- We have 181630 rows of data of various steel items having following features: 'Id', 'item_date', 'quantity tons', 'customer', 'country', 'status', 'item type', 'application', 'thickness', 'width', 'material_ref’, 'product_ref', 'delivery date', 'selling_price'.
- The project performs a regression analysis with selling_price feature to be predicted.
- We divide data into 90-10% train-test split respectively. Analysis and model is done on the train set and tested on unseen test data.
- Selected Model: XGBoost
- Hyperparameter search method: Bayesian Search CV
- Best r2 score:
- Test r2 score: 0.90
- Train r2 score: 0.92
- Features used:
- ‘quantity tons’
- ‘application’
- ‘thickness’
- ‘width’
- ‘product_ref’
- ‘delivery date’
- ‘status’ .
- Categorical features shortlisting criteria: ANOVA test F statistical scores. (for details, refer EDA notebook)
- High value of ANOVA F statistics shows that, the mean and spread of data among different categories are easily separable so such features gives better r2 score and explaining variance for model becomes easy.
- Numerical features were power transformed and standardized to get near gaussian like distribution.
- Categorical features were one-hot encoded.
- Total input features: 98
- Important Features:
- ‘status’
- ‘delivery date’
| Model | Train r2 score | Val r2 score | Test r2 score | Hyperparameter search method |
|---|---|---|---|---|
| Linear Regresion | 0.83 | 0.83 | 0.82 | Grid Search |
| Decision Tree | 0.86 | 0.80 | not performed | Grid Search |
| Random Forest | 0.70 | 0.69 | not performed | Bayesian Search |
| XGBoost | 0.92 | 0.89 | 0.90 | Bayesian Search |
| Neural Network | 0.90 | 0.89 | 0.89 | None |