I set my own umap (with random_state=0), hdbscan (with random_state=0), ctfidf_model, and representation_model, and then passed these four modules into the BERTopic() function. However, my results are different every time I run the code. What should I do?

[superseanyoung]

topic_model = BERTopic(embedding_model=transformer_model,
min_topic_size=3,
verbose=True,
umap_model=umap_model,
hdbscan_model=hdbscan_model,
ctfidf_model=ctfidf_model,
representation_model=representation_model,
#top_n_words=10,
#min_topic_size=10,
#nr_topics=None,
#low_memory=False,
#calculate_probabilities=True
)
topics,probs= topic_model.fit_transform(sentences,embeddings=embeddings)

[MaartenGr]

Could you create an issue for this as a bug and fill in all the information listed there? It's difficult to say what is happening without more context. Also, make sure you provide the full code.

[superseanyoung]

As long as the "random_state" parameter in the UMAP model is set, can it make the entire model non-random?

[MaartenGr]

Perhaps, but as mentioned I will need more context. I have no clue what's inside any of the models you passed to BERTopic, the version you are using, the OS between multiple runs, etc. Could you follow the steps I mentioned above?

[superseanyoung]

I solved the problem and found that as soon as you set random_state in umap, the randomness of the model is controlled, and the randomness of the subsequent outliers is controlled.

[superseanyoung]

Since I don't think this is a bug or anything, I just wanted to discuss it with you,

[MaartenGr]

That is correct, the random_state in UMAP it typically the one that controls the randomness but that might change if you use different submodels.