The present invention relates to an adaptive fraud detection system and method that leverages customer feedback and advanced machine learning techniques. The system comprises an upload interface for receiving customer-provided data on suspected fraudulent activities, an OCR module for extracting textual information from the data, a data analysis unit for generating fraud detection features, a feature repository for managing these features, a Graph Recurrent Neural Network (GraphRNN) model for predicting fraud patterns, a decision-making module for integrating component outputs, and an alert generation unit for issuing fraud alerts. The method involves receiving and securing customer data, extracting text using OCR, analyzing the text to generate fraud features, updating the feature repository, constructing graph representations of transactions, applying the GraphRNN model for fraud prediction, and generating alerts based on the predictions.

The presently disclosed subject matter relates to detection and mitigation of financial fraud, and in particular to implementation of systems for training machine learning models in such systems. It involves self-labelling of fraud risk in a transaction processing system.

Some aspects relate to technologies for probabilistic account linking, for instance, to perform fraud detection on online transaction platforms. In accordance with some configurations, linking strategies are defined for linking accounts based on account attributes. An average linking probability is generated for each linking strategy using account data for accounts on an online transaction platform, and the average linking probabilities are stored. To determine whether to link two accounts, linking strategies shared by the two accounts are identified, an account linking probability for the two accounts is generated using the average linking probabilities for the linking strategies shared by the two accounts, and the account linking probability is compared against a threshold. If the account linking probability satisfies the threshold, the accounts are linked and an action is taken based on the account linking.