2.1. Literature review

The importance of social media in promoting sustainability and its effect on users is a well-documented phenomenon. Zhao et al. analyzed the use of social media by sustainable fashion brands to communicate with consumers and promote their sustainability goals. Through a content analysis of the Instagram pages of the brands Verlaine, Patagonia, and Stella McCartney, the researchers found that posts with expressive and directive advertising were more likely to receive comments and likes from users (Zhao et al., Reference Zhao, Lee, Li and Sun2022). This research provides context for user response to social media marketing and how different approaches to social media marketing can impact consumer engagement. While our study focuses on user engagement on the topic of sustainability rather than analyzing different approaches to social media marketing, the study conducted by Zhao et al. established the effectiveness of social media for promoting sustainability, thus showing a relevant application for our research. Similarly, Wibowo et al. also showed the effectiveness of social media in promoting sustainability by exploring the impact of social media marketing activity and customer experience on customer behavior outcomes such as purchase intention, loyalty intention, and participation intention in the context of social networking sites. Through a sample of 413 surveys, the study showed that social media marketing activity and customer experience significantly impacted relationship quality and purchase intention (Wibowo et al., Reference Wibowo, Chen, Wiangin, Ma and Ruangkanjanases2020). These results indicate that social media platforms are an effective way of promoting sustainability. Additionally, a study conducted by Ana Adi provides insight into Twitter’s role in creating a community of interest around sustainability discussion under the sustainability hashtag. Through a collection of 15,000 tweets over three time periods, a network visualization was conducted to establish whether there is a sustainability community on Twitter, and emerging discourse was identified by looking at the top hashtags and a detailed look into the dataset of tweets collected (Adi, Reference Adi2017). This study concluded that the sustainability community on Twitter is largely driven by several accounts that largely shape the conversation through repeated postings and that content frequently revolves around people, planet, and products (Adi, Reference Adi2017). This study does not incorporate machine learning and looks at hashtags to determine topics of discourse related to sustainability; however, it provides valuable insight into factors within the sustainability community on Twitter, and what may shape the discussion under the hashtag.

Public opinion on sustainability is another critical factor to contextualize when researching patterns in sustainable content on Twitter. Sanders et al. analyzed the content of tweets posted to Twitter between July and December 2020 to examine the public conversation regarding the application of artificial intelligence (AI) in agriculture. According to the survey, there was a lot of support for employing AI in agriculture on Twitter (Sanders et al., Reference Sanders, Mayfield-Smith and Lamm2021). Although the study looks at public opinion on agriculture and not trending categories under the hashtag #sustainability, this study establishes that users on Twitter have a generally positive attitude toward sustainability content.

Significant research has also been conducted regarding the most discussed sustainability topics on Twitter. Pilař et al. identified the main topics of communication related to hashtag #sustainability on Twitter to understand how these topics were perceived by users worldwide. From 414,926 Twitter interactions by 223,476 users between April 17, 2018, and July 12, 2019, the study identified six key hashtags in the field of sustainability: Innovation, Environment, Climate Change, Corporate Social Responsibility, Technology, and Energy (Pilař et al., Reference Pilař, Kvasničková Stanislavská, Pitrová, Krejčí, Tichá and Chalupová2019). These results apply primarily to individuals and organizations looking to promote sustainability online by allowing them to be informed on topics that generally trend on Twitter. While the study explored a topic similar to our research, our study differs in that it looks at various sustainability hashtags and only uses the hashtag itself for content analysis. Similarly, Reyes-Menendez et al. (Reference Reyes-Menendez, Saura and Alvarez-Alonso2018) aimed to identify key factors related to sustainable development goals that are of concern to Twitter users. The authors used sentiment analysis and textual analysis to classify the tweets under the hashtag #WorldEnvironmentDay and identify public opinion on environmental and sustainability issues on that day. While this research only focuses on one specific day and does not categorize the major topics on Twitter relating to sustainability, it is helpful in contextualizing our research, as the study found that the World Environment Day hashtag was most often used to talk about climate change, global warming, extreme weather, water pollution, deforestation, climate risks, acid rain, and large-scale industrialization. Understanding what topics are discussed on one particular day can provide some insight into what we might find over a larger period of time.

Another #WorldEnvironmentDay study conducted by Zarrabeitia-Bilbao et al. explores the #WorldEnvironmentDay hashtag on Twitter and applies big data techniques and social network analysis to analyze over 400,000 tweets under the #WorldEnvironmentDay hashtag on June 5, 2021. This study analyzed the sentiment of tweets to be largely positive and conducted a semantic analysis to find words that are commonly mentioned among these tweets.

Additionally, different community leaders and topics of conversation on this day were identified using a machine learning analysis (Zarrabeitia-Bilbao et al., Reference Zarrabeitia-Bilbao, Rio-Belver, Alvarez-Meaza and de Alegría-Mancisidor2022). While this study only looked at #WorldEnvironmentDay rather than #sustainability, it revealed insights into the effectiveness of machine learning analysis to classify tweets related to sustainability. Furthermore, the study revealed that the World Environment Day hashtag was most often used to talk about official slogans from the UN, actions for the environment, social responsibility, Indian government, Pakistan, criticism of environmental degradation, restoration, and criticism of political responsibility for environmental degradation (Zarrabeitia-Bilbao et al., Reference Zarrabeitia-Bilbao, Rio-Belver, Alvarez-Meaza and de Alegría-Mancisidor2022).

2.2. Research gap

Preexisting research analyzes public discourse surrounding sustainability similarly to what we wish to explore. However, little to no research employs a combination of classical statistical analysis with the Latent Dirichlet Allocation methodology to analyze tweets under #sustainability on Twitter and view the intersection of these two methodologies. The aforementioned study conducted by Pilař et al. looks at #sustainability as does our research; however, the study does not use LDA analysis and only explores hashtags instead of overall topic analysis. The aforementioned study conducted by Reyes-Menendez et al. (Reference Reyes-Menendez, Saura and Alvarez-Alonso2018) analyzes major topics of discussion. However, the study explores only #WorldEnvironmentDay, and does not apply the same machine learning techniques and methodology as our study does. The aforementioned study conducted by Zarrabeitia-Bilbao et al. uses a similar methodology as our research does; however, the study explores only #WorldEnvironmentDay and does not use a keyword-frequency analysis and explore the intersection between classical statistical analysis and the machine learning model.

2.3. Objectives

This study aims to answer the research question: “What are the major topics of conversation under the #sustainability hashtag on Twitter?” by classifying the major topics of conversation to provide insight into user engagement patterns under the #sustainability hashtag on Twitter. These results provide valuable insights for those who wish to understand Twitter discourse surrounding sustainability and those who wish to use the results as a tool to promote sustainable content themselves and thus increase awareness for an important societal issue.

3.1. Research instruments

This study used Google Colab, a cloud-based Integrated Development Environment (IDE) in which all code was written and all data was analyzed. The only computer language used for all of the code was Python.

3.2. Data collection

To collect tweets from the Twitter platform, a web scraper Python library was used entitled “snscrape”. Snscrape is “a scraper for social networking services (SNS). It scrapes things like user profiles, hashtags, or searches and returns the discovered items, e.g. the relevant posts” (JustAnotherArchivist, 2023). Based on a Netflix web scraper that utilized snscrape (15EkbergC, 2022), the Python script collected 1000 random tweets per day. Only tweets written in English were collected.

The program to scrape Twitter posts collected the following information about each Twitter post: “UserID”, “Date”, “Username”, “TweetContent”, “Likes”, “Retweets”, “TimesQuoted”, “ReplyCount”, “FollowerCount”, “friendCount”, “mediaType”, “viewCount”, “Day”. UserID refers to the particular user’s identification number, which is given by default to all Twitter accounts; the date was collected following the format year-month-day; TweetContent referrers to the unfiltered content of the tweet; Likes was the number of likes the tweet received; Retweets was the number of retweets (shares) the tweet got; TimesQuoted was the number of replies that contained a quote from the post; ReplyCount is the number of replies; FollowerCount is the number of people the user was following; friendCount is the number of friends the user had; mediaType was simply an indication if the tweet had an image attached (most images were filtered out in preprocessing); viewCount is the number of views the tweet had, and Day was a relative day assigned to each tweet and can be disregarded.

3.3. Data preprocessing

To begin, all words in the tweetContent column of the dataset were converted to lowercase letters to ensure standardization among the data. Next, all tweets from users with less than 100 followers and 10 friends were removed in an attempt to remove any bots to get an accurate representation of what actual people are discussing. Next, all null values are cleared from the dataset to avoid problems with machine learning models or data analysis. URLs, Mentions, Emojis, and Reserved Words (RT, FAV) are removed, and Smileys are converted to text. Next, a word count column is created for the remaining tweet content using the unfiltered tweet content. Next, all stop words from the NLTK library are removed from the tweets, as they do not provide any insight into the topic of the tweets (Sebleier, 2010). Additional words removed include “sustainability”, “#sustainability”, “&”, “nan”, “rt”, “rt”, “fav”, “let”, “amp”, “re”, and “sustainable” as they all include words that are either non-descriptive, automatically generated by Twitter, or unhelpful for determining the topic (e.g., “sustainable” does not contribute to identifying posts under #sustainability). Next, all word endings are removed to revert the word to its lemma in a process called Lemmatization to improve the machine learning model. Otherwise, words such as “run” and “runs” would be identified as separate. Additionally, words are all separated from each other in a process called Tokenization, with words less than 3 characters being filtered out as they are not useful for analysis. Finally, special symbols and punctuation are removed as they are non-descriptive. All data are exported into a new file containing a “keywords” column with the preprocessed tweet content results stored for analysis.

3.4. Dataset analysis methodology

In order to gain an understanding of the data, basic descriptive statistics were calculated for the following attributes: Likes, Retweets, Times Quoted, Reply Count, View Count, Word Count, Follower Count, and Friend Count. These statistics include count (number of data points for each attribute), mean, standard deviation, minimum, maximum, 25%, 50%, and 75%. Next, histograms were plotted for the same metrics in order to understand the frequency of data points and the distribution of the data. Next, data points were plotted over time to identify any unusual changes on particular days in the dataset. Finally, notable correlations between variables were plotted on a scatter plot to help identify trends in the data.

3.5. Frequency analysis methodology

In the frequency analysis, the previously mentioned “keywords” column in the dataset was plotted to find the top 30 most frequent keywords. This allows for quantitative analysis of common words discussed. With the top 30 keywords, common words were grouped together into overall topics of discussion manually and used later in comparison to the output of the topic modeling machine learning model.

3.6. Topic modeling analysis methodology

In order to identify the major topics of conversation within the dataset, a machine learning model called Latent Dirichlet Allocation (LDA) was used. LDA is an unsupervised machine learning model that classifies major topics within a collection of words inputted. Unsupervised learning “uses machine learning algorithms to analyze and cluster unlabeled datasets. These algorithms discover hidden patterns or data groupings without the need for human intervention.” (IBM, 2023). The LDA model was first “trained” on the preprocessed data in the “keywords” column in the dataset. Next, The LDA algorithm was assigned a specific number of “topics” to separate the dataset into. The topics exported by the algorithm are not a traditional one-word classification of a common topic but rather the most relevant keywords that are deemed probable of belonging together in a topic (Kulshrestha, Reference Kulshrestha2020).

(1) $$ relevance\left( term\hskip0.35em w\hskip0.35em |\hskip0.35em topic\hskip0.35em t\right)={\lambda}^{\ast }p\left(w\hskip0.35em |\hskip0.35em t\right)+(1-\lambda {)}^{\ast }p\left(w\hskip0.35em |\hskip0.35em t\right)/p(w) $$

Keyword relevancy as seen in Equation (1) is a measure of the most relevant words within a topic (Sievert and Shirley, Reference Sievert and Shirley2014). Keyword relevancy within the model was automatically calculated using the LDA algorithm which functions as follows: first go through the collection of words and randomly assign each word in the collection to one of the topics (the number of topics is predefined when running the algorithm). Next in each document find the “proportion of words in [a] document that are assigned to [a] topic” (Kulshrestha, Reference Kulshrestha2020) next find the proportion of a word in a topic compared to the word in the overall document finally the probability of the word belonging to a topic is updated. This process is repeated several times until the assignment of words to a topic no longer changes to maximize accuracy (Kulshrestha, Reference Kulshrestha2020).

For this study, the number of topics to categorize the data into was set to 7, as trials were run with the number of topics set to 5–13, and the results for 7 topics yielded good “coherence” and “perplexity” scores, as well as the best results for distinct topics. Perplexity score “captures how surprised a model is of new data it has not seen before, and is measured as the normalized log-likelihood of a held-out test set.” (Kapadia, Reference Kapadia2022). Although this metric is not perfect for representing how accurate the model is, it is a good indicator of the model. The coherence score is a measure of “the degree of semantic similarity between high scoring words in the topic” (Kapadia, Reference Kapadia2022). The coherence score assists in evaluating how easy the topics are to identify. A high coherence and a low perplexity were optimal for this topic model analysis; however, topic model results and intertopic distance was also an important consideration when choosing the number of topics. Intertopic distance is “a visualization of the topics in a two-dimensional space” (SydneyF, 2021). The intertopic distance represents how distinct topics are from each other. Topics that are further away are more distinct in content. More distance in between topics represents a better fit of the model for the dataset.

4.1. Twitter metrics general statistics

Some The dataset collected after data pre-processing consisted of 37,728 rows (tweets) and 18 columns. The rows represent each individual Twitter post collected. The columns were each attribute, including “UserID”, “Date”, “Username”, “TweetContent”, “Likes”, “Retweets”, “TimesQuoted”, “ReplyCount”, “FollowerCount”, “friendCount”, “mediaType”, “viewCount”, “Day”, “wordCount”, and “keywords”. Analysis was first conducted on the following attributes: Likes, Retweets, Times Quoted, Reply Count, View Count, Word Count, Follower Count, and Friend Count.

First and foremost, all metrics have fully filled out data points, with 37,728 in each category, meaning the dataset is consistent with no missing values. In Table 1, the view count is the highest average metric, followed by the friend count, likes, and retweets. Average likes are notably higher than average retweets, suggesting that users are more likely to like tweets about sustainability than to share them with others. One possible inference is that people show interest in sustainability but do not often advocate for it on their platform. The majority of tweets (75%) had around 102 views, showing that the dataset consisted of a sufficient number of posts with the opportunity for user engagement.

Table 1. Standard twitter metrics data (n = 37,728)

Note. The initial number of tweets was 60,000. Of them, 22,272 posts did not meet the criteria, therefore leaving 37,728 posts.

4.2. Frequency analysis with histograms

Figure 1 displays histograms for each of the previous attributes mentioned. As can be seen, most figures (excluding the word count histogram) have a top frequency value of 0. This is likely due to the nature of the dataset, as it is a random sample. As is apparent, the majority of posts lie between 0 to around 15 likes, 0 to around 5 retweets, and 0 to around 4 replies. The majority of posts do not have any times quoted with the majority having 0, and some having 1 (keep in mind the y-axis scale is in the thousands). It is apparent that the majority of posts had between 20 and 40 words, with a large spike of posts with around 38 words.

Figure 1. Histograms of likes, retweets, times quoted, reply count, word count, and view count.

Note: The y-axis (Frequency) for each graph represents the number of tweets within the dataset.

4.3. Time series scatter plots

Figure 2 is a series of scatterplots displaying the Twitter metrics over the period of time they were gathered. As can be seen, the majority of graphs have a relatively steady trend for the value over time. This means that most days consisted of a relatively consistent number of likes, retweets, times quotes, replies, and views per day. However, some days had massive spikes in these metrics. The spikes suggest that a viral tweet was posted on that day. The word count graph remains consistently distributed over time, showing no change on any particular day in the amount people were typing, as expected.

Figure 2. Scattergrams of likes, retweets, times quoted, reply count, view count, word count over time.

Note: The x-axis displays the date in the format month-day that the posts were collected. The year for all of the posts was 2023.

4.4. Correlation analysis

Figures 3–6 show scatterplots of metrics plotted against each other. In Figure 3, although not extremely strong, there is a positive correlation between view count and likes due to likes generally increasing as view count increases. Likewise, Figure 4 shows a positive correlation between likes and retweets, as retweets increase proportionally to likes. A noticeable trait revealed in Figure 5 was that generally, follower count does not have much of an effect on view count. However, at around 1.25E06 followers, the view count does seem to spike in a few cases. A red dotted line represents this spike. Similarly, as shown in Figure 6, word count does not have much of an effect on likes; however, at 32 words, there seems to be a spike in likes (denoted by a red dotted line), indicating a sweet spot for how long of post viewers seem to enjoy.

Figure 3. View count plotted against likes.

Note: The graph on the left is a zoomed-out version of the graph on the right.

Figure 4. Likes plotted against retweets.

Note: The graph on the left is a zoomed-out version of the graph on the right.

Figure 5. View count plotted against follower count.

Note: The red dotted line represents an estimate of the x value where a spike occurs as referenced in the paper.

Figure 6. View count plotted against the word count.

Note: As referenced in the paper, the red dotted line represents an estimate of the x value where a spike occurs.