Effectivizing Bankruptcy Administration With AI

Sammanfattning

This thesis investigates the application of Machine Learning and Data Mining techniques to assist liquidators in their task of processing company bankruptcies. Through the use of different artificial intelligence models along with a transaction-filtering module based on keyword recognition, the objective is to reduce manual labor for the liquidators. The dataset was processed and scaled in preparation for anomaly detection. Entry encoding was also done to ensure that the features of the data was represented correctly. Synthetic data samples helped to decide which scaling methods worked best for the system. The proficiency of the different anomaly detection models was evaluated using synthetic data and through validation by a domain expert. Flagged entries with the highest probability of being fraudulent were reviewed by a domain expert, who found half of them to be relevant. Expert feedback suggested that additional contextual filtering could further improve the accuracy of the system, increasing the number of correctly flagged entries. The results indicate that there is promise in the use of anomaly detection methods to help liquidators find potential fraudulent data. The transaction-filtering module was deemed to be successful in effectively finding relevant assets. These findings show that using the developed system will have a potential of effectivizing bankruptcy administration.