Combining Prediction Technologies to Help Moderate Discussions

I came across a 2015 blog post by Vi­talik Bu­terin that con­tains some ideas similar to Paul Chris­ti­ano’s re­cent Crowd­sourc­ing mod­er­a­tion with­out sac­ri­fic­ing qual­ity. The ba­sic idea in both is that it would be nice to have a panel of trusted mod­er­a­tors care­fully pore over ev­ery com­ment and de­cide on its qual­ity, but since that is too ex­pen­sive, we can in­stead use some tools to pre­dict mod­er­a­tor de­ci­sions, and have the trusted mod­er­a­tors look at only a small sub­set of com­ments in or­der to cal­ibrate the pre­dic­tion tools. In Paul’s pro­posal the pre­dic­tion tool is ma­chine learn­ing (mainly us­ing in­di­vi­d­ual votes as fea­tures), and in Vi­talik’s pro­posal it’s pre­dic­tion mar­kets where peo­ple bet on what the mod­er­a­tors would de­cide if they were to re­view each com­ment.

It seems worth think­ing about how to com­bine the two pro­pos­als to get the best of both wor­lds. One fairly ob­vi­ous idea is to let peo­ple both vote on com­ments as an ex­pres­sion of their own opinions, and also place bets about mod­er­a­tor de­ci­sions, and use ML to set baseline odds, which would re­duce how much the fo­rum would have to pay out to in­cen­tivize ac­cu­rate pre­dic­tion mar­kets. The hoped for out­come is that the ML al­gorithm would make cor­rect de­ci­sions most of the time, but peo­ple can bet against it when they see it mak­ing mis­takes, and mod­er­a­tors would re­view com­ments that have the great­est dis­agree­ments be­tween ML and peo­ple or be­tween differ­ent bet­tors in gen­eral. Another part of Vi­talik’s pro­posal is that each com­menter has to make an ini­tial bet that mod­er­a­tors would de­cide that their com­ment is good. The ar­ti­cle notes that such a bet can also be viewed as a re­fund­able de­posit. Such forced bets /​ re­fund­able de­posits would help solve a se­cu­rity prob­lem with Paul’s ML-based pro­posal.

Are there bet­ter ways to com­bine these pre­dic­tion tools to help with fo­rum mod­er­a­tion? Are there other pre­dic­tion tools that can be used in­stead or in ad­di­tion to these?