Work done for my Bachelor's thesis.
Evaluation of clustering algorithms on biomedical abstracts using internal metrics (Silhouette, Calinski-Harabasz, Davies-Bouldin).
Different combinations of the following methods have been experimented.
| Step | Methods |
|---|---|
| Text embedding | Bag-of-words Word2vec GloVe fastText BioWordVec MiniLM BioBERT PubMedBERT |
| Dimensionality reduction | PCA LSA NMF UMAP |
| Clustering | K-means Bisecting k-means Agglomerative DBSCAN OPTICS HDBSCAN |
The evaluation dataset is created using PubMed abstracts. Each entry of the dataset contains articles related to a disease.
pmids.json contains the PMIDs of the articles of an already defined dataset.
To fetch the actual abstracts, run:
python dataset_fetch.py --pmids=pmids.json --output=<output-path>
To create a custom dataset with different diseases, create a file with the wanted diseases (one per line) and run:
python dataset_pmids.py --diseases-list=<path-to-list> --output=<output-path>
The output contains the PMIDs of the dataset, which can be passed to dataset_fetch.py.
To start the evaluation, run:
python eval.py --dataset=<path-to-dataset> --embedding=<embedding>
Available values of <embedding> are
bow, word2vec, glove, fasttext, biowordvec, minilm, biobert, pubmedbert.