tweetbotornot2
{tweetbotornot2} provides an out-of-the-box classifier for detecting Twitter bots that is easy to use, volume-friendly (classify over two hundred thousands accounts per day), accurate (see paper), robust and customizable (update the model with new or user-defined training data), and explainable (view feature contributions behind each estimated probability).
Installation
You can install the released version of tweetbotornot2 from CRAN with:
install.packages("tweetbotornot2")Predict
Use predict_bot() to use the built-in bot classifier
Enter screen names of Twitter users of interest and predict_bot()
returns the estimated probability that one or more Twitter accounts is a
bot.
## pass screen names to predict function
predict_bot(c("MagicRealismBot", "netflix_bot", "rdpeng", "hspter"))
#> user_id screen_name prob_bot
#> 1: 1203840834 netflix_bot 0.9974465
#> 2: 24228154 hspter 0.0038939
#> 3: 3701125272 MagicRealismBot 0.9999900
#> 4: 9308212 rdpeng 0.0171962predict_bot() also accepts previously collected Twitter data (e.g.,
data returned by rtweet::get_timelines())
## estimate bot probabilities
twtdat <- rtweet::get_timelines(
c(
## (these ones should be bots)
"American__Voter",
"MagicRealismBot",
"mitchhedbot",
"rstats4ds",
"thinkpiecebot",
"tidyversetweets",
"newstarsbot",
## (these ones should be Nots)
"AOC",
"kearneymw",
"kumailn",
"mindykaling",
"NateSilver538",
"realDonaldTrump",
"hspter",
"rdpeng"),
n = 200, check = FALSE)
## view output (order most likely to least like bot)
predict_bot(twtdat)[order(-prob_bot), ]
#> user_id screen_name prob_bot
#> 1: 3701125272 MagicRealismBot 0.99998999
#> 2: 780707721209188352 newstarsbot 0.99998736
#> 3: 935569091678691328 tidyversetweets 0.99997628
#> 4: 829792389925597184 American__Voter 0.99997056
#> 5: 3325527710 thinkpiecebot 0.99996030
#> 6: 1075011651366199297 rstats4ds 0.99983037
#> 7: 214244836 mitchhedbot 0.99942195
#> 8: 2973406683 kearneymw 0.05339408
#> 9: 25073877 realDonaldTrump 0.01949467
#> 10: 9308212 rdpeng 0.01719619
#> 11: 16017475 NateSilver538 0.00723628
#> 12: 24228154 hspter 0.00389387
#> 13: 138203134 AOC 0.00122775
#> 14: 23544596 mindykaling 0.00016858
#> 15: 28406270 kumailn 0.00016037Explain
Use explain_bot() to see the contributions made by each feature in the model
Examine prediction contributions for features in the model
## view top features in predictions of each user
explain_bot(twtdat)[feature %in% feature[1:10], .SD, on = "feature"][1:30, -1]
#> screen_name prob_bot feature value feature_description
#> 1: AOC 0.0012278 twt_srctw -2.349361 Tweet source of Twitter (official)
#> 2: AOC 0.0012278 usr_verif -1.288429 User verified
#> 3: AOC 0.0012278 usr_actyr -0.989596 User account age
#> 4: AOC 0.0012278 twt_wrdsd 0.467928 Tweet word count variation
#> 5: AOC 0.0012278 twt_hshsd -0.451500 Tweet hashtags variation
#> 6: AOC 0.0012278 usr_fllws -0.444115 User followers
#> 7: AOC 0.0012278 twt_srcna -0.434539 Tweet source of unknown
#> 8: AOC 0.0012278 BIAS 0.428031 Intercept (y when x=0)
#> 9: AOC 0.0012278 twt_wrdmn 0.422080 Tweet word count mean
#> 10: AOC 0.0012278 twt_wdtsd -0.391492 Tweet display widht variatiojn
#> 11: American__Voter 0.9999706 twt_srctw 2.447532 Tweet source of Twitter (official)
#> 12: American__Voter 0.9999706 usr_actyr 1.839742 User account age
#> 13: American__Voter 0.9999706 twt_srcna 0.995951 Tweet source of unknown
#> 14: American__Voter 0.9999706 BIAS 0.428031 Intercept (y when x=0)
#> 15: American__Voter 0.9999706 twt_wdtsd 0.294240 Tweet display widht variatiojn
#> 16: American__Voter 0.9999706 usr_fllws -0.170949 User followers
#> 17: American__Voter 0.9999706 twt_wrdmn -0.121823 Tweet word count mean
#> 18: American__Voter 0.9999706 twt_hshsd 0.113032 Tweet hashtags variation
#> 19: American__Voter 0.9999706 twt_wrdsd 0.052839 Tweet word count variation
#> 20: American__Voter 0.9999706 usr_verif 0.037238 User verified
#> 21: MagicRealismBot 0.9999900 twt_srctw 2.312730 Tweet source of Twitter (official)
#> 22: MagicRealismBot 0.9999900 usr_actyr 2.304153 User account age
#> 23: MagicRealismBot 0.9999900 twt_srcna 0.824347 Tweet source of unknown
#> 24: MagicRealismBot 0.9999900 BIAS 0.428031 Intercept (y when x=0)
#> 25: MagicRealismBot 0.9999900 twt_wrdmn 0.395463 Tweet word count mean
#> 26: MagicRealismBot 0.9999900 usr_fllws -0.351515 User followers
#> 27: MagicRealismBot 0.9999900 twt_wdtsd 0.325772 Tweet display widht variatiojn
#> 28: MagicRealismBot 0.9999900 twt_hshsd 0.143245 Tweet hashtags variation
#> 29: MagicRealismBot 0.9999900 twt_wrdsd 0.060281 Tweet word count variation
#> 30: MagicRealismBot 0.9999900 usr_verif 0.039289 User verified
#> screen_name prob_bot feature value feature_descriptionUpdate
Use sample_via_twitter_lists() to gather new users
sample_via_twitter_lists() leverages Twitter lists to snowball sample
from a few users (input) to hundreds or even thousands of users
(output). Use it to find and label “bot” and “not” (bot) Twitter
accounts via snowball sampling
## find up to 10,000 accounts similar to these well-known Twitter bots
bots <- sample_via_twitter_lists(c("netflix_bot", "American__Voter", "UTLEGtracker",
"JVLast", "EndlessJeopardy", "PossumEveryHour", "MagicRealismBot", "factbot1"),
n = 10000)
## retrieve user and timeline information for each user
bots_tmls <- tweetbotornot_collect(bots)Use preprocess_bot() to prepare user timeline data for modeling
Extract and transform numeric features from the data
## wrangle, munge, aggregate data for modelling
bots_tmls <- preprocess_bot(twtdat)Use tweetbotornot_update to train the model on new data
## train model
new_model <- tweetbotornot_update(newdata)
## use updated model to get new predictions
predict_bot(new_model, c("netflix_bot", "mindykaling"))About Model
Feature importance
The most influential features in the classifier
Feature contributions
How features contributed to predictions in the original training data:
