pareto-frontier-optimization

Grid Search Space

Create all combinations of:

Parameter	Range	Values
min_samples	3–9	[3, 4, 5, 6, 7, 8, 9]
epsilon	4–24 (step 2)	[4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24]
shape_weight	0.9–1.9 (step 0.1)	[0.9, 1.0, 1.1, 1.2, 1.3, 1.4, 1.5, 1.6, 1.7, 1.8, 1.9]

Total combinations: 7 × 11 × 11 = 847

Evaluation Pipeline

For each hyperparameter combination (min_samples, epsilon, shape_weight):

1. Loop over all unique images from expert dataset (using file_rad)

2. For each image:

   • Extract citizen science points for this image
   • If no citizen science points exist:
     • Set F1 = 0.0, delta = NaN
     • Continue to next image
   • Run DBSCAN with current hyperparameters on citizen science points
   • If no clusters found (all points noise or single cluster with < min_samples):
     • Set F1 = 0.0, delta = NaN
     • Continue to next image
   • Compute cluster centroids
   • Extract expert points for this image
   • Perform greedy matching of centroids to expert points
   • Compute F1 score and delta metric for this image

3. Aggregate across images:

   • Average F1: Include all F1 values (including 0.0)
   • Average delta: Only include non-NaN values
   • If all delta values are NaN (no matches found in any image), set average delta = NaN

4. Filter results:

   • Keep only results where average F1 > 0.5

Pareto Frontier

Identify Pareto-optimal solutions:

A solution is Pareto-optimal if:

• No other solution has both higher F1 and lower delta
• It is not dominated on both objectives

Optimization goals:

• Maximize F1 score (higher is better)
• Minimize delta (lower is better)

Parallelization

Recommended approach:

• Use multiprocessing.Pool or joblib.Parallel to evaluate hyperparameter combinations in parallel
• Each worker processes one or more complete combinations (all images for one hyperparameter set)
• Collect results and apply filtering/Pareto frontier detection on main process