Maintaining　accurate　and　fast　transient　stability　is　essential　for　safe　operation　of　the　power　system.　With　the　development　of　wide-area　measurement　system,　machine　learning-based　transient　stability　assessment　has　become　the　trend.　However,　in　realistic　application　of　the　power　system,　the　impacts　on　evaluation　rules　between　critical　samples　and　noncritical　samples　are　different.　Thus,　an　improved　cost-sensitive　coefficient　assignment　method　based　on　fault　severity　is　proposed.　First,　the　fault　severity　of　each　unstable　sample　is　calculated.　Then,　the　correction　coefficient　of　the　loss　function　of　the　unstable　sample　is　linearized　according　to　different　fault　severities.　The　closer　the　sample　is　to　the　critical　case,　the　higher　the　cost　coefficient　is.　Finally,　the　improved　cost-sensitive　method　is　combined　with　the　deep　learning　model　and　tested　in　the　IEEE-39　bus　system.　As　shown　in　the　results,　the　improved　cost-sensitive　method,　which　gives　different　correction　coefficients　to　samples　according　to　different　fault　severities,　has　better　performance.