This project focuses on developing a predictive model to determine a user's purchase intention when visiting an e-commerce website. Various machine learning algorithms and techniques are employed to analyze session data and predict whether a user will make a transaction.
The dataset comprises 12,330 sessions obtained from the UCI ML repository, representing user interactions on an e-commerce platform over a year. The dataset includes 10,422 negative class samples (no purchase) and 1,908 positive class samples (purchase).
- Feature Selection: Using SelectKBest to select the most informative features.
- Encoding: Categorical variables are encoded using LabelEncoder.
- Scaling: Numerical features are scaled using StandardScaler.
- Shuffling: Data is shuffled to remove bias.
The training dataset is divided into batches with different ratios of positive to negative samples (1:1, 1:2, 1:3, 1:4) to analyze model performance under varying class distributions.
| Batch | Algorithm | Accuracy | Precision | Recall | F1 Score |
|---|---|---|---|---|---|
| 1:1 | Decision Tree | 79.64% | - | - | - |
| 1:2 | Random Forest | 86.98% | 58.52% | 75.18% | 65.81% |
| 1:3 | XGBoost | 88.52% | 64.41% | 69.58% | 66.90% |
| 1:4 | Gradient Boosting | 87.88% | 64.21% | 61.55% | 62.86% |
- XGBoost consistently outperforms other algorithms across different batch sizes.
- Further optimization may enhance model performance.
- XGBoost is recommended for predicting purchase intent in e-commerce sessions.
- To run our Decision Tree Classifier - run /main/decision_tree.py
- To run our SVM Classifier - run /main/svmGK.py
- To run Scikit Deciison Tree Classifier - you can comment PART-a in the file /main/decision_tree.py and run PART-b
- To run Scikit Gradient Boosting - run /main/gradient_boosting.py
- To run Logistic Regression - run /main/logistic_regression.py
- To run Random Forest - run /main/random_forest.py
Note: The original dataset is online_shoppers_intension.csv. The /main/data folder contains all the batches data, and the /main/test folder contains separate test sets.