Top 10 2017 Dineout Vancouver Restaurants by Twitter Sentiment Analysis

I’m a huge fan of food. I like cooking it, eating it, taking pictures of it and talking about it. January is an exciting month in Vancouver for food lovers, because we have Dineout Vancouver, an annual event where hundreds of local restaurants create fixed priced menus showcasing their cuisine, for very reasonable prices.

But, with over 280 restaurants participating according to the Georgia Straight, and fierce competition for reservations, how are you supposed to choose where to spend your money?

Inspired by David Robinson’s post outlining sentiment analysis of Yelp data here, I decided to combine my love of food with my love of data science, and conduct a sentiment analysis of tweets about Dineout Vancouver and rank restaurants based on their average sentiment.

The first question you might have is: What is sentiment analysis? In short, it’s looking at text and specific words in text and labeling them as either positive or negative, to determine whether the message as a whole is positive or negative. For more information read this

Getting The Tweets

My first task was getting all the tweets. For this I used Python and the tweepy library. I did a quick Google and Twitter search for Dineout Vancouver to find relevant hashtags and user accounts. I found that most posts were using #DineOutVancouver, #DOVF, #dinearound2017 and @DineOutVanFest, the organization’s official Twitter handle. Next I connected to the twitter API using my credentials. If you’re following along here, I’m assuming that you already know how to register a new twitter application, and grab your credentials. If you don’t know how to do that read this.

I started by opening a connection using tweepy, and searching for tweets with the relevant words. I eliminated the tweets containing “RT” as these tended to be the restaurants themselves retweeting positive tweets. For each tweet, I grabbed the text of the tweet, the sender and the unique id. Then, I turned everything into a pandas dataframe and appended the output to a CSV file, as I did the rest of the analysis in R. Note that I appended here instead of overwriting, because the twitter API limits your tweets to the last week. I’d like to re-run this a number of times to get all the tweets during Dineout.

# load packages
import tweepy
import json
import pandas as pd

# put in your credentials
consumer_key = "your credentials here"
consumer_secret = "your credentials here"
access_token = "your credentials here"
access_secret = "your credentials here"

# authorize
authorization = tweepy.OAuthHandler(consumer_key=consumer_key, consumer_secret=consumer_secret)
authorization.set_access_token(access_token, access_secret)

# create the client
client = tweepy.API(authorization)

# initialize empty lists to store tweet info
tweets = list()
from_user = list()
tweet_id = list()

# for each tweet matching the search criteria
for tweet in tweepy.Cursor(client.search,
                           q='@DineOutVanFest OR #DineOutVancouver OR #DOVF OR #dinearound2017 exclude:retweets').items():
    
    # only get tweets that aren't retweets
    if "RT" not in tweet.text:
        tweets.append(tweet.text) # add the tweet text
        from_user.append(tweet.user.screen_name) # add the sender
        tweet_id.append(tweet.id_str) # add the tweet id

# turn everything into a dataframe
dineout_tweets = pd.DataFrame({'sender': from_user,
                                'tweet_text': tweets,
                               'tweet_id': tweet_id}, columns=['sender', 'tweet_text', 'tweet_id'])

# print the number of tweets grabbed
print("Number of Tweets Grabbed: %f" % (len(tweets)))

# append the tweets to the current file instead of overwriting to get cumulative tweets
# handle duplicates later
dineout_tweets.to_csv("../data/tweets.csv", mode='a', sep=',', index=False)

Necessary But Painful Data Entry

I’m not going to lie, this part wasn’t fun. I’m sure with enough effort I could have mostly automated all of this by writing a little program in scrapy to crawl search results and look for twitter handles. But, I felt that would take longer than doing it manually. So, I just bit the bullet and manually collected all of the twitter names of participating restaurants I could find. I used Google Sheets to compile this into a list and then save it as a CSV. I think it’s important to note here, that not every restaurant has a Twitter handle, and some restaurant chains such as Cactus Club or Glowbal Group, have multiple restaurants under one name. This isn’t ideal, as in the case of Glowbal Group, they have a bunch of different restaurants serving different food. So, a review of one restaurant will affect the others, but it’s the best I could do.

Data Wrangling and Analysis

First, I loaded all of the packages I needed for the analysis, including tidyverse. If you don’t have this installed, using install.packages("tidyverse") will install everything you need for data wrangling, and then you’ll need to install tidytext with install.packages("tidytext") which is a package in R for doing text analysis.

library(tidyverse)
library(stringr)
library(jsonlite)
library(tidytext)
library(knitr)

Next, I read in my list of participating restaurants, and eliminated the duplicates mentioned above. I use this later to filter out tweets to users who aren’t participating restaurants.

# load in the list of dineout restaurants
restaurants <- read_csv("../data/dineout_restaurants.csv")

# filter out the ones that don't have a twitter username
restaurants <- restaurants %>% 
  filter(!is.na(twitter_username))

# unique list of participating restaurant twitter names
restaurant_twitter_names <- unique(restaurants$twitter_username)

Here, I loaded in the list of tweets, and eliminated any duplicates. I did this because every time I pull new tweets I’m appending instead of overwriting, so there will always be duplicates. Then, I wrote the clean dataset back to the file. I also converted the id to a character string rather than a number.

# load in the list of tweets with either #dineoutvancouver or @dineoutvanfest
tweets <- read_csv(file = "../data/tweets.csv")

# eliminate duplicates
tweets <- distinct(tweets)

# write the eliminated duplicates back to file
write_csv(x = tweets, path = "../data/tweets.csv", col_names = TRUE)

# convert the tweets_ids to a character
tweets <- tweets %>% 
  mutate(tweet_id=as.character(tweet_id))

At this point, I had a list of unique tweets, but the usernames I needed were within the text of the tweet, and some tweets mention multiple users. So, I started by using map to create an extra column only containing the users mentioned in the tweet. Then, for each tweet I created a separate row containing the same tweet information for each user mentioned. This is important, because where multiple users are mentioned, I wanted the sentiment of each tweet to contribute to each of the mentioned user’s scores.

# pull the list of users sent the message from the text of the tweet and create a new column
users <- unlist(map(map(map(tweets$tweet_text, .f = str_extract_all, pattern="@\\w+"), unlist),paste, collapse = " "))
tweets$users <- users

# get one row per user
tweets <- tweets %>% 
  unnest_tokens(username, users)

Here’s where the magic of tidytext came in. First, the unnest_tokens() method allowed me to easily split each tweet into a separate row per word. The tidytext package also contains a list of pre-compiled stop words to help me filter out words not relevant to the sentiment analysis. So, my next step was removing these stop words from the words in each tweet. Next, I used the AFINN lexicon, also included in the tidytext package that includes a positivity score from -5 to positive 5 for each word. I joined the AFINN words using an inner join with the list of tweeted words, to get a sentiment score for each word. Then, I calculated the mean sentiment score for each tweet.

# remove stop words from the tweets
tweet_words <- tweets %>%
  select(username, tweet_text, tweet_id) %>%
  unnest_tokens(word, tweet_text) %>%
  filter(!word %in% stop_words$word,
         str_detect(word, "^[a-z']+$"))

# get a list of words with sentiment scores from the tidytext package
AFINN <- sentiments %>%
  filter(lexicon == "AFINN") %>%
  select(word, afinn_score = score)

# get the average sentiment of each tweet
tweet_sentiment <- tweet_words %>%
  inner_join(AFINN, by = "word") %>%
  group_by(username, tweet_id) %>% 
  summarize(sentiment = mean(afinn_score))

print(head(tweet_sentiment,10))

## Source: local data frame [10 x 3]
## Groups: username [7]
## 
##           username           tweet_id sentiment
##              <chr>              <chr>     <dbl>
## 1     aliencheftim 822521576134299649  3.000000
## 2  alportoitaliano 822621287864532992  1.000000
## 3        alumniubc 819705180627279872  2.000000
## 4     ancoradining 822680393455267840  4.000000
## 5      annalenayvr 821520226495799296  1.666667
## 6      annalenayvr 821949762035298305  2.000000
## 7      annalenayvr 822604362082029570  4.000000
## 8      annalenayvr 822637253591703552  3.000000
## 9      askforluigi 820358644185333760 -2.000000
## 10     basilmintbc 822248572951392256  2.000000

print(sprintf("Number of tweets with a sentiment score: %d", length(tweet_sentiment$tweet_id)))

## [1] "Number of tweets with a sentiment score: 453"

print(sprintf("Number of total tweets: %d", length(tweets$tweet_id)))

## [1] "Number of total tweets: 1434"

At this point, I had a list of tweets with an average sentiment score, based on the sentiment score of the individual words in each tweet. But, I had a bit of a problem here. There was only 453 out of 1434 tweets that contained any of the words in the AFINN lexicon. This meant many of the tweets didn’t contain negative or positive words. At this point, I needed to make a choice. I decided to count tweets with no negative or positive words as neutral by assigning them a zero value. To do this, I combined the tweets with sentiment scores with the original tweets again, and then I’m assigned 0 to the sentiment value of any tweet with NA.

# combine the sentiment with original tweet list
tweet_with_sentiment <- tweets %>% 
  left_join(tweet_sentiment)

# fill the tweets with no sentiment with zeros so they can be neutral
tweet_with_sentiment$sentiment[is.na(tweet_with_sentiment$sentiment)] <- 0

After that, I had a list of all the tweets again with sentiment scores for all of them. Next, I needed to determine a sentiment score for each user and filter out users mentioned that aren’t participating restaurants.

# get the average sentiment of each user
user_sentiment <- tweet_with_sentiment %>%
  group_by(username) %>%
  summarize(sentiment = mean(sentiment), n_tweets=n())

# filter only for participating restaurants and not other users.
participant_sentiment <- user_sentiment %>% 
  filter(username %in% restaurant_twitter_names) %>% 
  arrange(-sentiment)

# get the top ten users
top_ten_sentiment <- head(participant_sentiment, 10)

# format column names more descriptively
top_ten_sentiment <- top_ten_sentiment %>% 
  select(`Twitter Name`=username, Sentiment=sentiment, Tweets=n_tweets)

# print a nice table of the top restaurants
top_ten_table <- kable(x = top_ten_sentiment, align = 'c', caption = "Top Ten Dineout 2017 Restaurants by Twitter Sentiment")

Next, I printed the top ten restaurants in a table along with their sentiment score and number of tweets. I also produced a bar chart with the same information. Lastly, I created a scatter plot of the number of tweets versus the sentiment score to visualize any relationship between the two.

# reorder the restaurants by descending sentiment
plot_sentiment <- participant_sentiment %>%
  mutate(username=reorder(username, -sentiment)) %>%
  select(username, sentiment, n_tweets) %>%
  head(10) %>% 
  gather(key = measure, value = value, sentiment, n_tweets)

# plot the sentiment and number of tweets
top_ten_bar <- ggplot(plot_sentiment, aes(x=username, y=value, fill=measure)) + geom_bar(stat = "identity", position = "dodge") +
  theme(axis.text.x = element_text(angle = 45, hjust = 1)) +
  xlab("Twitter Username") +
  ylab("Sentiment / Number of Tweets") + 
  ggtitle(label = "Top Ten Dineout Vancouver Restaurants", subtitle = "by Twiiter Sentiment Score")

# plot to compare number of tweets v sentiment
tweets_v_sentiment <- ggplot(participant_sentiment, aes(x=n_tweets, y=sentiment, label = username)) + 
  geom_point() +
  geom_text(size = 3, check_overlap = TRUE, hjust = 0, nudge_x = 0.5) +
  xlab("Number of Tweets") +
  ylab("Sentiment") + 
  ggtitle(label = "Dineout Vancouver Restaurants", subtitle = "by Twiiter Sentiment Score and Number of Tweets")

Results

Drumroll please. Here are the top ten Dineout Vancouver 2017 restaurants by sentiment score.

Discussion

This small example was pretty exciting to me. In a few hours worth of work using the Twitter API and packages such as tweepy for Python and tidytext for R, I was able to conduct a real life sentiment analysis of something I care about, to help me make a purchase decision. But, there’s definitely some limitations of this analysis and room for improvement. Firstly, there isn’t a lot of data here, as Dineout Vancouver 2017 only began on Friday, January 20th. One positive or negative tweet can really sway the sentiment score. Also, looking at some of the tweets that are negative, the algorithm definitely isn’t perfect. Here’s a few examples:

# get the worst tweets
worst_tweets <- tweet_with_sentiment %>%
  filter(username %in% restaurant_twitter_names & sentiment < 0) %>% 
  arrange(sentiment) %>%
  select(`Tweet Text`=tweet_text, Sentiment=sentiment)

# print them in a nice table
kable(format = 'html', x = worst_tweets, caption = "Worst Tweets by Sentiment")

Most of these tweets are actually negative, but there are some that aren’t. Also, there’s a lot of negative tweets about @EdibleCanada, but they are all related to the moral issue of them serving seal meat, rather than how good the food was or how good the service was. I think another improvement that could be made here, would be to create or find a list of positive/negative words specifically related to food, that may not be in the list I used from tidytext.

If you would like to use/adapt the code from this post, I’ve included all of my code and data files in this github repository

If you enjoyed my post, have any suggestions or would like to see me write about anything in the future, please let me know.