GIC

GIC refines cluster assignments by iteratively improving proximity matrices derived from decision tree ensembles to generate more accurate dissimilarity measures for clustering unlabeled biological data.

Key Features:

• Iterative refinement: Repeatedly updates the proximity matrix and cluster assignments through successive cycles until convergence.
• Random Forest (RF) labeling: Trains a Random Forest using labels that differentiate between real and synthetic data units to produce the initial proximity matrix.
• Proximity-to-dissimilarity conversion: Converts the RF-derived proximity matrix into a dissimilarity matrix for downstream clustering.
• Cluster-guided retraining: Uses initial cluster assignments as labels to refine the RF model and update the proximity matrix.
• Compatibility with other ensembles: Applies the same iterative approach to other base procedures such as the Extremely Randomized Tree ensemble.
• Quantitative evaluation: Assesses cluster cohesion and separation using the Silhouette Score.
• Focus on non-geometric relationships: Targets dissimilarity measurement for unlabeled units where geometric measures like Euclidean distance may not capture complex relationships.

Scientific Applications:

• Clustering of unlabeled biological data: Enhances cluster discovery in complex biological datasets using ensemble-derived proximities.
• Improving non-Euclidean clustering: Provides improved dissimilarity measures when traditional geometric distances (e.g., Euclidean) are insufficient.
• Ensemble-proximity based analyses: Integrates with analyses that leverage Random Forest or Extremely Randomized Tree proximities for downstream clustering evaluation.

Methodology:

Train a Random Forest with labels distinguishing real and synthetic units to obtain a proximity matrix; convert the proximity matrix to a dissimilarity matrix; derive initial cluster assignments; use those cluster labels to retrain the RF and update the proximity matrix; iterate this cycle until convergence; evaluate cluster quality with the Silhouette Score; the same procedure can be applied using an Extremely Randomized Tree ensemble.