Unsupervised segmentation of customers of an e-commerce site

Overview

Olist is a Brazilian company that offers a sales solution for online marketplaces. The aim of this project is to:

• create an actionable customer segmentation that its e-commerce teams can use on a daily basis for their communication campaigns;
• analyse the stability of the segments over time, in order to determine the frequency at which the segmentation must be updated to remain relevant, (for a maintenance contract)

Motivation

This is project 5 for the Master in Data Science (in French, BAC+5) from OpenClassrooms.

The project demonstrates :

• clustering algorithms (Kmeans, Agglomerative Clustering, DBSCAN)
• evaluation of cluster consistency, shape and stability
• visualisation of clusters in 2 dimensions (t-SNE, PCA)
• evolution of clusters over time

Requirements

To run the notebooks, the dataset must be placed in a DATA_FOLDER ('data/raw'). Python libraries are listed in requirements.txt. Each notebook also includes a list of its own requirements, and a procedure for pip install of any missing libraries.

Data: Olist has provided an anonymized database at https://www.kaggle.com/olistbr/brazilian-ecommerce, (~130Mb) with information on order history, products purchased, satisfaction comments, and customer location since January 2017.

Python libraries : numpy, pandas, matplotlib, seaborn, scikit-learn, scipy, missingno, dython, squarify, yellowbricks, holoviews, bokeh

Files

Notes : Files are in French. As requested for the project, the jupyter notebooks have not been "cleaned up" : the focus is on setting up, tuning, visualising and evaluating unsupervised learning algorithms.

Custom functions created in this project for data preprocessing, statistical analysis and data visualisation are encapsulated within each notebook, to avoid importing and versioning custom libraries. Open https://nbviewer.org/ and paste notebook GitHub url if GitHub takes too long to render.

• P5_01_analyse.ipynb: Data cleaning, exploratory analysis, feature engineering

• P5_02_essais.ipynb: Trials of different modeling approaches and selection of final model

• P5_03_simulation.ipynb: Simulation to determine the necessary frequency of updating the segmentation model (maintenance contract)

• P5_04_support.pdf: Presentation and conclusion

Approach

Data Cleaning

• joining 8 tables: aggregation by unique customer, elimination of incomplete orders

Exploration of available dimensions & Feature Engineering

Creation / selection of indicators of customer behaviour :

• Customer and vendor geolocation : distance between customer and vendor
• Delivery delay: delivery date - forecast date
• Customer satisfaction : Average rating of customer reviews
• Product category simplification; customer favourite product category
• Customer number of commands, preferred payment method and number of payments

Segmentation (clustering)

• RFM Segmentation (Recency, Frequency, Money)
• KMeans segmentation - choice of number of clusters K:
  • Consistency (inertia/distortion, Calinski Harabasz, Davies_Bouldin)
  • Shape (silhouette score)
  • Stability (reproducibility on sub-samples, seasonality)
  • Fit time
• (Hierarchical) Agglomerative Clustering: Ward, Average, Complete Linkage, Dendrograms
• DBSCAN (Density-Based Spatial Clustering of Applications with Noise): Hyperparameter tuning (epsilon, min samples)

Segment visualisation

• Silhouette plots, Multi Dimensional Scaling (MDS) plot of intercluster distance
• Radar (spider) plots, line (snake) plots, bubble plots, box plots
• Feature importance: PCA visualisation, t-SNE plots coloured by segment / feature

Stability Simulation

• evaluation of stability (last year of data), by 1-month to 6-month timesteps
• ARI score (adjusted rand index) to measure cluster similarity
• Stability visualisation (Sankey diagrams)

Conclusions

• Most distinct segmentation based on 5 features : ['MonetaryValue', 'Frequency', 'review_score', 'mean_nb_payments', 'delivery_delay']: less important features add "noise".
• Best consistency, silhouette scores and stability for 7 customer segments, characterised by small / medium / large purchase, frequent / infrequent purchase, and level of satsifaction
• Satisfaction very sensitive to departure from forecast delivery time
• Segments are stable for 2 months; stability depends on season

Suggestions for Improvement

• Better understand why customers are dissatisfied: specific to certain vendors, product categories?
• Explore further the characteristics of each segment: favourite categories, favoured payment methods, number of payments
• Add categorical variables to segmentation: use Kprototype with one-hot encoded favorite categories
• conduct a NLP (Natural Language) analysis of customer review comments and titles
• Segmentation by Kmeans seems very unstable: Check segment stability for different start dates
• 80% of customer purchases concentrated in only 3 of 26 states (SP, PR, MG) - why?

Features (keywords)

• clustering algorithms (Kmeans, Agglomerative Clustering, DBSCAN)
• evaluation of cluster consistency(), shape (silhouette) and stability
• visualisation of clusters in 2 dimensions (t-SNE, PCA, MDS)
• visual comparison of cluster features (radar plots, line plots)
• evolution of clusters over time (Sankey diagram)

Skills acquired

• Set up the unsupervised learning model adapted to the business problem
• Transform relevant variables of an unsupervised learning model
• Adapt the hyperparameters of an unsupervised algorithm in order to improve it
• Evaluate the performance of an unsupervised learning model