MarketingCampaign

A customer segmentation project using unsupervised and supervised learning from Udacity's Machine Learning Engineer nanodegree.

Background

This project was proposed by Arvato Financial services. Arvato Bertelsmann is a company dedicated to help its customers with several financial aspects of organizations. Such aspects include invoicing & accounting, debt collection, credit risk & fraud management, among others. Its solutions have been applied in industries ranging from e-commerce to telecommunications and banking.

Marketing is essential for any sales-related business; however, marketing campaigns are expensive and sometimes even risky. In this project we contribute to create an effective marketing campaign that uses machine learning techniques to estimate which individuals can be potential clients (according to some data). The final solution can help to reduce the cost of marketing while increasing the chance of having new customers.

Problem statement

From a business perspective, the problem statement of this project is to answer the following question: “How can our client (a mail-order company of organic products) acquire new clients more efficiently?”

Datasets and inputs

Arvato has provided four datasets with attributes and demographic information from existing clients, and from a sample of the general population in Germany.

• Udacity_AZDIAS_052018.cscv: Demographics data for the general population of Germany; 891,211 subjects with 366 features each.
• Udacity_CUSTOMERS_052018.csv: Demographics data from past customers of the mail-order company; 191,654 subjects with 369 features each.
• Udacity_MAILOUT_052018_TRAIN.csv: Train subset of targets individuals of a marketing campaign; 42,982 subjects with 367 features each.
• Udacity_MAILOUT_052018_TEST.csv: Test subset of targets individuals of a marketing campaign; 42,982 subjects with 367 features each.

Solution statement

The solution is straightforward: based on the features of a person’s features, we have to develop a model to predict whether or not such person has a higher chance to become a customer of the mail-order company. However, since the number of subjects and features is high and, since we have additional useful information (from the existing clients), we can first investigate which features are more useful to predict our goal. The first part of the solution is to use an unsupervised learning algorithm to match the features of the existing clients with those from the general population in Germany, the goal of this is to characterize the profile of potential customers. As a second step, we can proceed to train a classification model to predict whether or not a person has a high possibility of becoming a client. In the first case, a convenient algorithm to try is KMeans, because it is an efficient model and performs well with medium size datasets. If necessary, we can use a dimensionality reduction technique (such as PCA) to improve the performance of the model. In the second case, it can be useful to try with a simple linear model (such as logistic regression) and also try with a more complex algorithm such as a RandomForest classifier.

Notebooks descriptions

This project is divided in three notebooks called "Arvato Project Workbook pt 1-3":

• In Workbook 1, data exploration and data cleaning are performed on data coming from customers and from the general population.
• In Workbook 2 we perform a population segmentation using PCA as well as an unsupervised learning algorithm
• Finally, workbook 3, trains three differente supervised learning models.

A more detailed report for this project can be found in the following link