Federated learning enabled hotel customer classification towards imbalanced data

Hotel customer classification is the basis of customer profiling, which can significantly benefit a hotel by providing more appropriate services for targeted customers. However, imbalanced data distribution from individual hotels cannot support a reliable classification result, and sharing personal information among hotels is not allowed. In this paper, we propose to achieve a privacy-preserved hotel customer classification model via federated learning. A significant challenge is that hotels with different star ratings or distributed in different city regions usually serve specific customer groups, resulting in imbalanced data that degrade classification accuracy. We introduce an attention mechanism and design a client selection strategy to balance global and local performance upon imbalanced data. Due to privacy issues, we evaluate our solution's communication cost and accuracy on public imbalanced datasets and demonstrate the real-world customer classification results. Extensive experiments show that our solution performs better than the COTA method.