ML practice projects

A repository containing some (interesting) machine learning projects done for practice and kept for reference. Each project has it's own README providing more details.

Overview

• churn_featuretools_pipeline: Contrary to the other projects here, this is actually a "real life" use case from my job. And it shows the data load and feature preprocessing pipeline only. The interesting part is the experimentation with and then application of the featuretools library for feature generation.

• customer_lifetime_value_modelling: A project predicting (residual) customer lifetime value (clv) in non-contracutal business. Methodology is based on paper by Dr. Peter Fader of Wharton, the underlying math is handled by the lifetimes package. The project highlights some problems of this approach.

• customer_segmentation_RFM_method: Performing customer segmentation with the RFM method (Recency, Frequency, Monetary Value). Two approaches have been tested, one with k-means clustering and one with the 'classic' quantile method. The results have been compared, but without deeper analysis - this project is focused on code implementation.

• customer_segmentation_unsupervised_kmeans: Finding customer segments with PCA and k-means. Note: Everything in this project has been applied on a much more complex dataset in the Starbucks' segmentation challenge project. In the resources folder are several notebooks on different unsupervised algorithms from Udacity's MLE nanodegree course.

  • market_basket_analysis: A repository containing code for implemenation of association rules mining (ARM, also called frequent itemset mining) with the apriori algorithm. There is an implementation from scratch and a real use case on a b2b retail dataset using the MLextnd package. In a third notebook the rules are further analyzed with graph analytics with help of the networkX library.

• nlp_classification_webapp_disaster_response: Pt. 1: NLP Classification with Multi-Label Target Pt. 2: Creating a Web-App for message classification. The classification happens in a neat end-to-end sklearn pipeline. The resources folder contains notebooks on bow / tfidf and on nlp pipelining from Udacity's DS nanondegree course.

• price_prediction_ames_housing: Supervised regression modelling to predict house prices. Comparision of ElasticNet and XGBoost. The focus of this project is on finding the right feature preprocessing and most of all on inspecting feature importances (among others with eli5 library).

• recommender_systems_IBM-Watson: More like a showcase of different approaches for creating a recommender system. (rank-based, content-based (with NLP part), user-based, Funk-SVD). Extensive resources section from Udacity's DS nanodegree course, also containing a complete recommender class.

• survival_analysis_lifetimes: Exploration of the functionality of Cameron Davidson's lifelines package. Goal is to construct survival functions / survival curves for customers (as a whole, in cohorts and individual). This is an alternative approach to churn analysis, based on the durations of customer relationship.

• uplift_modelling_imbalanced_starbucks: uplift modelling project dealing with very imbalanced target classes. The main point was to experiment with the imblearn package and to create a pipeline that oversamples the minority target class with SMOTENC.

• web_FastAPI_deployment: Mini-Project to introduce myself to the FastAPI library. We train a super simple Random Forest Classifier and build a REST API that exposes the predictions for new observations based on that model.