Question about reproducibility

[powerhorse1986]

Dear Maarten,

Thanks so much for developing this great tool. Recently, I am working on some abstracts downloaded from PubMed using BERTopic. I noticed the number of topics were different on different computers. On my work laptop, which is a DELL windows machine, only 3 topics were detected. However, I run the same code on my own machine, a MacBook Pro with M1 chip, I got 18 topics. The code was shown as the following:

`umap_model = UMAP(n_neighbors = 5, n_components = 3, min_dist = 0.0, metric = 'cosine', random_state = 42)
hdbscan_model = HDBSCAN(min_cluster_size = 15, metric = 'euclidean', cluster_selection_method = 'eom',
prediction_data = True)

vectorizer

vectorizer_model = CountVectorizer(stop_words="english")

Text generation with chatGPT

client = openai.OpenAI(api_key = token)
#chatgpt = OpenAI(client, model = 'gpt-3.5-turbo', prompt = prompt, chat = True)
chatgpt = OpenAI(client, model = 'gpt-4-turbo', prompt = prompt, chat = True)

All representation models

representation_model = {
#'KeyBERT': keybert,
'ChatGPT4_Turbo': chatgpt
#'MMR': mmr
}
topic_model = BERTopic(
# Vectorizer
vectorizer_model = vectorizer_model,

# Sub-models
embedding_model = embedding_model,
umap_model = umap_model,
hdbscan_model = hdbscan_model,
representation_model = representation_model,
calculate_probabilities = True,

# Hyperparameters
top_n_words = 100,
verbose = True

)`

I guess maybe different computer architectures caused this problem. May I know your opinion on this problem? Thank you.

[MaartenGr]

Ah, different OS will indeed result in different results. I'm not entirely sure but I remember that either was a result of HDBSCAN or UMAP, not any of the other models I believe.

[powerhorse1986]

Hi Maarten,

Thank you so much for your response.