Datarock Coding Test

Aim

Can we use the same geochemical data and labels to generate a predictive model for future drill holes which can label samples on whether they are in class A or class B?

More data has been acquired since the geochemist completed her work — can we predict labels onto these data points (labelled “?”)?

Read Data

• Understand its structure and content
• Check for errors, missing values, and outliers
• Clean the data and prepare it for modelling

Fields (type):

• Unique_ID (object)
• holeid (object)
• from (int64)
• to (float64)
• As (float64)
• Au (object)
• Pb (float64)
• Fe (float64)
• Mo (float64)
• S (float64)
• Cu (float64)
• Zn (float64)
• Class (object)

Dataset:

• 140 Drillholes
• 4,004 Classified data points
• 767 Unclassified data points

Notes:

• Missing collar coordinates
• Missing units
• Some missing values (As Missing 1503 values)
• Not all intervals present

QAQC:

• Missing collar coordinates
• Missing units
• <0.005 in Au replaced with half detection limit 0.0025
• Au converted from object to float
• -999 values removed
• Some missing values (e.g. in As)
• Not all intervals present

Visualisation

• Plotted holeid along the x-axis and depth on the y-axis
• Created two versions: one coloured by Pb, one by Class

Figure: Drillholes coloured by Pb values

Figure: Drillholes coloured by known class

Analysis

Pb Cut-off Grade

• Iterated through Pb cut-off values to find the one best aligned with Class
• Pb = 160 gave the highest accuracy: 78.4%

Figure: Pb cutoff accuracy test

Figure: Predicted Classes using Pb cut-off 160

Weighted Scoring

• Used a weighted sum of key elements (Pb, Mo, Au, As) based on correlation
• Elements S, Zn, Cu, and Fe were excluded due to low correlation
• Accuracy improved slightly to 78.91%

Figure: Weighted scoring accuracy

Figure: Predicted Classes using weighted scoring

XGBoost

• Applied XGBoost to model non-linear relationships and interactions
• Validated using 20% random test split
• Accuracy improved to 84.52%

Figure: Confusion matrix for XGBoost model

Figure: Predicted classes using XGBoost

XGBoost + Depth Intervals

• Added depth interval as an additional parameter to model spatial relationships
• Final model accuracy: 93.51%

Figure: Predicted classes using XGBoost model with depth intervals

Figure: Predicted classes using XGBoost model with depth intervals

Further Improvements

• Apply collar location to plot drillholes in 3D space and better understand spatial relationships
• Validate models by removing entire drillholes instead of random samples
• Compare with other ML methods (e.g. Random Forest, SVM)
• Include geological data such as lithology, alteration, and structure