Gold Recovery Optimization 🏭✨

A Machine Learning project to predict gold recovery in an industrial process (rougher and final stages) to support production optimization and better decision-making.

---

Goal 🎯

Predict:

• rougher.output.recovery
• final.output.recovery

Using process features (concentrations, particle sizes, feed rates, etc.).

---

Data 📦

Files used:

• gold_recovery_train.csv → (16860, 86)
• gold_recovery_test.csv → (5856, 52)
• gold_recovery_full.csv → (22716, 86)

⚠️ The test set has fewer columns, so features are aligned using the intersection of columns between train and test (52 shared features).

---

Recovery Formula Check ✅

I validated the project’s recovery formula against rougher.output.recovery:

• MAE ≈ 9.30e-15 (basically 0)
  This confirms the target column is consistent with the expected calculation.

---

Metric 📏

Evaluation uses sMAPE and a weighted final score:

• sMAPE(rougher)
• sMAPE(final)
• sMAPE_final = 0.25 * sMAPE(rougher) + 0.75 * sMAPE(final)

---

Approach 🧠

1. Load + explore the datasets
2. Clean data (missing values, consistency checks, distribution review)
3. Align features between train and test
4. Train a multi-output model (2 targets)
5. Evaluate with train/validation split + cross-validation
6. Hyperparameter tuning with GridSearchCV (nested-CV style)

---

Results 📈

Main model: RandomForestRegressor (multi-output)

• Baseline (no tuning):

  • sMAPE_final ≈ 10.1897

• Best tuned model:

  • n_estimators=700, max_depth=None, min_samples_split=2
  • sMAPE(rougher) ≈ 8.035
  • sMAPE(final) ≈ 6.9762
  • ✅ sMAPE_final ≈ 7.2409

---

Suggested Repo Structure 🗂️