I think you may be going into precog territory while looking for a needle in a haystack. Some of the common patterns seem fairly well documented: disaffected male, mental issues, history of abuse, affiliation with radical islam. But most experts seem to believe finding a key indicator before hand is really impossible since you will also label so many people who will actually resist the urge to violence. The Orlando killer popped on the FBIs radar twice and they couldn’t make the call with all the expertise and experience the organization has. Machine learning a data model to predict violence will have so many false positives as to be either useless or used as a totalitarian hammer. [1, 2, 3, 4]