Identifying SME customers from click feedback on mobile banking apps: Supervised and semi-supervised approaches

Area of Science:

• Computer Science
• Data Science
• Banking Technology

Background:

• The banking industry is increasingly shifting towards mobile applications, necessitating advanced customer understanding.
• Customer segmentation is crucial for banks to gain insights into user lifestyles and behaviors for personalized services.

Purpose of the Study:

• To develop and evaluate a method for classifying mobile banking users into SME (Small and Medium-sized Enterprises) and Non-SME groups based on click behavior.
• To improve the identification of anonymous users and enable targeted service and product offerings.

Main Methods:

• Extracted hand-crafted features from click log data and evaluated using Extreme Gradient Boosting (XGBoost).
• Converted user click logs into image representations to capture temporal information, reducing the need for manual feature engineering.
• Trained Convolutional Neural Networks (CNNs), specifically ResNet-18, on these image representations.
• Evaluated a semi-supervised learning model using image data, combining labeled and unlabeled images.

Main Results:

• The CNN model (ResNet-18) achieved an average accuracy of 71.69%, outperforming XGBoost (61.70%).
• The semi-supervised learning model achieved 73.12% accuracy using only half the labeled images compared to the fully supervised CNN.
• Image-based representations proved effective for training with semi-supervised algorithms, yielding superior results with less labeled data.

Conclusions:

• Converting mobile banking click logs into images is an effective strategy for customer segmentation.
• Image-based representations enable successful application of semi-supervised learning, improving classification accuracy and reducing reliance on labeled data.
• This novel approach enhances the understanding of mobile banking user behavior and offers a new method for developing customer segmentation classifiers.