This paper outlines the process of creating a collaborative filter for a recommender system. The intent is for that recommender system to be used by a convenience store chain to increase in-store sales to their loyalty program members. The focus of the study is on cleaning and preparing data for analysis, creation of six different recommender systems, and comparing them on the basis of Root Mean Squared Error. 

The process included the loading of client provided excel sheets into Python. These were then transformed using basic cleaning techniques and changed according to the specific needs of the client. After preparing the data, an implicit ratings system was created that formed the backbone of the recommender models that were created. 

The recommender models were then created in Python using functions, and built-in functionality from the Surprise package. Each of these models had their own strengths and weaknesses that are further explored in the study. After creation, each model was scored using Root Mean Squared Error as a way of measuring accuracy of predicted ratings.  

After comparison, it was found that both a Mean Approach Model and a Singular Value Decomposition Model performed almost equally well in predicting customer ratings. Given their similarity in performance, and the simplicity and lower cost of the Mean Approach Model, it was recommended that the Client use a Mean Approach Model in making personalized recommendations for their loyalty program