Efficient predictive classification model using CACP and GRASP

The volume of historical purchasing data has become huge, and it includes many kinds of data attributes. Specifically, categorical data, such as product codes, are difficult to handle. If the product is purchased repeatedly, we can aggregate the data and use the product data as a numerical attribute. However, if the item was purchased only once, we can get only very basic information, such as whether it was purchased or not. To use the information more effectively, we can use a subset of these purchased items as a purchasing pattern within the set of items. Some classification predictive models that use these patterns were proposed, including the classification by aggregating contrast patterns (CACP). However, the model sometimes produces too many specific patterns. This is not a problem for predictions, but interpreting the model can become too complicated to implement efficiently. In this paper, we propose a method to decrease the number of patterns in the classification model for CACP. The proposed method uses the meta-heuristics algorithm known as greedy randomized adaptive search procedure (GRASP). A computational experiment shows that we can remove extra patterns and construct the model, while maintaining its performance level.