A link prediction-based recommendation system using transactional data

Area of Science:

• Data Science
• Machine Learning
• Graph Theory

Background:

• Increasing data volume necessitates effective recommendation systems.
• Transaction datasets (e.g., credit card, e-commerce) offer valuable user-item interaction signals.
• Understanding user interests is key for relevant item recommendations.

Purpose of the Study:

• To propose a link prediction-based recommendation system for transaction datasets.
• To leverage graph representation learning and gradient boosting classifiers.
• To predict future user purchasing behavior, specifically merchant selection.

Main Methods:

• Constructing a user-item interaction network.
• Applying graph representation learning algorithms for node embeddings.
• Utilizing gradient boosting classifiers for link prediction.
• Evaluating performance against matrix factorization methods.

Main Results:

• The proposed system demonstrated superior performance in merchant prediction.
• Key metrics included receiver operating characteristic curves and area under the curve.
• Mean average precision also indicated the model's effectiveness.
• Transactional data proved powerful for generating relevant recommendations.

Conclusions:

• The proposed link prediction-based system excels with transaction data.
• Graph learning and gradient boosting offer a robust approach to recommendations.
• This method provides a powerful alternative for enhancing user experience in e-commerce and financial systems.