Project Summary

Consider a system for predicting default risk used by Buy-now-pay-later (BNPL) lenders. An accurate predictor reduces the lenders’ risk and safeguards customer interests. Developing this predictor requires a large pool of customer profiles and their default records. However, sharing such data across corporations is impossible due to privacy agreements and business interests. This proposal aims to address this by building a distributed machine learning solution, training a predictor while keeping customer data local to the source institute ensuring privacy.

Now, think of a scenario where a child clicks on an advertisement and unknowingly ends up applying for a BNPL credit unknowingly, or when you return an item and don’t want the BNPL lender to retain your personal details. In these cases, among many others, financial institutes need the crucial ability to remove customers from their knowledge base and modify models, such as the above default risk predictor. Machine unlearning algorithms enable an already trained model to forget a selective subset of training samples. While existing methods lack research in applying unlearning to distributed learning, vital for this application, our project will build a distributed learning system with unlearning capabilities for default risk prediction.

Results

Partner Organisation