Online Shoppers Purchasing Intention Clustering

Isaiah Jenkins

Project Overview

This project analyzes the Online Shoppers Purchasing Intention dataset from the UC Irvine Machine Learning Repository to segment users into shoppers and non-shoppers using clustering techniques. The dataset comprises 12,330 sessions, with 84.5% (10,422) non-shoppers (Revenue=False) and 15.5% (1,908) shoppers (Revenue=True). Using Python libraries (Pandas, NumPy, Scikit-learn, Matplotlib, Seaborn), the analysis involves data preprocessing, applying clustering models (KMeans and Agglomerative Clustering with ward, complete, average, and single linkages), and evaluating their ability to distinguish shoppers from non-shoppers despite class imbalance.

Dataset

The dataset, online_shoppers_intention.csv, includes 18 features (10 numerical, 8 categorical):

• Numerical Features:
  • Administrative, Administrative_Duration, Informational, Informational_Duration, ProductRelated, ProductRelated_Duration: Page visits and time spent in each category.
  • BounceRates, ExitRates, PageValues: Google Analytics metrics for bounce rate, exit rate, and page value.
  • SpecialDay: Proximity to special days (e.g., Valentine’s Day).
• Categorical Features:
  • Month, OperatingSystems, Browser, Region, TrafficType, VisitorType, Weekend: General session and user information.
• Target Variable:
  • Revenue: Binary label indicating purchase (True) or no purchase (False).
• Preprocessing:
  • Dropped less relevant features: SpecialDay, Month, Browser, Region, TrafficType, VisitorType, Weekend.
  • Applied log transformation to handle skewed numerical features.
  • Scaled features using StandardScaler for clustering.

Analysis

The analysis followed these steps:

1. Data Exploration: Examined dataset structure (12,330 rows, 18 columns), confirmed no missing values, and computed descriptive statistics and value counts.
2. Data Visualization: Used an elbow plot to assess optimal cluster numbers, selecting 2 clusters to align with the goal of separating shoppers from non-shoppers.
3. Feature Engineering: Dropped 7 features, applied log transformations, and scaled numerical features.
4. Clustering Models:
   • KMeans: Configured with 2 clusters; results indicated 1,440 shoppers and 5,744 non-shoppers.
   • Agglomerative Clustering: Tested with ward, complete, average, and single linkages; results ranged from 0 to 1,908 shoppers identified.
5. Evaluation: Compared cluster assignments to the Revenue label to assess segmentation effectiveness.

Summary of Models

The analysis evaluated KMeans and Agglomerative Clustering (ward, complete, average, single linkages) for segmenting shoppers from non-shoppers. KMeans identified 1,440 shoppers (75.5% of 1,908 true shoppers) and 5,744 non-shoppers, achieving a balanced cluster ratio. Agglomerative Clustering with ward and average linkages grouped 1,905 shoppers (99.8%) into one cluster but with high imbalance (94.3% of sessions), while complete linkage identified 1,360 shoppers (71.3%) with better balance but lower accuracy. Single linkage was ineffective, assigning 99.99% of sessions (1,908 shoppers, 10,421 non-shoppers) to one cluster. KMeans outperformed Agglomerative Clustering variations by providing more balanced and effective segmentation, leveraging its simplicity to better handle class imbalance.

Key Findings

• Model Performance: KMeans achieved the best balance, identifying 75.5% of shoppers with a cluster ratio closer to the dataset’s distribution. Agglomerative Clustering, especially with single linkage, struggled with severe imbalance, often failing to separate shoppers effectively.
• Feature Insights: Features like ProductRelated, ProductRelated_Duration, and PageValues likely drove shopper identification due to their correlation with purchasing behavior.
• Challenges: The dataset’s class imbalance (84.5% non-shoppers) hindered Agglomerative Clustering’s ability to form distinct shopper clusters.
• Takeaway: KMeans’ centroid-based approach proved more effective than hierarchical methods, highlighting the value of simplicity in handling imbalanced data.

Installation

To run this project, install the required dependencies:

pip install pandas numpy matplotlib seaborn scikit-learn

Download the online_shoppers_intention.csv dataset from the UCI Machine Learning Repository and place it in the data/ directory.

Usage

1. Clone the repository:

   git clone https://github.com/your-username/online-shoppers-clustering.git
   cd online-shoppers-clustering

2. Set up the dataset:

   • Place online_shoppers_intention.csv in the data/ directory.

3. Run the Jupyter Notebook:

   jupyter notebook PurchasingIntentionClustering-2.ipynb

4. Follow the notebook to explore data, visualize the elbow plot, apply clustering models, and review results.

Next Steps

• Preprocessing Enhancements: Explore robust scaling or outlier trimming beyond log transformation to address skewed features.
• Model Exploration: Test additional clustering algorithms (e.g., DBSCAN, Gaussian Mixture Models) to improve shopper segmentation.
• Class Imbalance: Downsample non-shoppers or apply techniques like SMOTE to balance the dataset.
• Feature Engineering: Create interaction terms or ratios (e.g., ProductRelated_Duration per page) to enhance clustering.
• Dimensionality Reduction: Apply PCA or t-SNE to reduce feature space and improve cluster separation.