I would imagine the problem to be very similar in spirit to developing a Huffman code given symbol probabilities. Given the distribution of file access probabilities, the number of scans+clicks in the folder hierarchy is captured by the encoding string length/cost. (But I’m being a bit loose)