Link to the GitHub repository

Customer Category Predictor - Process Overview

Below is an overview of the process used in this project:

• Data Preparation
  • Imported the dataset (train.csv) and examined its structure.
  • Handled missing values by creating dedicated dummy variables to indicate whether a cell was NaN or not.
  • Converted categorical variables into multiple dummy features to capture all distinct categories.
• Feature Engineering
  • Retained only relevant numerical and dummy variables for modeling.
  • Removed unnecessary identifiers (like customer_id) and original categorical columns.
• Train/Test Split
  • Separated the dataset into training (80%) and testing (20%) subsets.
  • Ensured a fixed random state for reproducibility.
• Modeling
  • Logistic Regression
    • Served as a baseline model to quickly assess predictive performance.
    • Provided straightforward interpretability of coefficients.
  • Support Vector Machine (SVM)
    • Conducted hyperparameter tuning with GridSearchCV to identify optimal C and gamma values.
    • Evaluated multiple parameter combinations for improved performance.
• Performance Evaluation
  • Computed confusion matrices to examine true positives, true negatives, false positives, and false negatives.
  • Calculated accuracy, recall, precision, and F1 scores to provide a comprehensive view of each model’s effectiveness.
• ROC Curve Analysis
  • Plotted ROC curves for both Logistic Regression and SVM on the same graph.
  • Measured the Area Under the Curve (AUC) to summarize model performance at various classification thresholds.
  • A higher AUC indicates that the model is generally better at distinguishing between the classes across different thresholds.
  • The diagonal line (from (0,0) to (1,1)) in the ROC plot represents a random guess baseline. The closer a model’s ROC curve is to the top-left corner, the more effective it is at discriminating between positive and negative classes.