This is one of the Udacity Data Scientist Nanodegree Project. This project aims to use unsupervised learning techniques to identify segments of the population from the core customer base for a mail-order sales company in Germany. Therefore, these segments can then be used to direct marketing campaigns towards audiences with the highest expected rate of returns.

The techniques I used in this project include:

1. Data cleaning
2. Encoding and processing mixed-type feature
3. Feature Scaling and Dimensionality Reduction
4. Clustering
5. Performance improvement with OpenBLAS

You can find the full analysis in my GitHub repo.

Data

The data files associated with this project (not included in this repository):

• Udacity_AZDIAS_Subset.csv: Demographics data for the general population of Germany; 891,211 persons (rows) x 85 features (columns).
• Udacity_CUSTOMERS_Subset.csv: Demographics data for customers of a mail-order company; 191,652 persons (rows) x 85 features (columns).
• Data_Dictionary.md: Detailed information file about the features in the provided datasets.
• AZDIAS_Feature_Summary.csv: Summary of feature attributes for demographics data; 85 features (rows) x 4 columns

Analysis Structure

1. Data exploration and data cleaning (85% of the analysis)
2. Feature Engineering (One Hot, Scaling, and PCA)
3. Clustering with k-means

Conclusion

I use the elbow method to find that 6 is the optimal number for clustering, which means the model segments customers and the general population into 6¹ groups. We can find the proportion of customers in cluster 2 is higher than the general population, which suggests people in cluster 2 are the target audience. We also find that, in cluster 3, the customer is underrepresented, which means people in that group are outside of the target demographics.

Comparing 2 segments, we can find there are some key differences. For example:

• Distance from building to point of sale: People in cluster 3 are closer to Pos then people in cluster 2.
• Wealth / Life Stage Typology: More people in cluster 2 are upper class
• Type of Building: Most builds in cluster 2 are residential build.
• Social status: Most people in cluster 2 are top earners and most people in cluster 3 are house owners.

1

conda create -y -n p38openblas -c conda-forge "python=3.8" scipy "blas=*=*openblas"

Licensing and Acknowledgements

Udacity Data Scientist provided the starting code for this project. Udacity partners at Bertelsmann Arvato Analytics provided the data.