This project applies multiple machine learning models to predict customer satisfaction scores using behavioral and interaction data.

The project explores two main questions:

1. Which model best predicts customer satisfaction?
2. Which customer behavior variables are most important in predicting satisfaction?

To answer these questions, several models were implemented, including both custom-built algorithms and library-based machine learning models.

The dataset used is the Customer Experience Dataset from Kaggle.

It contains 1000 customer observations and several variables describing user interactions, behavior, and satisfaction.

In this project, the following variables are used.

• Num_Interactions
• Feedback_Score
• Products_Purchased
• Products_Viewed
• Time_Spent_on_Site

• Satisfaction_Score

The goal is to predict Satisfaction_Score using these five behavioral features.

Missing values are handled by filling them with the column mean.

The dataset is split into:

• 70% training data
• 30% testing data

Z-score normalization is applied:

X_normalized = (X - mean) / std

The mean and standard deviation are calculated only from the training data and then applied to both training and test datasets.

Five models are implemented and compared.

A multivariate linear regression model implemented manually using gradient descent.

The model iteratively updates parameters until the loss improvement falls below a threshold.

This model serves as a baseline model.

A custom implementation of KNN regression.

Steps:

1. Compute Euclidean distance between test and training samples
2. Select the k nearest neighbors
3. Predict using the average of their target values

In this project:

k = 3

Polynomial regression implemented using Scikit-learn.

PolynomialFeatures with degree 2 is used to allow the model to capture nonlinear relationships between variables.

A regression decision tree implemented manually.

The splitting rule is based on variance reduction, where the algorithm selects the feature and threshold that reduce the variance of the target variable the most.

This model is also used to compute feature importance.

A gradient boosting model implemented using XGBoost.

Key parameters include:

• n_estimators = 500
• learning_rate = 0.03
• max_depth = 3
• subsample = 0.8
• colsample_bytree = 0.8

XGBoost is used both for prediction performance comparison and feature importance analysis.

Model performance is evaluated using Root Mean Squared Error (RMSE).

RMSE = sqrt(mean((y - y_hat)^2))

Lower RMSE indicates better prediction performance.

Both training RMSE and testing RMSE are compared for all models.

The results show that:

• Linear Regression performs best on the test dataset
• Polynomial Regression performs similarly but slightly worse
• KNN shows some overfitting behavior
• Decision Tree performs moderately
• XGBoost provides competitive predictive performance

Feature importance is evaluated using two tree-based models:

• Decision Tree
• XGBoost

Both models suggest that the most important predictor of customer satisfaction is:

Time_S