Abstract
Traditional Customer Relationship Management (CRM) focuses on customer acquisition and retention. However, identifying high-risk, potential-default customers is equally critical for long-term business sustainability, particularly in the financial sector. This paper proposes an intelligent CRM system utilizing a Support Vector Machine (SVM) ensemble. By integrating a "Choose From Subspace" (CFS) selection strategy with Bayesian-based aggregation, the proposed framework achieves a classification accuracy of 87.06%, significantly outperforming standard neural networks and logistic regression models.
Introduction
In industries such as banking and insurance, the ability to identify "delinquent" customers is a vital component of risk management. While most CRM research focuses on marketing and sales optimization, this study addresses the technical challenges of risk identification. Single "unitary" data mining models, such as individual Neural Networks, often suffer from inherent biases or architectural limitations. To overcome these, the authors employ ensemble learning—combining multiple diverse models to create a "wisdom of the crowd" effect that is more robust and accurate than any individual component.
Methodology
The proposed intelligent CRM system follows a four-step framework:
- Data Preparation: Standardizing and preprocessing raw customer data for consistency.
- Single Classifier Creation: To ensure ensemble effectiveness, individual models must be diverse. Diverse SVMs were generated by varying training subsets, kernel functions (e.g., Polynomial vs. Gaussian), and hyperparameters like the penalty parameter $C$ and kernel width $\sigma^2$.
- Ensemble Member Selection: The authors evaluated three strategies for selecting the final ensemble members:
- Principal Component Analysis (PCA): Selection based on error matrix eigenvalues.
- Choose the Best (CTB): Selecting models with the highest individual accuracy.
- Choose From Subspace (CFS): Selecting the best-performing model from each specific parameter space to ensure maximum diversity.
- Ensemble Construction: Two aggregation methods were compared: Majority Voting (simple democracy) and Bayesian-based ensemble (probabilistic weighting).
Experimental Setup
The study utilized a real-world credit dataset from a UK financial service company, consisting of 1,871 samples (800 for training and 1,071 for testing). The proposed SVM ensemble was benchmarked against Logistic Regression (LogR) and Artificial Neural Network (ANN) ensembles to validate its performance.
Results and Discussion
The experimental results demonstrate a clear hierarchy in performance across different model types and strategies. The authors found that diversity is mandatory; as per the Hansen and Salamon (1990) principle, ensemble members must be both accurate and diverse to outperform single models.
As shown in the data, the "Choose From Subspace" (CFS) strategy consistently outperformed PCA and CTB, suggesting that maintaining a representative group of different model types is superior to simply picking top-performing individuals that may share similar biases. Furthermore, the Bayesian-based aggregation consistently outperformed simple majority voting. The optimal configuration—SVM combined with CFS and Bayesian aggregation—reached an accuracy of 87.06%, a substantial improvement over the ~60% accuracy typical of traditional logistic regression approaches.
Conclusion
This research highlights the superiority of SVM ensembles for high-risk customer identification. While the proposed model offers significant accuracy gains, it does come with higher computational overhead compared to single models. Future work should address the transparency of specific predictive features and explore the model's performance on temporal data to identify shifting risk profiles over time.
Summary updated