DBSCAN-Based Customer Segmentation (Wholesale Dataset)

A machine learning project for segmenting wholesale customers using the DBSCAN density-based clustering algorithm, with a focus on identifying natural customer groups, detecting outliers, and analyzing spending behaviors across various product categories.

---

Overview

This project applies DBSCAN, a density-based clustering algorithm, to the popular Wholesale Customers Dataset.
The goal is to:

• Discover natural customer groups
• Detect abnormal high-spending outliers
• Analyze differences in purchase patterns
• Understand spending behavior across categories such as Milk, Grocery, Fresh, Frozen, and Detergents_Paper

This type of segmentation is widely used in real-world applications like customer profiling, market analysis, and retail intelligence.

---

Dataset

• Name: Wholesale Customers Dataset
• Source: Data set
• Records: 440 customers
• Features:
  • Channel (Hotel/Restaurant/Retail)
  • Region
  • Fresh
  • Milk
  • Grocery
  • Frozen
  • Detergents_Paper
  • Delicassen

These features represent annual spending levels for each customer.

---

Project Workflow

1) Exploratory Data Analysis

• Previewed the dataset and inspected NaN values
• Basic visualizations:
  • Scatterplots (Milk vs Grocery)
  • Histograms colored by Channel
  • PairPlot by Region
  • Correlation heatmap
  • Seaborn Clustermap for feature similarity

2) Preprocessing

Since the spending categories vary greatly (some up to 100,000 units), Standard Scaling was applied.

3) DBSCAN Modeling

Tested multiple epsilon values:

for eps in np.linspace(0.001, 3, 50):
    DBSCAN(eps=eps).fit_predict(scaled_data)

For each model, recorded:

• Number of clusters
• Number of outliers
• Percentage of outliers

This allowed identifying a stable and meaningful eps region.

Final eps chosen: eps = 2

4) Cluster Visualization

Created cluster-colored scatterplots:

• Grocery vs Milk
• Milk vs Detergents_Paper
• Other spending relationships

5) Cluster Profiling

A new column Labels stores DBSCAN cluster assignments.

Then compared cluster means:

cat_means = df.groupby("Labels").mean()

Outlier cluster (-1) represented extremely high spenders.
Clusters 0 and 1 represent medium and low spenders.

6) Normalized Cluster Heatmap

Used MinMaxScaler to normalize group means and produce a 0–1 heatmap.

This clearly reveals:

• Outlier cluster has extremely high spending in all categories
• Cluster 0: Average spending, moderate levels of Detergents_Paper
• Cluster 1: Low-volume grocery/milk customers

---

Libraries Used

• numpy
• pandas
• seaborn
• matplotlib
• scikit-learn (DBSCAN, StandardScaler, MinMaxScaler)

---

How to Run

Clone the repository:

github

Install dependencies:

pip install -r requirements.txt

• requirements.txt → file

or directly:

pip install numpy pandas seaborn matplotlib scikit-learn

Run all cells to train and evaluate the model.

This project is implemented as a Python script.
(No Jupyter Notebook version yet)

---

Results Summary

The DBSCAN model revealed three meaningful behavioral groups:

Cluster -1 (Outliers)

• Extremely high spenders across all categories
• Represent unique customer behavior
• Most likely wholesale distributors or very large retailers

Cluster 0

• Medium-to-high grocery and detergent spending
• Likely restaurants or hotels

Cluster 1

• Low spending across all features
• Small shops or low-volume buyers

---

Conclusion

This project demonstrates the power of density-based clustering for customer segmentation:

• No need to predefine K
• Natural detection of non-linear clusters
• Automatic identification of extreme-value customers
• Clear interpretation via cluster means and heatmaps

DBSCAN provides robust segmentation for noisy, high-variance datasets like wholesale customer data.

---

Author ✍️

Author: Ali
Field: Data Science & Machine Learning Student
Email: ali.hz87980@gmail.com
GitHub: ali-119