Amazon topic modelling

This is a data science project based on the Amazon Product Reviews Dataset. Our goal is to identify the major topics in the customer reviews of various products.

The data

The input folder contains the following data.

1. 7817_1.csv, which contains the Amazon product reviews data.
2. amazon_topic_model, which contains the optimised version of the BERTopic model that we built during this project. This was saved as a file via the joblib package and can be loaded with joblib.load(). Visualisation and analysis of this model can be found in the Conclusion section of the notebook amazon-topic-modelling.ipynb. Note: You need to download the notebook and view it locally in order to see the Plotly visualisations.
3. clean_docs, which contains the cleaned version of the Amazon product reviews data, as a list of strings. This was saved as a file using joblib.

Methodology

Below is a brief summary of the methodology. A more detailed methodology, complete with code snippets and plots, can be found in the notebook amazon-topic-modelling.ipynb.

Our goal is to perform topic modelling on the Amazon product reviews using BERTopic, identifying an optimal number of topics to split the data into, and identifying what the major topics in the reviews are. We apply the following techniques to achieve this.

1. Preprocessing. We first identify the title and body of each review in the dataset and compile them into single documents. We then remove non-alphabetical characters, remove duplicate reviews, identify the languages of the reviews and remove any reviews that are not in English. Finally, we correct any typos/misspellings in the reviews.
2. Topic modelling. We use BERTopic for our topic modelling. This involves multiple stages, each with its own separate model: embedding via SentenceTransformer using the all-MiniLM-L6-v2 LLM; dimensionality reduction via UMAP, clustering via HDBSCAN, tokenisation via CountVectorizer, weighting via ClassTfidfTransformer, and topic representation via MaximalMarginalRelevance. We start by using the default settings for each model to establish a baseline topic model.
3. Hyperparameter tuning. We tune the hyperparameters of the models listed above by iterating through 1000 random combinations of values, feeding them into BERTopic, and calculating the coherence score of the resulting topics. We pick the five highest coherence scores and display visualisations for the corresponding topic models, so that we can manually pick the best one.

Results

Running the script topic-modeller.py consistently achieves coherence scores between 0.71 and 0.76.

We identify six major topics: Voice Assistants, Television, Reading, Tablets, Accessories, and Audio. These are separated into distinct clusters in the dataset, with a potential Remote Controls sub-cluster of the Television topic.

See the amazon_topic_model file in the inputs folder for a sample optimised model constructed using this code. Load the amazon_topic_model and clean_docs files with joblib.load() and use the visualize_barchart() and visualize_documents() methods in BERTopic to visualise the topics. You can also see these visualisations in the Conclusion section of the notebook amazon-topic-modelling.ipynb (you need to download the notebook and view it locally in order to see the Plotly visualisations).