I don't know of any deeper interpretation of blending. I guess one could look to the 'wisdom of crowds' for anecdotal evidence. There might be connections with consensus and voting systems, but those are mostly discrete AFAIK. It is at least a pragmatic way of exploiting the different biases of the members in the ensemble. In our case, the Netflix predictions we made based on Wikipedia data scored worse on their own than ratings-only predictions, but they attracted some weight in the blend and made the overall score (marginally) better.