Learning Decisions Offline from Censored Observations with {\epsilon}-insensitive Operational Costs