During the fall of 2018, as part of my data journalism research for Storybench, a "cookbook for digital storytelling" by Northeastern University's School of Journalism, I collected Twitter data from 2018 midterm election candidates (with R) to see the possible correlation between the sentiment of those tweets and election results. To better understand these words, a machine learning model was used to predict tweet sentiment, and results are visualized (with Python) and published here.
Data were scraped from Twitter with rtweet, a package developed by Michael W. Kearney from the University of Missouri. Data are either from twitter accounts of candidate themselves or their official campaign accounts.
library("rtweet")
allcandidates <- get_timelines(
c("SenatorCantwell", "Susan4Senate", "jontester", "MattForMontana", "SenFeinstein", "kdeleon", "RosenforNevada", "SenDeanHeller", "kyrstensinema", "RepMcSally", "MittRomney", "JennyWilsonUT", "SenJohnBarrasso", "MartinHeinrich", "MickRich4Senate", "GovGaryJohnson", "maziehirono", "rcurtis808", "tedcruz", "BetoORourke", "FLGovScott", "SenBillNelson", "RepKevinCramer", "SenatorHeitkamp", "SenatorFischer", "JaneRaybould", "amyklobuchar", "NewbergerJim", "TinaSmithMN", "KarinHousley", "tammybaldwin", "LeahVukmir", "stabenow", "JohnJamesMI", "HawleyMO", "clairecmc", "RogerWicker", "dbaria", "MarshaBlackburn", "PhilBredesen", "braun4indiana", "JoeforIndiana", "SenSherrodBrown", "JimRenacci", "SenBobCasey", "louforsenate", "timkaine", "CoreyStewartVA", "Sen_JoeManchin", "MorriseyWV", "SenatorCardin", "Campbell4MD", "SenatorCarper", "RobArlett", "SenatorMenendez", "BobHugin", "SenGillibrand", "CheleFarley", "ChrisMurphyCT", "MattCoreyCT", "elizabethforma", "RepGeoffDiehl", "SenWhitehouse", "flanders4senate", "SenAngusKing", "RingelsteinME", "SenSanders", "ZupanForSenate"), retryonratelimit = TRUE, n = 3000, include_rts = FALSE
)
write_as_csv(allcandidates, "/your_own_file_directory/all_tweets.csv", prepend_ids = TRUE, na = "", fileEncoding = "UTF-8") #saving data in a csv fileThe model is developed by Jeremy Howard, a deep learning researcher and educator, as part of his course fast.ai. All code below (Python) were run on paperspace.com. Instructions on how to set up a machine on paperspace can be found here.
To train the model, I ran Jeremy's code line by line on paperspace. His model was trained by a large movie review dataset containing 50,000 reviews from IMDB.
# loading required packages
%reload_ext autoreload
%autoreload 2
%matplotlib inline
from fastai.learner import *
import torchtext
from torchtext import vocab, data
from torchtext.datasets import language_modeling
from fastai.rnn_reg import *
from fastai.rnn_train import *
from fastai.nlp import *
from fastai.lm_rnn import *
import dill as pickle
import spacy
import numpy as np
import pandas as pd
# path of data
PATH='data/aclImdb/'
# defining required variables, parameters and functions, and loading the model
em_sz = 200 # size of each embedding vector
nh = 200 # number of hidden activations per layer
nl = 3 # number of layers
opt_fn = partial(optim.Adam, betas=(0.7, 0.99))
TEXT = data.Field(lower=True, tokenize=spacy_tok)
TEXT = pickle.load(open(f'{PATH}models/TEXT.pkl','rb'))
IMDB_LABEL = data.Field(sequential=False)
bs=64; bptt=70
splits = torchtext.datasets.IMDB.splits(TEXT, IMDB_LABEL, 'data/') #this splits all the words in the model into positive and negative
md2 = TextData.from_splits(PATH, splits, bs)
m3 = md2.get_model(opt_fn, 1500, bptt, emb_sz=em_sz, n_hid=nh, n_layers=nl,
dropout=0.1, dropouti=0.4, wdrop=0.5, dropoute=0.05, dropouth=0.3)
m3.reg_fn = partial(seq2seq_reg, alpha=2, beta=1)
m3.load_encoder(f'adam3_10_enc')
m3.load_cycle('imdb2', 4)
accuracy(*m3.predict_with_targs()) #accuracy of my model is 0.90849358974358974
# loading data and storing data in separate panda dataframes
AllCandidates = pd.read_csv("data/all_tweets.csv")
candidates = [1]*63 # creating 63 lists because I am analysing 63 candidates' tweets; 1 is just a randomly chosen number
names = AllCandidates['screen_name'].unique() #putting all candidates' Twitter handles in "names"
for i in range (0,63):
candidates[i] = AllCandidates[AllCandidates['screen_name'].str.contains(names[i])] # by doing this, data for each candidate are stored in separate dataframes. For example, candidates[0] stores all data for "SenatorCantwell"; candidates[1] for "Susan4Senate"; candidates[2] for "jontester"...
# predicting sentiment for each candidate's tweets
m = m3.model #this is the model
m[0].bs=1 #this is the batch size. so, analyzing one tweet at a time
prediction = []
for i in range(candidates[j].values[:,4].shape[0]): #replace j with 0, 1, 2, 3...62 for each candidate
ss = candidates[j]['text'][i+a] #saving the ith tweet here; replace a with 0 (Cantwell's tweets start here), 3030 (Susan's tweets start here)...
s = [spacy_tok(ss)] #tokenizing the words
t = TEXT.numericalize(s) #giving the words a number
m.eval()
m.reset() #reset it so it can analyze the next tweet
res,*_ = m(t)
prediction.append(IMDB_LABEL.vocab.itos[to_np(torch.topk(res[-1],1)[1])[0]])
# showing results
positive = [i for i,x in enumerate(prediction) if x == 'pos'] # all positive tweets here
negative = [i for i,x in enumerate(prediction) if x == 'neg'] # all negative tweets here
len(positive) #number of positive tweets by each candidate
len(negative) #number of negative tweets by each candidateThe closer to the election date, the more candidates tweeted.
import pandas as pd
import matplotlib.pyplot as plt
import numpy as np
import seaborn as sns
from scipy import stats
sns.set()
candidates = pd.read_csv("Updated Candidates.csv")
percentage_of_positive_tweets = [72.77215502,72.89565754,74.76340694,77.25617934,77.09820567,78.4372093,75.56086797,75.48902875,77.11086923,75.85608332,77.11526349,81.2329108]
month = [1,2,3,4,5,6,7,8,9,10,11,12]
fig=plt.figure(figsize=(20,10))
plt.scatter(month,percentage_of_positive_tweets, s=500)
plt.ylabel('% of positive tweets from all candidates', fontsize = 25)
plt.yticks(fontsize =27)
plt.xlabel('month and year', fontsize=25)
plt.xticks(month,
['Dec 2017', 'Jan 2018', 'Feb 2018', 'March 2018', 'April 2018', 'May 2018', 'June 2018', 'July 2018', 'Aug 2018', 'Sept 2018', 'Oct 2018', 'Nov 2018'], fontsize = 16)
fig.savefig('positive_tweets.jpg')
Candidates were generally more positive as they approached election day.
fig=plt.figure(figsize=(20,10))
plt.scatter([1,2,3,4,5,6,7,8,9,10,11,12],[4051,4859,4121,5219,5183,5375,5438,5879,6431,6337,9526,4023], s=500)
plt.ylabel('number of tweets from all candidates', fontsize = 30)
plt.xlabel('month and year', fontsize=25)
plt.yticks(fontsize =27)
plt.xticks(month,
['Dec 2017', 'Jan 2018', 'Feb 2018', 'March 2018', 'April 2018', 'May 2018', 'June 2018', 'July 2018', 'Aug 2018', 'Sept 2018', 'Oct 2018', 'Nov 2018'], fontsize = 16)
fig.savefig('tweets_over_time.jpg')
In general, the more candidates tweeted, the more votes they got. And Democrats tweeted more and got more votes.
fig=plt.figure(figsize=(20,10))
plt.scatter(np.log(candidates[candidates['Party (D/R/L/I)'] == 'D']['Total number of tweets']),candidates[candidates['Party (D/R/L/I)'] == 'D']['percentage of vote'], s=500, color='blue')
plt.scatter(np.log(candidates[candidates['Party (D/R/L/I)'] == 'R']['Total number of tweets']),candidates[candidates['Party (D/R/L/I)'] == 'R']['percentage of vote'], s=500, color='red')
plt.scatter(np.log(candidates[candidates['Party (D/R/L/I)'] == 'I']['Total number of tweets']),candidates[candidates['Party (D/R/L/I)'] == 'I']['percentage of vote'], s=500, color='purple')
plt.scatter(np.log(candidates[candidates['Party (D/R/L/I)'] == 'L']['Total number of tweets']),candidates[candidates['Party (D/R/L/I)'] == 'L']['percentage of vote'], s=500, color='yellow')
plt.ylabel('% of vote', fontsize = 30)
plt.xlabel('log(number of tweets)', fontsize=25)
plt.yticks(fontsize =27)
plt.xticks(fontsize =25)
plt.legend(('Democrats', 'Republicans', 'Independents', 'Libertarians'), loc='upper left', fontsize=21)
fig.savefig('tweets_and_votes.jpg')
In general, the more followers candidates had, the more votes they got. Again, Democrats had more followers and more votes.
fig=plt.figure(figsize=(20,10))
plt.scatter(np.log(candidates[candidates['Party (D/R/L/I)'] == 'D']['Follower count']),candidates[candidates['Party (D/R/L/I)'] == 'D']['percentage of vote'], s=500, color='blue')
plt.scatter(np.log(candidates[candidates['Party (D/R/L/I)'] == 'R']['Follower count']),candidates[candidates['Party (D/R/L/I)'] == 'R']['percentage of vote'], s=500, color='red')
plt.scatter(np.log(candidates[candidates['Party (D/R/L/I)'] == 'I']['Follower count']),candidates[candidates['Party (D/R/L/I)'] == 'I']['percentage of vote'], s=500, color='purple')
plt.scatter(np.log(candidates[candidates['Party (D/R/L/I)'] == 'L']['Follower count']),candidates[candidates['Party (D/R/L/I)'] == 'L']['percentage of vote'], s=500, color='yellow')
plt.ylabel('% of vote', fontsize = 30)
plt.xlabel('log(number of twitter followers)', fontsize=30)
plt.yticks(fontsize =27)
plt.xticks(fontsize =25)
plt.legend(('Democrats', 'Republicans', 'Independents', 'Libertarians'), loc='upper left', fontsize=30)
fig.savefig('followers_and_votes.jpg')
For Democrats, the more positive tweets they had, the fewer votes they had. The opposite was true for Republicans.
fig=plt.figure(figsize=(20,10))
x1 = candidates[candidates['Party (D/R/L/I)'] == 'D']['% of positives']
x2 = candidates[candidates['Party (D/R/L/I)'] == 'R']['% of positives']
x3 = candidates[candidates['Party (D/R/L/I)'] == 'I']['% of positives']
x4 = candidates[candidates['Party (D/R/L/I)'] == 'L']['% of positives']
y1 = candidates[candidates['Party (D/R/L/I)'] == 'D']['percentage of vote']
y2 = candidates[candidates['Party (D/R/L/I)'] == 'R']['percentage of vote']
y3 = candidates[candidates['Party (D/R/L/I)'] == 'I']['percentage of vote']
plt.scatter(candidates[candidates['Party (D/R/L/I)'] == 'D']['% of positives'],candidates[candidates['Party (D/R/L/I)'] == 'D']['percentage of vote'], s=500, color='blue')
plt.scatter(candidates[candidates['Party (D/R/L/I)'] == 'R']['% of positives'],candidates[candidates['Party (D/R/L/I)'] == 'R']['percentage of vote'], s=500, color='red')
plt.scatter(candidates[candidates['Party (D/R/L/I)'] == 'I']['% of positives'],candidates[candidates['Party (D/R/L/I)'] == 'I']['percentage of vote'], s=500, color='purple')
plt.scatter(candidates[candidates['Party (D/R/L/I)'] == 'L']['% of positives'],candidates[candidates['Party (D/R/L/I)'] == 'L']['percentage of vote'], s=500, color='yellow')
plt.legend(('Democrats', 'Republicans', 'Independents', 'Libertarians'), loc='lower right', fontsize=16)
slope1, intercept1, r_value1, p_value1, std_err1 = stats.linregress(x1,y1)
line1 = slope1*x1+intercept1
plt.plot(x1,line1)
slope2, intercept2, r_value2, p_value2, std_err2 = stats.linregress(x2,y2)
line2 = slope2*x2+intercept2
plt.plot(x2,line2)
slope3, intercept3, r_value3, p_value3, std_err3 = stats.linregress(x3,y3)
line3 = slope3*x3+intercept3
plt.plot(x3,line3)
plt.ylabel('% of vote', fontsize = 30)
plt.xlabel('% of positive tweets', fontsize=25)
plt.yticks(fontsize =27)
plt.xticks(fontsize =25)
fig.savefig('% of positives and % of vote.jpg')