Déjà vu: Assessing similarity between service contracts for risk prediction

Abstract

Major IT service providers typically manage a large portfolio of contracts with a variety of customers. To ensure smooth delivery and continuous profitability, it is critical for the service providers to leverage the experiences and lessons learnt from the historical contracts and prevent similar issues from reoccurring in the future. In this context, we investigate how to predict potential risks for new contracts based on their similarities with existing ones. A critical challenge along this line is to effectively measure the similarity between the contracts. To this end, extending from the Mahalanobis distance metric learning framework, we develop a new approach to gauge contract similarity using expert assessment data collected prior to contract signing (so called 'contract fingerprints'). A key advantage of the proposed method is the ability to train model with not only continuous distance measures between contract pairs, but also the binary side information of dissimilar pairs. Finally, experimental results on real-world service contract data show that our proposed approach greatly outperforms existing benchmarks, and can provide more accurate contract risk assessment.