Overview

In this step, DORA compares your embedded input data to clustered learning points, which are grouped for training and validation. This process drives the supervised learning model that predicts the likelihood of mineralization across your Area of Interest (AOI), directly influencing accuracy and generalization.

What is a Predictive Model?

In this step, DORA uses a two-class classification model to learn from your embedded rasters and labeled learning points. The goal is to predict where similar mineralization patterns might occur across your AOI.

To make these predictions, DORA divides your learning points into spatial clusters, which are groups of geographically close data points. These clusters are used to separate training and testing data in a way that reflects real-world exploration: the model learns from one area and is then tested on a different, geologically distinct one.

DORA applies a spatial cross-validation technique, rotating through each cluster by holding it out for testing while training on the others. This clustering ensures that training and testing doesn't occur on neighboring points, which helps avoid misleading results due to spatial correlation in the data.

Why This Step Matters

Geoscience Perspective

Clustering helps the model account for spatial variation in geology. Adjusting the Minimum Cluster Size and Minimum Number of Points allows you to tailor the model to reflect local complexity or uniformity. For example, smaller clusters in structurally complex settings can help the model detect subtle mineral signatures.

AI Perspective

This step defines how your model learns and how its accuracy is validated. By dividing learning points into spatially distinct clusters, DORA prevents training and testing on neighboring data, reducing the risk of overfitting and improving the model’s ability to generalize across the AOI.

DORA uses spatial cross-validation, rotating through each cluster to test the model on unseen data. This process provides a more realistic measure of performance by evaluating how well the model predicts true and false positives and negatives across different regions. The quality of your clustering directly impacts how transferable and trustworthy your predictions are in new, un-sampled areas.

💡 Tip: Understanding the Confusion Matrix

One of the three output graphs generated in the next step is the Prediction Accuracy/Confusion Matrix, alongside the VRIFY Prospectivity Score (VPS). It summarizes how well your model predicts true positives and negatives across all train-test clusters. This is especially helpful for spotting areas where the model performs well overall but may struggle in zones that are geologically important.

Step by Step Instructions

Open Build Predictive Model
- From the experiment setup panel, click Step 5: Build Predictive Model.
Set the Minimum Cluster Size
- This parameter controls how far the model can search to group data points together:
Set the Minimum Number of Points
- This parameter defines the fewest samples required to form a valid cluster:
(Optional) Use Advanced Settings
- These are pre-set to recommended defaults for most experiments. See Tips & Considerations below for more information on adjusting these settings:
Generate Your VRIFY Prospectivity Score (VPS)
- Click Generate VRIFY Prospectivity Score to create your prospectivity predictions.

Tips & Considerations

Default Settings Recommended, But Experimentation is Encouraged

For most users, the default Advanced Settings are sufficient. These settings are designed to help fine-tune the model and address overfitting or underfitting issues. That said, you won’t break anything by experimenting! If you're curious, try adjusting the settings and observe how they affect your results.

Learn more in the Advanced Features for Building Predictive Models article or click the associated tooltip.