Predicting customers' health supplement purchase propensities for precision marketing at MeCome pharmacy

Project Details


Fall 2019


William Feng, Silvia Yang, Vivian Lee, Thompson Lin





Surviving in this competitive pharmaceutical market is not as simple as one would think it to be. Faced with a range of strong competitors, MeCome’s main business problem is to find a solution for its low customer loyalty. Therefore, our group proposes to use naive, XGBoost, and feature engineering as our primary data mining tools for precision marketing at MeCome. After deciding our tools, we moved on to data partition, data preprocessing, and RFM scores for better exploration. Furthermore, to simplify our search on a pharmacy that provides a variety of products, we have narrowed our research by building a predictive model for a sub-categorical product “2-06-01” (Natural enzyme products). The outcome variable is binary, 1 represents customers who will buy health supplements in the following month, while 0 means will not buy for the next month.

Through rigorous cleaning of the data, we then tested the naive, XGBoost, and feature engineering. After our explorations with different data mining strategies, we discovered that for data analysis, feature engineering performed poorly and naive has the best prediction performance. Additionally, the best tool to use to target potential customers that have high purchase propensity but does not shop in MeCome is by using machine learning through different features.

In terms of identifying customers with high health supplement purchase propensities, our recommendation to MeCome is to focus on deployment and future implementations. For deployment, we advise the company to send text messages to the high purchase propensity customers identified in our analysis. Additionally, for future implementation, a one-time analysis should be conducted before each promotion by the frequency of 2-3 months.

Application Area: