Use tmtoolkit to fit multiple LDA models in parallel.

[SeppeDeWinter]

Using this change it is easier for a user to explore multiple parameters for LDA modeling.

The function run_topic_modeling now accepts a list of values for

• n_topics
• alpha
• eta

When a list of values is given for one or more of these parameters multiple models will be fit in parallel to allow the user to explore the most optimal hyperparameters. This is done under the hood using tmtoolkit. The function now returns a list of topic models along with quality metrics.

After a model has been selected this can be added to the AnnData object using the new function add_topic_modeling_result.

tmtoolkit provides multiple functionalities to evaluate topic models, see https://tmtoolkit.readthedocs.io/en/latest/topic_modeling.html#Evaluation-of-topic-models.

For this reason the loglikelihood function is removed given that it is already implement in tmtoolkit.

This change does introduce a new dependency. We could consider making the topic modeling dependencies optional given that it is a more advanced use case.

pip install tfmindi[topic] for instance.

[LukasMahieu]

Okay, looks interesting, will have to test this.
We already have an "evaluate_topic_models", which does something similar but unoptimized and only for a range over n_topics. Maybe this functionality fits better there?

[SeppeDeWinter]

Okay, looks interesting, will have to test this. We already have an "evaluate_topic_models", which does something similar but unoptimized and only for a range over n_topics. Maybe this functionality fits better there?

True! That way we don't have to change the API.
Not sure how easy it is to automatically detect the optimal parameters though, although I also have not fully explored how strong the results differ between parameters.