This is the summary of a behavioral targeting article by Diana L. Huerta – Muñoz, Roger Z Rios-Mercado, and Ruben Ruiz. A beverage distribution company wants to partition its customers into those sharing marketing and geographical attributes. In addition, the segments must be compact enough as desired. This paper addresses this customer segmentation problem. The original pdf file can be accessed here: An Iterated Greedy Heuristic for a Market Segmentation Problem with Multiple Attributes.
Market segmentation is dividing a market into smaller segments of customers sharing similar service or product needs. This strategy is done by a company to know more about its customers for customer satisfaction and thus, profit gain.
In solving problems involving market segmentation, one of the most popular methods is clustering algorithms, which can be either heirarchical or partitional. Partitional methods are effective for large data, because then construction of tree diagrams (or dendograms) in hierarchical methods become difficult to compute. Here a single partition is defined either globally or locally. Various states are created and run to determine which is the best configuration and therefore the best output clustering.
In this study, a beverage distribution business aims to apply market segmentation, wherein each segment consists of customers having the same average purchase volume, type of store and contract. Segments also need to be as compact as possible. The results of this study will allow the company to try out effective product marketing strategies. A model was created which generated feasible solutions and Variable Neighborhood Search (VNS) was used to further improve the quality of the solutions.
Heuristics are trial-and-error methods of problem solving because an algorithmic approach is not practical. The heuristics in this study are called IGACS or Iterated Greedy Algorithm for Customer Segmentation. Using this approach, a sequence of solutions is generated by repeating over greedy constructive heuristics. It has two main phases, destruction and reconstruction.
Using p-means algorithm, a solution is obtained. In order to select the best solution, destruction involves removing some elements in this solution. Then, it is reconstructed by applying another greedy constructive heuristic. Partitions are improved in this study by adding a local search procedure.
For the initial partition, the method used in this study is called MPM-Modified p-means.
Computational Results and Statistical Analysis
Six different vectors are tested in this study using IGACS. The vectors are named: ‘equal weights,’ ‘commercial weights,’ ‘dispersion,’ ‘purchase volume,’ ‘type of contract,’ and ‘type of store,’with each name corresponding to which property is given preference. ‘Equal weights’ gives equal preference to all properties, while ‘commercial weights’ refer to weights set by the company.
Applying the variable neighborhood search procedure and the MPM algorithm in this study results in more dissimilar partitions. In addition, IGACS has improved the dissimilarity even more.