This study applied the Latent Dirichlet Allocation (LDA) algorithm for topic modelling, a well-established technique in computational content analysis. Introduced by Blei et al., LDA has been widely recognised for its effectiveness in uncovering thematic structures from large datasets of textual data. A key advantage of LDA is its robust performance on datasets comprising short texts, rendering it especially suitable for analysing social media data, which typically consists of brief, varied-length documents.

In evaluating topic models, the number of topics was adjusted from 10 to 60 in increments of 10. Utilising Gensim (open-source library) a computational evaluation of the topic models was performed using coherence scores, guiding the selection of an optimal model of 10 topics. A qualitative analysis was also conduced to ensure the topics made sematic sense to domain experts.

Building on the topic modelling analysis, sentiment analysis was performed to explore the emotional valence associated with each identified topic. This analysis was performed using the Valence Aware Dictionary and sEntiment Reasoner (VADER) sentiment analysis tool, a lexicon and rule-based sentiment analysis framework specifically attuned to sentiments expressed in social media. Each document within the dataset was assigned a compound sentiment score, a normalised, weighted composite score, encapsulating the overall sentiment of a document, ranging from-1 (most negative) to +1 (most positive). The analysis was extended to compute the average sentiment score for documents within each topic, which had been identified through the earlier LDA analysis.

A qualitative examination of the top 5 most positive and negative posts for each topic was performed, based on sentiment scores. Systematic identification and qualitative assessment of these posts allowed a deeper understanding of the underlying themes and emotional expressions that contributed to their extreme sentiment scores.

To further dissect the emotional undertones of each topic, a model was employed to identify words most strongly associated with positive and negative sentiments within each topic category. This approach aimed to pinpoint specific language cues that signal emotional valence in the context of the topics derived from the LDA analysis. Two models were made for both negative and positive sentiment, Term Frequency-Inverse Document Frequency (TF-IDF) vectorisation was applied to transform the text data. The TF-IDF model, configured to disregard the most common English stop words and to limit the analysis to the 5000 most relevant features, allowed quantification of the importance of words in relation to their sentiment scores. A linear regression model was then trained for each topic, using the TF-IDF vectors as predictors for the sentiment scores.

To better understand the relationships in the data and their potential influence on content and sentiment, basic network analysis was performed on the data using the popular and versatile Python pack NetworkX, designed for the construction, manipulation, and analysis of the structure and dynamics of complex networks. A directed graph is a type of graph where each edge has a direction, indicating a one-way relationship between two nodes (users). In this case, the direction represents posts and interaction directed at a user in the subreddit community; therefore, it can be seen as proxy for their popularity within a social network. A directed graph was created across the whole network, as well as by the 10 topics identified by the topic modelling. Therefore, 11 graphs were analysed in total. In-degree centrality statistics, which calculate how many interactions are directed at a user, were performed for each of the graphs. Furthermore, the dataset was categorised into ‘top users’ and ‘all other users’ for the purpose of analysing discourse. This was performed across all graphs, including the entire network and by topic.