Profit-driven fusion framework based on bagging and boosting classifiers for potential purchaser prediction

Abstract

Accurately identifying potential purchasers (PPers) is pivotal for enhancing an enterprise's core competitiveness in a competitive market. Although existing research focused on individual classifiers for PPer prediction, there is a notable gap in the integration of the bagging and boosting algorithms, resulting in suboptimal performance. This study introduces a novel fusion framework for profit-oriented PPer prediction that combines the strengths of the bagging (specifically, random forest, RF) and boosting (utilizing categorical boosting, CatBoost) algorithms. CatBoost replaces the original base learner in RF, leveraging the advantages of both classifiers to reduce the variance and bias. To optimize the proposed RF-CatBoost-based fusion framework for profit maximization, we employ a grid search to fine-tune hyperparameters. This approach aligns with enterprises' profit-driven objectives. The experimental results, statistical tests, and Bayesian A/B tests collectively demonstrate that the proposed framework outperforms all comparative classifiers, yielding the highest profits. Furthermore, an interpretability analysis reveals the significant factors influencing the prediction results, providing valuable insights for decision makers in identifying PPers within customer groups.