 70% for
scores in the (0.8,1.0) range (mostly bots). Accuracy for ac-
counts in the grey-area range (0.4,0.8) fluctuates between
60% and 80%. Intuitively, this range contains the most chal-
lenging accounts to label, as reflected also in the low inter-
annotators overlap in this region. When the accuracy of each
bin is weighted by the population density in the large dataset
from which the manually annonated has been extracted, we
obtain 86% overall classification accuracy.
We also compare annotator agreement scores for the ac-
counts in each bot-score decile. We observe that agreement
scores are higher for accounts in the (0.0,0.4) range and
lower for accounts in the (0.8,1.0) range, indicating that it
is more difficult for human annotators to identify bot-like as
opposed to human-like behavior.
We observe a similar pattern for the amount of time re-
quired on average to annotate human and bot accounts. An-
notators employed on average 33 seconds to label human
accounts and 37 seconds for bot accounts.
Fig. 1 shows the results of experiments designed to in-
vestigate our ability to detect manually annotated bots. The
baseline ROC curve is obtained by testing the honeypot
model on the manually annotated dataset. Unsurprisingly,
the baseline accuracy (0.85 AUC) is lower than that obtained
cross-validating on the honeypot data (0.95 AUC), because
the model is not trained on the newer bots.
Dataset Effect on Model Accuracy
We can update our models by combining the manually-
annotated and honeypot datasets. We created multiple bal-
anced datasets and performed 5-fold cross-validation to
evaluate the accuracy of the corresponding models:
Annotation: We trained and tested a model by only using
annotated accounts and labels assigned by the majority of
annotators. This yields 0.89 AUC, a reasonable accuracy
considering that the dataset contains recent and possibly
sophisticated bots.
Merged: We merged the honeypot and annotation
datasets for training and testing. The resulting classifier
achieves 0.94 AUC, only slightly worse than the honey-
pot (training and test) model although the merged dataset
contains a variety of more recent bots.
Figure 1: ROC curves of models trained and tested on dif-
ferent datasets. Accuracy is measured by AUC.
Figure 2: Distribution of classifier score for human and bot
accounts in the two datasets.
Mixture: Using mixtures with different ratios of accounts
from the manually annotated and honeypot datasets, we
obtain an accuracy ranging between 0.90 and 0.94 AUC.
In Fig 2, we plot the distributions of classification scores
for human and bot accounts according to each dataset. The
mixture model trained on 2K annotated and 10K honeypot
accounts is used to compute the scores. Human accounts
in both datasets have similar distributions, peaked around
0.1. The difference between bots in the two datasets is more
prominent. The distribution of simple, honeypot bots peaks
around 0.9. The newer bots from the manually annotated
dataset have typically smaller scores, with a distribution
peaked around 0.6. They are more sophisticated, and ex-
hibit characteristics more similar to human behavior. This
raises the issue of how to properly set a threshold on the
score when a strictly binary classification between human
and bots is needed. To infer a suitable threshold, we com-
pute classification accuracies for varying thresholds consid-
ering all accounts scoring below each threshold as human,
and then select the threshold that maximizes accuracy.
We compared scores for accounts in the manually anno-
tated dataset by pairs of models (i.e. trained with different
Figure 3: Comparison of scores for different models. Each
account is represented as a point in the scatter plot with a
color determined by its category. Test points are randomly
sampled from our large-scale collection. Pearson correla-
tions between scores are also reported, along with estimated
thresholds and corresponding accuracies.
mixtures) for labeled human, bot, and a random subset of
accounts (Fig. 3). As expected, both models assign lower
scores for humans and higher for bots. High correlation co-
efficients indicate agreement between the models.
Feature Importance Analysis
To compare the usefulness of different features, we trained
models using each class of features alone. We achieved the
best performance with user meta-data features; content fea-
tures are also effective. Both yielded AUC above 0.9. Other
feature classes yielded AUC above 0.8.
We analyzed the importance of single features using the
Gini impurity score produced by our Random Forests model.
To rank the top features for a given dataset, we randomly se-
lect a subset of 10,000 accounts and compute the top features
across 100 randomized experiments. The top 10 features are
sufficient to reach performance of 0.9 AUC. Sentiment and
content of mentioned tweets are important features along
with the statistical properties of retweet networks. Features
of the friends with whom a user interacts are strong predic-
tors as well. We observed the redundancy among many cor-
related features, such as distribution-type features (cf. Ta-
ble 1), especially in the content and sentiment categories.
Further analysis of feature importance is the subject of on-
going investigation.
False Positive and False Negative Cases
Neither human annotators nor machine-learning models per-
form flawlessly. Humans are better at generalizing and learn-
ing new features from observed data. Machines outperform
human annotators at processing large numbers of relations
and searching for complex patterns. We analyzed our an-
notated accounts and their bot scores to highlight when
disagreement occurs between annotators and classification
models. Using an optimal threshold, we measured false pos-
itive and false negative rates at 0.15 and 0.11 respectively in
our extended dataset. In these experiments, human annota-
tion is considered as ground truth.
We identified the cases when the disagreement between
classifier score and annotations occurs. We manually exam-
ined a sample from these accounts to investigate these er-
rors. Accounts annotated as human can be classified as bot
when an account posts tweets created by connected appli-
cations from other platforms. Some unusually active users
are also classified as bots. Those users tend to have more
retweets in general. This is somewhat intuitive as retweet-
ing has lower cost than creating new content. We encoun-
tered examples of misclassification for organizational and
promotional accounts. Such accounts are often operated by
multiple individuals, or combinations of users and automatic
tools, generating misleading cues for the classifiers. Finally,
the language of the content can also cause errors: our models
tend to assign high bot scores to users who tweet in multiple
languages. To mitigate this problem, the public version of
our system now includes a classifier that ignores language-
dependent features.
Estimation of Bot Population
In a 2014 report by Twitter to the US Securities and Ex-
change Commission, the company put forth an estimate that
between 5% and 8.5% of their user base consists of bots.1
We would like to offer our own assessment of the propor-
tion of bot accounts as measured with our approach. Since
our framework provides a continuous bot score as opposed
to a discrete bot/human judgement, we must first determine
an appropriate bot-score threshold separating human and bot
accounts to estimate the proportion of bot accounts.
To infer a suitable threshold, we computed classification
accuracies for varying thresholds considering all accounts
scoring below each threshold as human. We then selected the
threshold yielding maximum accuracy (see insets of Fig. 4).
We estimated the population of bots using different mod-
els. This approach allows us to identify lower and upper
bounds for the prevalence of Twitter bots. Models trained us-
ing the annotated dataset alone yield estimates of up to 15%
of accounts being bots. Recall that the honeypot dataset was
obtained earlier and therefore does not include newer, more
sophisticated bots. Thus models trained on the honeypot data
alone are less sensitive to these sophisticated bots, yielding a
more conservative estimate of 9%. Mixing the training data
from these two sources results in estimates between these
bounds depending on the ratio of the mixture, as illustrated
in Fig. 4. Taken together, these numbers suggest that esti-
mates about the prevalence of Twitter bots are highly depen-
dent on the definition and sophistication of the bots.
Some other remarks are in order. First, we do not exclude
the possibility that very sophisticated bots can systemati-
cally escape a human annotator’s judgement. These complex
bots may be active on Twitter, and therefore present in our
datasets, and may have been incorrectly labeled as humans,
making even the 15% figure a conservative estimate. Sec-
ond, increasing evidence suggests the presence on social me-
dia of hybrid human-bot accounts (sometimes referred to as
cyborgs) that perform automated actions with some human
supervision (Chu et al. 2012; Clark et al. 2016). Some have
been allegedly used for terrorist propaganda and recruitment
purposes. It remains unclear how these accounts should be
labeled, and how pervasive they are.
Characterization of User Interactions
Let us next characterize social connectivity, information
flow, and shared properties of users. We analyze the cre-
ation of social ties by accounts with different bot scores, and
their interactions through shared content. We also cluster
accounts and investigate shared properties of users in each
cluster. Here and in the remainder of this paper, bot scores
are computed with a model trained on the merged dataset.
Social connectivity
To characterize the social connectivity, we collected the so-
cial networks of the accounts in our dataset using the Twitter
API. Resulting friend and follower relations account for 46
billion social ties, 7 billion of which represent ties between
the initially collected user set.
Our observations on social connectivity are presented in
Fig. 5. We computed bot-score distributions of friends and
followers of accounts for each score interval. The dark line
in the top panel shows that human accounts (low score)
mostly follow other human accounts. The dark line in the
bottom panel shows a principal peak around 0.1 and a sec-
ondary one around 0.5. This indicates that humans are typ-
ically followed by other humans, but also by sophisticated
bots (intermediate scores). The lines corresponding to high
scores in the two panels show that bots tend to follow other
bots and they are mostly followed by bots. However sim-
ple bots (0.8–1.0 ranges) can also attract human attention.
This happens when, e.g., humans follow benign bots such as
those that share news. This gives rise to the secondary peak
of the red line in the bottom panel. In summary, the creation
of social ties leads to a homophily effect.
Fig. 6 illustrates the extent to which connections are re-
ciprocated, given the nature of the accounts forming the ties.
The reciprocity score of a user is defined as the fraction of
friends who are also followers. We observe that human ac-
counts reciprocate more (dark line). Increasing bot scores
correlate with lower reciprocity. We also observe that sim-
ple bot accounts (0.8–1.0 ranges) have bimodal reciprocity
distributions, indicating the existence of two distinct behav-
iors. The majority of high-score accounts have reciprocity
score smaller than 0.2, possibly because simple bots follow
users at random. The slight increase as the reciprocity score
approaches one may be due to botnet accounts that coordi-
nate by following each other.
Information flow
Twitter is a platform that fosters social connectivity and the
broadcasting of popular content. In Fig. 7 we analyze infor-
mation flow in terms of mentions/retweets as a function of
the score of the account being mentioned or retweeted.
Simple bots tend to retweet each other (lines for scores
in the 0.8–1.0 ranges peak around 0.8 in the bottom panel),
while they frequently mention sophisticated bots (peaking
Figure 4: Estimation of bot population obtained from models with different sensitivity to sophisticated bots. The main charts
show the score distributions based on our dataset of 14M users; accounts identified as bots are highlighted. The inset plots show
how the thresholds are computed by maximizing accuracy. The titles of each subplot reflect the number of accounts from the
annotated and honeypot datasets, respectively.
Figure 5: Distributions of bot scores for friends (top) and
followers (bottom) of accounts in different score intervals.
Figure 6: Distribution of reciprocity scores for accounts in
different score intervals.
around 0.5 in the top panel). More sophisticated bots (scores
in the 0.5–0.7 ranges) retweet, but do not mention humans.
They might be unable to engage in meaningful exchanges
with humans. While humans also retweet bots, as they may
Figure 7: Bot score distributions of users mentioned (top)
and retweeted (bottom) by accounts with different scores.
post interesting content (see peaks of the dark lines in the
bottom panel), they have no interest in mentioning bots di-
rectly (dark lines in the top panel).
Clustering accounts
To characterize different account types, let us group ac-
counts into behavioral clusters. We apply K-Means to nor-
malized vectors of the 100 most important features selected
by our Random Forests model. We identify 10 distinct clus-
ters based on different evaluation criteria, such as silhouette
scores and percentage of variance explained. In Fig 8, we
present a 2-dimensional projection of users obtained by a di-
mensionality reduction technique called t-SNE (Maaten and
Hinton 2008). In this method, the similarity between users
is computed based on their 100-dimensional representation
in the feature space. Similar users are projected into nearby
points and dissimilar users are kept distant from each other.
Let us investigate shared cluster properties by manual in-
spection of random subsets of accounts from each cluster.
Figure 8: t-SNE embedding of accounts. Points are colored
based on clustering in high-dimensional space. For each
cluster, the distribution of scores is presented on the right.
Three of the clusters, namely C0C2, have high average bot
scores. The presence of significant amounts of bot accounts
in these clusters was manually verified. These bot clusters
exhibit some prominent properties: cluster C0, for exam-
ple, consists of legit-looking accounts that are promoting
themselves (recruiters, porn actresses, etc.). They are con-
centrated in the lower part of the 2-dimensional embedding,
suggesting homogeneous patterns of behaviors. C1 contains
spam accounts that are very active but have few followers.
Accounts in C2 frequently use automated applications to
share activity from other platforms like YouTube and Insta-
gram, or post links to news articles. Some of the accounts in
C2 might belong to actual humans who are no longer active
and their posts are mostly sent by connected apps.
Cluster C3 contain a mix of sophisticated bots, cyborg-
like accounts (mix of bot and human features), and human
users. Clusters of predominantly human accounts, namely
C4C9, separate from one another in the embedding due to
different activity styles, user popularity, content production
and consumption patterns. For instance, accounts in C7 en-
gage more with their friends, unlike accounts from C8 that
mostly retweet with little other forms of interaction. Clusters
C5,C6, and C9 contain common Twitter users who produce
experiential tweets, share pictures, and retweet their friends.
Related Work
Also known as “sybil” accounts, social bots can pollute on-
line discussion by lending false credibility to their messages
and influence other users (Ferrara et al. 2016a; Aiello et
al. 2012). Recent studies quantify the extent to which au-
tomated systems can dominate discussions on Twitter about
topics ranging from electronic cigarettes (Clark et al. 2015)
to elections (Bessi and Ferrara 2016). Large collections of
social bots, also known as botnets, are controlled by botmas-
ters and used for coordinated activities. Examples of such
botnets identified for advertisement (Echeverr´
ıa and Zhou
2017) and influence about Syrian civic war (Abokhodair,
Yoo, and McDonald 2015). Social bots also vary greatly
in terms of their behavior, intent, and vulnerabilities, as il-
lustrated in a categorization scheme for bot attacks (Mitter,
Wagner, and Strohmaier 2013).
Much of the previous work on detecting bots is from
the perspective of the social network platform operators,
implying full access to all data. These studies focus on
collecting large-scale data to either cluster behavioral pat-
terns of users (Wang et al. 2013a) or classify accounts
using supervised learning techniques (Yang et al. 2014;
Lee, Eoff, and Caverlee 2011). For instance, Beutel et al. de-
composed event data in time, user, and activity dimensions
to extract similar behaviors (Beutel et al. 2013). These tech-
niques are useful to identify coordinated large-scale attacks
directed at a common set of targets at the same time, but
accounts with similar strategies might also target different
groups and operate separately from each other.
Structural connectivity may provide important cues. How-
ever, Yang et al. studied large-scale sybil attacks and ob-
served sophisticated sybils that develop strategies for build-
ing normal-looking social ties, making themselves harder to
detect (Yang et al. 2014). Some sybil attacks analyze the
social graph of targeted groups to infiltrate specific organi-
zations (Elyashar et al. 2013). SybilRank is a system devel-
oped to identify attacks from their underlying topology (Cao
et al. 2012). Alvisi et al. surveyed the evolution of sybil de-
fense protocols that leverage the structural properties of the
social graph (Alvisi et al. 2013).
The work presented here follows several previous contri-
butions to the problem of social bot detection that leverage
learning models trained with data collected from human and
bot accounts. Chu et al. built a classification system identify-
ing accounts controlled by humans, bots, and cyborgs (Chu
et al. 2010; Chu et al. 2012). Wang et al. analyzed sybil
attacks using annotations by experts and crowd-sourcing
workers to evaluate consistency and effectiveness of differ-
ent detection systems (Wang et al. 2013b). Clark et al. la-
beled 1,000 accounts by hand and found natural language
text features to be very effective at discriminating between
human and automated accounts (Clark et al. 2016). Lee et
al. used a honeypot approach to collect the largest sample
of bot accounts available to date (Lee, Eoff, and Caverlee
2011). That study generated the honeypot dataset used in the
present paper. Here, we extend this body of prior work by
exploring many different categories of features, contribut-
ing a new labeled dataset, estimating the number of bot ac-
counts, analyzing information flow among accounts, iden-
tifying several classes of behaviors, and providing a public
bot detection service.
An alternative approach to study social bots and sybil at-
tacks is to understand what makes certain groups and indi-
viduals more appealing as targets. Wald et al. studied the
factors affecting the likelihood of a users being targeted by
social bots (Wald et al. 2013). These approaches point to ef-
fective strategies that future social bots might develop.
Recently, we have observed efforts to facilitate research
collaborations on the topic of social bots. DARPA organized
a bot detection challenge in the domain of anti-vaccine cam-
paigns on Twitter (Subrahmanian et al. 2016). We released
our Twitter bot detection system online for public use (Davis
et al. 2016). Since its release, our system has received mil-
lions of requests and we are improving models based on
feedback we received from our users. The increasing avail-
ability of software and datasets on social bots will help de-
sign systems that are capable of co-evolving with recent so-
cial bots and hopefully mitigating the effects of their mali-
cious activities.
Social media make it easy for accounts controlled by hy-
brid or automated approaches to create content and interact
with other accounts. Our project aims to identify these bots.
Such a classification task could be a first step toward study-
ing modes of communication among different classes of en-
tities on social media.
In this article, we presented a framework for bot detec-
tion on Twitter. We introduced our machine learning sys-
tem that extracts more than a thousand features in six dif-
ferent classes: users and friends meta-data, tweet content
and sentiment, network patterns, and activity time series. We
evaluated our framework when initially trained on an avail-
able dataset of bots. Our initial classifier achieves 0.95 AUC
when evaluated by using 5-fold cross validation. Our analy-
sis on the contributions of different feature classes suggests
that user meta-data and content features are the two most
valuable sources of data to detect simple bots.
To evaluate the performance of our classifier on a more re-
cent and challenging sample of bots, we randomly selected
Twitter accounts covering the whole spectrum of classifi-
cation scores. The accuracy of our initial classifier trained
on the honeypot dataset decreased to 0.85 AUC when tested
on the more challenging dataset. By retraining the classifier
with the two datasets merged, we achieved high accuracy
(0.94 AUC) in detecting both simple and sophisticated bots.
We also estimated the fraction of bots in the ac-
tive English-speaking population on Twitter. We classified
nearly 14M accounts using our system and inferred the op-
timal threshold scores that separate human and bot accounts
for several models with different mixes of simple and so-
phisticated bots. Training data have an important effect on
classifier sensitivity. Our estimates for the bot population
range between 9% and 15%. This points to the importance
of tracking increasingly sophisticated bots, since deception
and detection technologies are in a never-ending arms race.
To characterize user interactions, we studied social
connectivity and information flow between different user
groups. We showed that selection of friends and followers
are correlated with accounts bot-likelihood. We also high-
lighted how bots use different retweet and mention strategies
when interacting with humans or other bots.
We concluded our analysis by characterizing subclasses
of account behaviors. Clusters identified by this analysis
point mainly to three types of bots. These results emphasize
that Twitter hosts a variety of users with diverse behaviors;
this is true for both human and bot accounts. In some cases,
the boundary separating these two groups is not sharp and
an account can exhibit characteristics of both.
Acknowledgments. We thank M. JafariAsbagh, P. Shiralkar
for helpful discussions. We also want to thank undergraduate
students A. Toms, A. Fulton, A. Witulski, and M. Johnston for
contributing data annotation. This work was supported in part
by ONR (N15A-020-0053), DARPA (W911NF-12-1-0037), NSF
(CCF-1101743), and the J.S. McDonnell Foundation.
[Abokhodair, Yoo, and McDonald 2015] Abokhodair, N.; Yoo, D.;
and McDonald, D. W. 2015. Dissecting a social botnet: Growth,
content and influence in twitter. In Proc. of the 18th ACM Conf.
on Computer Supported Cooperative Work & Social Computing,
839–851. ACM.
[Agarwal et al. 2011] Agarwal, A.; Xie, B.; Vovsha, I.; Rambow,
O.; and Passonneau, R. 2011. Sentiment analysis of Twitter data.
In Proc. of the Workshop on Languages in Social Media, 30–38.
[Aiello et al. 2012] Aiello, L.; Deplano, M.; Schifanella, R.; and
Ruffo, G. 2012. People are strange when you’re a stranger: Im-
pact and influence of bots on social networks. In Proc. 6th Intl.
AAAI Conf. on Weblogs & Soc. Media (ICWSM).
[Alvisi et al. 2013] Alvisi, L.; Clement, A.; Epasto, A.; Lattanzi, S.;
and Panconesi, A. 2013. Sok: The evolution of sybil defense via
social networks. In Proc. IEEE Symposium on Security and Privacy
(SP), 382–396.
[Bakshy et al. 2011] Bakshy, E.; Hofman, J. M.; Mason, W. A.; and
Watts, D. J. 2011. Everyone’s an influencer: quantifying influence
on Twitter. In Proc. 4th ACM Intl. Conf. on Web Search and Data
Mining, 65–74.
[Berger and Morgan 2015] Berger, J., and Morgan, J. 2015. The
isis twitter census: Defining and describing the population of isis
supporters on twitter. The Brookings Project on US Relations with
the Islamic World 3:20.
[Bessi and Ferrara 2016] Bessi, A., and Ferrara, E. 2016. Social
bots distort the 2016 us presidential election online discussion.
First Monday 21(11).
[Bessi et al. 2015] Bessi, A.; Coletto, M.; Davidescu, G. A.; Scala,
A.; Caldarelli, G.; and Quattrociocchi, W. 2015. Science vs con-
spiracy: Collective narratives in the age of misinformation. PLoS
ONE 10(2):e0118093.
[Beutel et al. 2013] Beutel, A.; Xu, W.; Guruswami, V.; Palow, C.;
and Faloutsos, C. 2013. Copycatch: stopping group attacks by
spotting lockstep behavior in social networks. In Prov. 22nd Intl.
ACM Conf. World Wide Web (WWW), 119–130.
[Bollen, Mao, and Zeng 2011] Bollen, J.; Mao, H.; and Zeng, X.
2011. Twitter mood predicts the stock market. Journal of Com-
putational Science 2(1):1–8.
[Boshmaf et al. 2011] Boshmaf, Y.; Muslukhov, I.; Beznosov, K.;
and Ripeanu, M. 2011. The socialbot network: when bots social-
ize for fame and money. In Proc. 27th Annual Computer Security
Applications Conf.
[Botta, Moat, and Preis 2015] Botta, F.; Moat, H. S.; and Preis, T.
2015. Quantifying crowd size with mobile phone and twitter data.
Royal Society open science 2(5):150162.
[Briscoe, Appling, and Hayes 2014] Briscoe, E.; Appling, S.; and
Hayes, H. 2014. Cues to deception in social media communi-
cations. In Hawaii Intl. Conf. on Syst Sci.
[Cao et al. 2012] Cao, Q.; Sirivianos, M.; Yang, X.; and Pregueiro,
T. 2012. Aiding the detection of fake accounts in large scale so-
cial online services. In 9th USENIX Symp on Netw Sys Design &
Implement, 197–210.
[Chavoshi, Hamooni, and Mueen 2016] Chavoshi, N.; Hamooni,
H.; and Mueen, A. 2016. Identifying correlated bots in twitter.
In Social Informatics: 8th Intl. Conf., 14–21.
[Chu et al. 2010] Chu, Z.; Gianvecchio, S.; Wang, H.; and Jajodia,
S. 2010. Who is tweeting on twitter: human, bot, or cyborg? In
Proc. 26th annual computer security applications conf., 21–30.
[Chu et al. 2012] Chu, Z.; Gianvecchio, S.; Wang, H.; and Jajodia,
S. 2012. Detecting automation of twitter accounts: Are you a hu-
man, bot, or cyborg? IEEE Tran Dependable & Secure Comput
[Clark et al. 2015] Clark, E.; Jones, C.; Williams, J.; Kurti, A.; Nor-
totsky, M.; Danforth, C.; and Dodds, P. 2015. Vaporous marketing:
Uncovering pervasive electronic cigarette advertisements on twit-
ter. arXiv preprint arXiv:1508.01843.
[Clark et al. 2016] Clark, E.; Williams, J.; Jones, C.; Galbraith, R.;
Danforth, C.; and Dodds, P. 2016. Sifting robotic from organic text:
a natural language approach for detecting automation on twitter.
Journal of Computational Science 16:1–7.
[Danescu-Niculescu-Mizil et al. 2013] Danescu-Niculescu-Mizil,
C.; West, R.; Jurafsky, D.; Leskovec, J.; and Potts, C. 2013. No
country for old members: user lifecycle and linguistic change in
online communities. In Proc. of the 22nd Intl. Conf. on World
Wide Web, 307–318.
[Das et al. 2016] Das, A.; Gollapudi, S.; Kiciman, E.; and Varol,
O. 2016. Information dissemination in heterogeneous-intent net-
works. In Proc. ACM Conf. on Web Science.
[Davis et al. 2016] Davis, C. A.; Varol, O.; Ferrara, E.; Flammini,
A.; and Menczer, F. 2016. BotOrNot: A system to evaluate social
bots. In Proc. 25th Intl. Conf. Companion on World Wide Web,
ıa and Zhou 2017] Echeverr´
ıa, J., and Zhou, S. 2017.
The ‘star wars’ botnet with ¿350k twitter bots. arXiv preprint
[Elyashar et al. 2013] Elyashar, A.; Fire, M.; Kagan, D.; and
Elovici, Y. 2013. Homing socialbots: intrusion on a specific or-
ganization’s employee using socialbots. In Proc. IEEE/ACM Intl.
Conf. on Advances in Social Networks Analysis and Mining, 1358–
[Ferrara and Yang 2015] Ferrara, E., and Yang, Z. 2015. Quantify-
ing the effect of sentiment on information diffusion in social media.
PeerJ Comp. Sci. 1:e26.
[Ferrara et al. 2016a] Ferrara, E.; Varol, O.; Davis, C.; Menczer, F.;
and Flammini, A. 2016a. The rise of social bots. Comm. ACM
[Ferrara et al. 2016b] Ferrara, E.; Varol, O.; Menczer, F.; and Flam-
mini, A. 2016b. Detection of promoted social media campaigns.
In Proc. Intl. AAAI Conference on Web and Social Media.
[Ferrara et al. 2016c] Ferrara, E.; Wang, W.-Q.; Varol, O.; Flam-
mini, A.; and Galstyan, A. 2016c. Predicting online extremism,
content adopters, and interaction reciprocity. In Social Informat-
ics: 8th Intl. Conf., SocInfo 2016, Bellevue, WA, USA, 22–39.
[Ghosh, Surachawala, and Lerman 2011] Ghosh, R.; Surachawala,
T.; and Lerman, K. 2011. Entropy-based classification of retweet-
ing activity on twitter. In Proc. of KDD workshop on Social Net-
work Analysis.
[Gjoka et al. 2010] Gjoka, M.; Kurant, M.; Butts, C. T.; and
Markopoulou, A. 2010. Walking in facebook: A case study of
unbiased sampling of osns. In Proc. IEEE INFOCOM, 1–9.
[Haustein et al. 2016] Haustein, S.; Bowman, T. D.; Holmberg, K.;
Tsou, A.; Sugimoto, C. R.; and Larivi`
ere, V. 2016. Tweets as
impact indicators: Examining the implications of automated “bot”
accounts on twitter. Journal of the Association for Information
Science and Technology 67(1):232–238.
[Kloumann et al. 2012] Kloumann, I. M.; Danforth, C. M.; Harris,
K. D.; Bliss, C. A.; and Dodds, P. S. 2012. Positivity of the english
language. PLoS ONE 7(1):e29484.
[Lee, Eoff, and Caverlee 2011] Lee, K.; Eoff, B. D.; and Caverlee,
J. 2011. Seven months with the devils: A long-term study of con-
tent polluters on twitter. In Proc. 5th AAAI Intl. Conf. on Web and
Social Media.
[Letchford, Moat, and Preis 2015] Letchford, A.; Moat, H. S.; and
Preis, T. 2015. The advantage of short paper titles. Royal Society
Open Science 2(8):150266.
[Lokot and Diakopoulos 2016] Lokot, T., and Diakopoulos, N.
2016. News bots: Automating news and information dissemina-
tion on twitter. Digital Journalism 4(6):682–699.
[Maaten and Hinton 2008] Maaten, L. v. d., and Hinton, G. 2008.
Visualizing data using t-sne. Journal of Machine Learning Re-
search 9(Nov):2579–2605.
[McAuley and Leskovec 2013] McAuley, J., and Leskovec, J.
2013. From amateurs to connoisseurs: modeling the evolution of
user expertise through online reviews. In Proc. 22nd Intl. ACM
Conf. World Wide Web, 897–908.
[Mislove et al. 2011] Mislove, A.; Lehmann, S.; Ahn, Y.-Y.; On-
nela, J.-P.; and Rosenquist, J. N. 2011. Understanding the de-
mographics of Twitter users. In Proc. of the 5th Intl. AAAI Conf.
on Weblogs and Social Media.
[Mitchell et al. 2013] Mitchell, L.; Harris, K. D.; Frank, M. R.;
Dodds, P. S.; and Danforth, C. M. 2013. The geography of happi-
ness: Connecting Twitter sentiment and expression, demographics,
and objective characteristics of place. PLoS ONE 8(5):e64417.
[Mitter, Wagner, and Strohmaier 2013] Mitter, S.; Wagner, C.; and
Strohmaier, M. 2013. A categorization scheme for socialbot attacks
in online social networks. In Proc. of the 3rd ACM Web Science
[Mocanu et al. 2013] Mocanu, D.; Baronchelli, A.; Perra, N.;
Gonc¸alves, B.; Zhang, Q.; and Vespignani, A. 2013. The Twitter
of Babel: Mapping world languages through microblogging plat-
forms. PLoS ONE 8(4):e61981.
[Morstatter et al. 2013] Morstatter, F.; Pfeffer, J.; Liu, H.; and Car-
ley, K. 2013. Is the sample good enough? comparing data from
twitter’s streaming api with twitter’s firehose. In 7th Int Conf on
Weblogs & Soc Med.
[Pedregosa et al. 2011] Pedregosa, F.; Varoquaux, G.; Gramfort, A.;
Michel, V.; Thirion, B.; Grisel, O.; et al. 2011. Scikit-learn: Ma-
chine learning in Python. Journal of Machine Learning Research
[Ratkiewicz et al. 2011] Ratkiewicz, J.; Conover, M.; Meiss, M.;
Goncalves, B.; Flammini, A.; and Menczer, F. 2011. Detecting
and tracking political abuse in social media. In 5th Int Conf on
Weblogs & Soc Med, 297–304.
[Savage, Monroy-Hernandez, and H¨
ollerer 2016] Savage, S.;
Monroy-Hernandez, A.; and H¨
ollerer, T. 2016. Botivist: Calling
volunteers to action using online bots. In Proceedings of the 19th
ACM Conference on Computer-Supported Cooperative Work &
Social Computing, 813–822. ACM.
[Subrahmanian et al. 2016] Subrahmanian, V.; Azaria, A.; Durst,
S.; Kagan, V.; Galstyan, A.; Lerman, K.; Zhu, L.; Ferrara, E.; Flam-
mini, A.; Menczer, F.; et al. 2016. The DARPA Twitter Bot Chal-
lenge. IEEE Computer 6(49):38–46.
[Wald et al. 2013] Wald, R.; Khoshgoftaar, T. M.; Napolitano, A.;
and Sumner, C. 2013. Predicting susceptibility to social bots on
twitter. In Proc. 14th Intl. IEEE Conf. on Information Reuse and
Integration, 6–13.
[Wang et al. 2013a] Wang, G.; Konolige, T.; Wilson, C.; Wang, X.;
Zheng, H.; and Zhao, B. Y. 2013a. You are how you click: Click-
stream analysis for sybil detection. In Proc. USENIX Security, 1–
15. Citeseer.
[Wang et al. 2013b] Wang, G.; Mohanlal, M.; Wilson, C.; Wang,
X.; Metzger, M.; Zheng, H.; and Zhao, B. Y. 2013b. Social turing
tests: Crowdsourcing sybil detection. In Proc. of the 20th Network
& Distributed System Security Symposium (NDSS).
[Warriner, Kuperman, and Brysbaert 2013] Warriner, A. B.; Kuper-
man, V.; and Brysbaert, M. 2013. Norms of valence, arousal, and
dominance for 13,915 english lemmas. Behavior research methods
[Wilson, Wiebe, and Hoffmann 2005] Wilson, T.; Wiebe, J.; and
Hoffmann, P. 2005. Recognizing contextual polarity in phrase-
level sentiment analysis. In ACL Conf on Human Language Techn
& Empirical Methods in NLP, 347–354.
[Yang et al. 2014] Yang, Z.; Wilson, C.; Wang, X.; Gao, T.; Zhao,
B. Y.; and Dai, Y. 2014. Uncovering social network sybils in the
wild. ACM Trans. Knowledge Discovery from Data 8(1):2.
... Social bots imitate human behaviors on social networks such as Twitter, Facebook, Instagram, etc. [43]. Millions of bots, typically controlled by automated programs or platform APIs [1], attempt to sneak into genuine users as a disguise to pursue malicious goals such as actively engaging in election interference [11,17], misinformation dissemination [8], and privacy attacks [37]. Bots are also involved in spreading extreme ideologies [3,18], posing threats to online communities. ...
Social bot detection is of paramount importance to the resilience and security of online social platforms. The state-of-the-art detection models are siloed and have largely overlooked a variety of data characteristics from multiple cross-lingual platforms. Meanwhile, the heterogeneity of data distribution and model architecture makes it intricate to devise an efficient cross-platform and cross-model detection framework. In this paper, we propose FedACK, a new federated adversarial contrastive knowledge distillation framework for social bot detection. We devise a GAN-based federated knowledge distillation mechanism for efficiently transferring knowledge of data distribution among clients. In particular, a global generator is used to extract the knowledge of global data distribution and distill it into each client's local model. We leverage local discriminator to enable customized model design and use local generator for data enhancement with hard-to-decide samples. Local training is conducted as multi-stage adversarial and contrastive learning to enable consistent feature spaces among clients and to constrain the optimization direction of local models, reducing the divergences between local and global models. Experiments demonstrate that FedACK outperforms the state-of-the-art approaches in terms of accuracy, communication efficiency, and feature space consistency.
... . Bot accounts have also been leveraged to promote extremism by adopting terrorism ideologies [12,32,46,65,89]. Several security threats and attacks such as worms, phishing links, and hijacking have been attributed to the bot accounts [37,72,78,82,95,107]. Reports have suggested that roughly 15% of all active accounts on Twitter are bot accounts [65]. ...
Full-text available
With the recognition and influence of Twitter on modern society, an enormous amount of multimedia information is regularly generated and rapidly disseminated on the platform. These characteristics have caught the attention of automated accounts called bots that are frequently leveraged for malevolent activities. From distorting the political elections to crashing the stock markets to spreading conspiracy theories and fake news, various bot accounts have become a source of grave concern. Particularly, spambots have been known to mimic the behaviour of a legitimate user, making them almost insurmountable to detect. Of late, DNA inspired behaviour encoding of Twitter accounts such as B3Type, and B3Content DNA has achieved promising results in detecting the social spambots. However, the evolving nature of spambots drives academia and the industries to devise adaptive strategies to keep pace with the progressing capabilities of these spambots and curtail the menace caused by them. Therefore, this study proposes a hybrid technique utilizing digital DNA as a base approach and augmenting it with the state-of-the-art BERT model pre-trained on the sentiment classification task. The proposed hybrid encoding is termed B3Sentiment DNA. Further, the study extends B3Content encoding and proposes B5Content DNA encoding to make spambot detection more robust. B5Content encoding achieved an accuracy of 75.23%, surpassing B3Content encoding, while the proposed hybrid approach, B3Sentiment DNA, achieved an accuracy of 85.79%, significantly outperforming all the DNA encoding techniques considered in this study.
... Social bots spread large amounts of information to achieve their purpose: supporting political parties, attacking opponents, suppressing dissidents, and creating differences [22]. Studies have shown that social bots account for 9-15% of active Twitter accounts [24]. In some specific topics, the proportion of published content can even reach more than 50% [25]. ...
Full-text available
Social bots have already infiltrated social media platforms, such as Twitter, Facebook, and so on. Exploring the role of social bots in discussions of the COVID-19 pandemic, as well as comparing the behavioral differences between social bots and humans, is an important foundation for studying public health opinion dissemination. We collected data on Twitter and used Botometer to classify users into social bots and humans. Machine learning methods were used to analyze the characteristics of topic semantics, sentiment attributes, dissemination intentions, and interaction patterns of humans and social bots. The results show that 22% of these accounts were social bots, while 78% were humans, and there are significant differences in the behavioral characteristics between them. Social bots are more concerned with the topics of public health news than humans are with individual health and daily lives. More than 85% of bots’ tweets are liked, and they have a large number of followers and friends, which means they have influence on internet users’ perceptions about disease transmission and public health. In addition, social bots, located mainly in Europe and America countries, create an “authoritative” image by posting a lot of news, which in turn gains more attention and has a significant effect on humans. The findings contribute to understanding the behavioral patterns of new technologies such as social bots and their role in the dissemination of public health information.
The paper considers the ability to describe malicious bots using their characteristics, which can be the basis for building models for recognising bot parameters and qualitatively analysing attack characteristics in social networks. The following metrics are proposed using the characteristics of VKontakte social network bots as an example: trust, survivability, price, seller type, speed, and expert quality. To extract these metrics, an approach is proposed that is based on the methods of test purchases and the Turing test. The main advantage of this approach is that it proposes to extract features from the data obtained experimentally, thereby obtaining a more reasonable estimation than the expert approach. Also, an experiment on extracting metrics from malicious bots of the VKontakte social network using the proposed approach is described, and an analysis of the metrics' dependence is carried out. The experiment demonstrates the possibility of metrics extracting and analysis. In general, the proposed metrics and the approach to their extraction can become the basis for the transition from binary attack detection in social networks to a qualitative description of the attacker and his capabilities, as well as an analysis of the evolution of bots.
Nowadays, social media platforms are thronged with social bots spreading misinformation. Twitter has become the hotspot for social bots. These bots are either automated or semi-automated, spreading misinformation purposefully or not purposefully is influencing society’s perspective on different aspects of life. This tremendous increase in social bots has aroused huge interest in researchers. In this paper, we have proposed a social bot detection model using Random Forest Classifier, we also used Extreme Gradient Boost Classifier, Artificial Neural Network, and Decision Tree Classifier on the top 8 attributes, which are staunch. The attribute is selected after analyzing the preprocessed data set taken from Kaggle which contains 37446 Twitter accounts having both human and bots. The overall accuracy of the proposed model is above 83%. The result demonstrated that the model is feasible for high-accuracy social bot detection.KeywordsSocial botsBot detectionFeature selectionRandom forest classifierXGBoostANNDecision tree classifier
The number of malicious bots is increasing rapidly with the growing popularity of social media. We evaluate the importance of 19 commonly used features for Twitter bot detection. Our goal is to propose a set of minimal user-specific features for developing scalable Twitter bot detection systems. To identify the most important features, we apply three model inspection methods - Permutation Importance (PI), SHapely Additive exPlanation (SHAP), and Local Interpretable Model-agnostic Explanations (LIME). We find that the number of followers, friends, and favourites, and the rate of Tweets, making friends and liking Tweets are the most important user-specific features for Twitter bot detection. We apply the Wilcoxon signed rank test to compare the performance of the models trained using all features, using the important features and the features not found as important in our evaluation, respectively. We observe that there are no significant differences between the performance of the models trained using all features and the models trained using the important features. On the other hand, the models using the unimportant features by our evaluation show statistically significant poor performance. We demonstrate that the above six features are sufficient to identify Twitter bots.
While social bots can be used for various good causes, they can also be utilized to manipulate people and spread malware. Therefore, it is crucial to detect bots running on social media platforms. However, social bots are increasingly successful in creating human-like messages with the recent developments in artificial intelligence. Thus, we need more sophisticated solutions to detect them. In this study, we propose a novel deep learning architecture in which three long short-term memory (LSTM) models and a fully connected layer are utilized to capture complex social media activity of humans and bots. Since our architecture involves many components connected at different levels, we explore three learning schemes to train each component effectively. In our extensive experiments, we analyze the impact of each component of our architecture on classification accuracy using four different datasets. Furthermore, we show that our proposed architecture outperforms all baselines used in our experiments.
Conference Paper
Full-text available
We develop a technique to identify abnormally correlated user accounts in Twitter, which are very unlikely to be human operated. This new approach of bot detection considers cross-correlating user activities and requires no labeled data, as opposed to existing bot detection techniques that consider users independently, and require large amount of recently labeled data. Our system uses a lag-sensitive hashing technique and a warping-invariant correlation measure to quickly organize the user accounts in clusters of abnormally correlated accounts. Our method is 94% precise and detects unique bots that other methods cannot detect. Our system produces daily reports on bots at a rate of several hundred bots per day. The reports are available online for further analysis.
Technical Report
Full-text available
While most online social media accounts are controlled by humans, these platforms also host automated agents called social bots or sybil accounts. Recent literature reported on cases of social bots imitating humans to manipulate discussions, alter the popularity of users, pollute content and spread misinformation, and even perform terrorist propaganda and recruitment actions. Here we present BotOrNot, a publicly-available service that leverages more than one thousand features to evaluate the extent to which a Twitter account exhibits similarity to the known characteristics of social bots. Since its release in May 2014, BotOrNot has served over one million requests via our website and APIs.
Full-text available
Vast numbers of scientific articles are published each year, some of which attract considerable attention, and some of which go almost unnoticed. Here, we investigate whether any of this variance can be explained by a simple metric of one aspect of the paper's presentation: the length of its title. Our analysis provides evidence that journals which publish papers with shorter titles receive more citations per paper. These results are consistent with the intriguing hypothesis that papers with shorter titles may be easier to understand, and hence attract more citations.
Conference Paper
Full-text available
As popular tools for spreading spam and malware, Sybils (or fake accounts) pose a serious threat to online communities such as Online Social Networks (OSNs). Today, sophisticated attackers are creating realistic Sybils that effectively befriend legitimate users, rendering most automated Sybil detection techniques ineffective. In this paper, we explore the feasibility of a crowdsourced Sybil detection system for OSNs. We conduct a large user study on the ability of humans to detect today's Sybil accounts, using a large corpus of ground-truth Sybil accounts from the Facebook and Renren networks. We analyze detection accuracy by both " experts " and " turkers " under a variety of conditions, and find that while turkers vary significantly in their effectiveness , experts consistently produce near-optimal results. We use these results to drive the design of a multi-tier crowd-sourcing Sybil detection system. Using our user study data, we show that this system is scalable, and can be highly effective either as a standalone system or as a complementary technique to current tools.
A large number of Twitter users are bots. They can send spam, manipulate public opinion, and contaminate the Twitter API stream that underline so many research works. One of the major challenges of research on Twitter bots is the lack of ground truth data. Here we report our discovery of the Star Wars botnet with more than 350k bots. We show these bots were generated and centrally controlled by a botmaster. These bots exhibit a number of unique features, which reveal profound limitations of existing bot detection methods. Our work has significant implications for cybersecurity, not only because the size of the botnet is larger than those analysed before, but also because it has been well hidden since its creation in 2013. We argue that more research is needed to fully understand the potential security risks that a large, hidden botnet can pose to the Twitter environment, and research in general.
Social media have been extensively praised for increasing democratic discussion on social issues related to policy and politics. However, what happens when this powerful communication tools are exploited to manipulate online discussion, to change the public perception of political entities, or even to try affecting the outcome of political elections? In this study we investigated how the presence of social media bots, algorithmically driven entities that on the surface appear as legitimate users, affect political discussion around the 2016 U.S. Presidential election. By leveraging state-of-the-art social bot detection algorithms, we uncovered a large fraction of user population that may not be human, accounting for a significant portion of generated content (about one-fifth of the entire conversation). We inferred political partisanships from hashtag adoption, for both humans and bots, and studied spatio-temporal communication, political support dynamics, and influence mechanisms by discovering the level of network embeddedness of the bots. Our findings suggest that the presence of social media bots can indeed negatively affect democratic political discussion rather than improving it, which in turn can potentially alter public opinion and endanger the integrity of the Presidential election. © 2016, Alessandro Bessi and Emilio Ferrara. All Rights Reserved.
Conference Paper
Many qualitative studies of communication practices on social media have recognized that people's motivation for participating in social networks can vary greatly. Some people participate for fame and fortune, while others simply wish to chat with friends. In this paper, we study the implications of such heterogeneous intent for modeling information diffusion in social networks. We experiment with user-level perception of messages, analyze large-scale information cascades, and model information diffusion in heterogeneous-intent networks. We perform carefully designed user studies to establish the relationship between the intent and language style of a message sender. Style of the user appear to adapt their language to achieve different intents. We perform a large-scale data analysis on Twitter message cascades and confirm that message propagation through a network is correlated with historical representations of individuals' intents. Finally, we posit a simple analytical model of information diffusion in social networks that takes heterogeneous intents into account and find that this model is able to explain empirically observed properties of structural virality that are not explained by current models.
Conference Paper
We present a machine learning framework that leverages a mixture of metadata, network, and temporal features to detect extremist users, and predict content adopters and interaction reciprocity in social media. We exploit a unique dataset containing millions of tweets generated by more than 25 thousand users who have been manually identified, reported, and suspended by Twitter due to their involvement with extremist campaigns. We also leverage millions of tweets generated by a random sample of 25 thousand regular users who were exposed to, or consumed, extremist content. We carry out three forecasting tasks, (i) to detect extremist users, (ii) to estimate whether regular users will adopt extremist content, and finally (iii) to predict whether users will reciprocate contacts initiated by extremists. All forecasting tasks are set up in two scenarios: a post hoc (time independent) prediction task on aggregated data, and a simulated real-time prediction task. The performance of our framework is extremely promising, yielding in the different forecasting scenarios up to 93% AUC for extremist user detection, up to 80% AUC for content adoption prediction, and finally up to 72% AUC for interaction reciprocity forecasting. We conclude by providing a thorough feature analysis that helps determine which are the emerging signals that provide predictive power in different scenarios.
To help activists call new volunteers to action, we present Botivist: a platform that uses Twitter bots to find potential volunteers and request contributions. By leveraging different Twitter accounts, Botivist employs different strategies to encourage participation. We explore how people respond to bots calling them to action using a test case about corruption in Latin America. Our results show that the majority of volunteers (>80%) who responded to Botivist's calls to action contributed relevant proposals to address the assigned social problem. Different strategies produced differences in the quantity and relevance of contributions. Some strategies that work well offline and face-to-face appeared to hinder people's participation when used by an online bot. We analyze user behavior in response to being approached by bots with an activist purpose. We also provide strong evidence for the value of this type of civic media, and derive design implications.