rf_model

Amazon Order Total Prediction & Segmentation

This project applies end-to-end data science techniques to an e-commerce (Amazon-like) dataset in order to:

1.Predict the final order total (TotalAmount)** using regression models.
2.Segment orders/customers into spending tiers** (low / mid / high) using classification.

The work includes data cleaning, exploratory data analysis (EDA), feature engineering, clustering, regression modeling, classification modeling, and model deployment.

1. Dataset Description

Source: Synthetic Amazon-style orders dataset (Amazon.csv). Size: 100,000 rows. Input features (examples): Product: ProductID, ProductName, Category, Brand Pricing: UnitPrice, Discount, Tax, ShippingCost Quantity: Quantity Transaction info: OrderDate, PaymentMethod, OrderStatus, City, Country, SellerID

• Target (Regression):
  • TotalAmount – final amount paid for the order.

The main research question: Can we accurately predict the final order total (TotalAmount) of an Amazon order using customer, product, pricing, and transaction-level features?**

1. Exploratory Data Analysis (EDA)

Key steps:

Data cleaning:

• Parsed OrderDate as datetime.
• Created time-based features such as Week (ISO week number).
• Checked for missing values, duplicated rows, and basic consistency.

Descriptive statistics:

• UnitPrice:
  • Mean ≈ 303, Std ≈ 172, Min = 5, Max ≈ 600.
  • Strong positive correlation with TotalAmount (~0.72).
• Quantity:
  • Mostly between 1 and 5 units.
  • TotalAmount grows almost linearly with Quantity.

-Outlier analysis:

• Some very high totals due to combination of high UnitPrice & Quantity.

• Kept most outliers, since they represent genuine high-value orders, which are important for modeling.

• Seasonality / Black Friday analysis:

  • Computed Week from OrderDate.
  • Found that Week 47 (Black Friday) and Week 48 (Cyber Monday) have slightly higher average TotalAmount than other weeks.
  • Week 47 shows the clearest spike relative to the global weekly average.

1. Feature Engineering I added multiple engineered features to improve predictive power:

3.1 Time-based features

• Week: ISO week number for each order.
• IsHighSeason: 1 if Week in {47, 48} (Black Friday/Cyber Monday), else 0.

3.2 Price-related flags

• IsExpensiveItem: 1 if UnitPrice above median unit price, else 0.

Note: To avoid data leakage, we did not include direct formulas that reconstruct TotalAmount (e.g., ValueBeforeTax = UnitPrice * Quantity * (1 − Discount)) in the final models.

3.3 Clustering-based features (Unsupervised Learning)

I applied K-Means clustering on standardized numerical features:

• Features used for clustering:

  • UnitPrice, Quantity, Discount, Tax, ShippingCost

• Steps:

  1. Standardized the features with StandardScaler.
  2. Trained KMeans(n_clusters=5, random_state=42).
  3. Added:
     • ClusterID – the cluster assignment for each order.
     • DistanceToCentroid – Euclidean distance from each point to its cluster center.

• PCA Visualization:

  • Reduced the clustering space to 2D using PCA (PCA1, PCA2) and plotted the clusters.
  • Clusters showed clear separation, corresponding roughly to:
    • High unit price & high quantity orders (high spenders).
    • Low unit price & low quantity orders (low spenders).
    • Orders with strong discounts.
    • Orders with higher taxes/shipping.

These cluster-derived features were then used as additional inputs to the regression and classification models.

1. Regression Modeling (Predicting TotalAmount)

4.1 Baseline Linear Regression

• Features used:
  • Numeric: UnitPrice, Quantity, Discount, Tax, ShippingCost, Week
  • Categorical: Category (one-hot encoded)
• Train/Test Split: 80% / 20%, random_state=42
• Performance:
  • R² ≈ 0.909
  • MAE ≈ 166
  • RMSE ≈ 217

Interpretation:
The baseline model already explains ≈91% of the variance in TotalAmount, mainly driven by Quantity, UnitPrice, and Tax. Category has only a minor effect.

4.2 Linear Regression with Feature Engineering & Clusters

Using:

• All original numerical features
• Week, IsHighSeason, IsExpensiveItem, ClusterID, DistanceToCentroid
• One-hot encoded Category

Performance:

• R² ≈ 0.920
• MAE ≈ 156
• RMSE ≈ 204

This shows a modest improvement over the baseline, primarily due to cluster-based segmentation and simple time-based flags.

4.3 Tree-Based Regression Models

I then trained two advanced models on the engineered dataset:

1. Random Forest Regressor
2. Gradient Boosting Regressor

Results (approximate):

• Linear Regression (engineered):
  • R² ≈ 0.92, RMSE ≈ 204
• Random Forest Regressor:
  • R² ≈ 0.9999, RMSE ≈ 6.6
• Gradient Boosting Regressor:
  • R² ≈ 0.996, RMSE ≈ 46.1

Because the underlying relationship between features and TotalAmount is nearly deterministic (a pricing formula applied consistently), tree-based models can almost perfectly reconstruct the function mapping inputs to target.

Winning Regression Model: Random Forest Regressor — near-perfect performance, robust to non-linearities and feature interactions.

5. Classification Task: Regression-to-Classification

I reframed the problem as a multi-class classification task, predicting spending tier class instead of continuous TotalAmount.

5.1 Target Transformation (Part 7.1)

I converted TotalAmount into three classes using quantile-based binning on the training set:

• Class 0: bottom 33% (TotalAmount < 443.53)
• Class 1: middle 33% (443.53 ≤ TotalAmount < 1088.52)
• Class 2: top 33% (TotalAmount ≥ 1088.52)

This ensures balanced classes and has a meaningful business interpretation: low-, mid-, and high-spending orders.

5.2 Class Balance Check (Part 7.2)

Class distribution:

• Train:
  • Class 0 ≈ 33%
  • Class 1 ≈ 33%
  • Class 2 ≈ 34%
• Test:
  • Very similar proportions (≈33% each)

Since the classes are well-balanced, accuracy and macro F1 are both appropriate evaluation metrics.

---

6. Classification Models (Part 8)

I trained three different classifiers using the same engineered feature set:

• Numeric: UnitPrice, Quantity, Discount, Tax, ShippingCost, Week, IsHighSeason, IsExpensiveItem, ClusterID, DistanceToCentroid
• Categorical: Category (one-hot encoded)

All models used a ColumnTransformer + Pipeline for preprocessing and training.

6.1 Logistic Regression (Multinomial)

• Reasonable performance as a baseline.
• Main issues:
  • Frequently confuses the middle class (1) with lower (0) and higher (2) tiers.
• Limitation:
  • Struggles to capture non-linear relationships in the data.

6.2 Random Forest Classifier

• Best overall performance:
  • Very high accuracy and macro F1.
  • Confusion matrix shows almost perfect classification.
• Interpretation:
  • Effectively captures complex patterns and interactions.
  • Leverages clustering-based features (ClusterID, DistanceToCentroid) and raw numeric inputs.

6.3 Gradient Boosting Classifier

• Strong performance:
  • High precision and recall across classes.
  • Slightly more errors than Random Forest, mainly between neighboring classes (1 ↔ 2).
• Still a very competitive model, but not the top performer.

Critical mistakes considered: In a business context, false negatives on the high-spending class (Class 2 → Class 0/1) are more costly than false positives, because they represent missed opportunities for targeting valuable customers.

Winning Classification Model: Random Forest Classifier — highest accuracy, macro F1, and most balanced confusion matrix.

7. Exported Models

The following models were exported as .pkl files for deployment:

• winning_model_random_forest.pkl
  → Random Forest regression model (pipeline including preprocessing).
• rf_classifier_model.pkl
  → Random Forest classification model (pipeline including preprocessing).