TL;DR
This paper introduces a novel method for classifying bank customers by integrating both static demographic data and dynamic transaction behavior. By using a hybrid genetic algorithm to cluster dynamic data and then feeding these clusters along with static attributes into a neural network, the authors achieve a significant improvement in classification accuracy compared to methods relying solely on static data.
What Problem Does This Paper Address?
Customer relationship management (CRM) in banking heavily relies on understanding and segmenting customers. Traditional approaches often focus on static customer attributes like age, income, and gender. However, these static features alone may not fully capture a customer's evolving behavior and preferences. Dynamic data, such as transaction history, offers a richer, albeit more complex, view of customer behavior. The challenge lies in effectively combining these two distinct types of data to create more accurate and insightful customer classifications. Existing methods often struggle with the privacy and update frequency issues of demographic data, or they fail to adequately integrate dynamic behavioral patterns.
Methods
The proposed method consists of three main phases:
- Customer Data Preparation (Dynamic Data): The paper focuses on preprocessing dynamic transaction data. This involves weighting transactions by profit to create profit-weighted time series, which are then used alongside original time series to cluster buying behaviors. Descriptive statistical features (trend, average, variance, kurtosis, skewness) are extracted from these time series to reduce data dimensionality and avoid assumptions about time series properties.
- Customer Transaction Data Cluster Using Hybrid GA: A hybrid Genetic Algorithm (GA) is employed for clustering the dynamic data. This hybrid approach combines GA with a Simulated Annealing (SA) algorithm to prevent premature convergence and enhance effectiveness.
- Coding Method: Chromosomes represent cluster centers, composed of real numbers. Each customer's dynamic data is represented by 10 features, so a chromosome for K clusters has 10*K genes.
- Fitness Function: The fitness is inversely proportional to the total measure of distance among cluster centers, aiming to minimize intra-cluster distances.
- Genetic Operators: Standard roulette wheel selection and a one-point crossover are used. A novel mutation operator is introduced that adjusts genes based on their deviation from the maximum and minimum cluster center measures, ensuring centers stay within data ranges and protecting well-fitted chromosomes.
- Customer Classification Using Static and Dynamic Features: The clustered dynamic data (from the GA) is combined with static customer attributes. A Back Propagation (BP) neural network with one hidden layer is used for the final classification. The input layer receives both static attributes and the dynamic cluster labels, while the output layer represents the customer group number.
Datasets
The authors use a sample dataset from a bank branch in China.
- Dataset Size: 1932 customer records.
- Time Period: 2002/1/1 to 2003/12/13.
- Customer Groups: Pre-classified into high value, middle value, and low value groups.
- Static Attributes: 16 attributes were selected and transformed into numerical representations (e.g., age ranges, gender coded as 1 or 2).
- Dynamic Attributes: Derived from 23 kinds of financial products and services, with transaction records including date, code, amount, and profit rate.
Key Insights and Ideas
- Synergy of Static and Dynamic Data: The core insight is that combining static demographic profiles with dynamic behavioral patterns provides a much richer and more accurate understanding of customers than either data type alone.
- Hybrid GA for Robust Clustering: Using a hybrid GA with SA helps overcome the limitations of traditional clustering algorithms and standard GAs, leading to more effective and stable clustering of complex dynamic data.
- Descriptive Statistics for Time Series: Employing descriptive statistical features (trend, average, etc.) for time series data simplifies the handling of dynamic data, making it more manageable for clustering without requiring strict assumptions about time series properties.
- Neural Networks for Integrated Classification: BP neural networks are well-suited for integrating the outputs of the dynamic clustering with static attributes to perform the final, more nuanced customer segmentation.
Key Results
- The proposed integrated method significantly improved classification accuracy.
- Training Accuracy: Increased from 73.1% (static data only) to 85.3% (combined data), a 17% improvement.
- Verifying Accuracy: Increased from 69.4% (static data only) to 82.9% (combined data), a 19% improvement.
- The dynamic data cluster results were found to be a valuable input to the neural network, confirming the effectiveness of the integration.
Limitations and Future Work
- Limitations: The paper acknowledges that the dynamic data cluster results cannot be used as the sole input to the neural network, implying that static data still plays a crucial role. The specific details of the "privacy of and update speed issues of customers information" are mentioned as a motivation but not deeply explored as a limitation of the proposed method itself.
- Future Work: The authors suggest that for further study, more features can be included in the clustering process based on the specific application's characteristics. They also mention the possibility of incorporating "chaos feature" to further enhance classification rates.
Summary updated