Unsupervised random forests

Area of Science:

• Machine Learning
• Bioinformatics
• Data Mining

Background:

• Unsupervised machine learning algorithms are crucial for identifying patterns in complex datasets.
• Existing methods may struggle with mixed data types (categorical and continuous).

Purpose of the Study:

• Introduce sidClustering, a new random forests-based unsupervised machine learning algorithm.
• Demonstrate its effectiveness in identifying clusters from mixed data types.

Main Methods:

• sidClustering employs feature sidification to create mutually exclusive ranges and interaction features.
• A multivariate random forest model predicts these sidified features.
• Multivariate impurity splitting is utilized for cluster identification.

Main Results:

• The sidification process is unique and reproducible.
• sidClustering successfully identifies clusters arising from both categorical and continuous variables.
• The algorithm retains the advantages of traditional random forests.

Conclusions:

• sidClustering offers a robust approach for unsupervised clustering with mixed data types.
• The method is validated on simulated and real-world datasets, including cancer and cardiovascular patient data.