wolffd@0: function [sim, dissim, confidence] = sim_from_comparison_naive(comparison, comparison_ids, symmetrical)
wolffd@0: %
wolffd@0: %  [sim, dissim, confidence] = sim_from_comparison_naive(comparison)
wolffd@0: %
wolffd@0: % derives symmetric, absolute similarity measurements
wolffd@0: % from relative magnatagatune comparisons
wolffd@0: % naive implementation for first tests of the ITML algorithm
wolffd@0: %
wolffd@0: 
wolffd@0: % reindex comparison for more simple evaluation
wolffd@0: % makro_prepare_comparison
wolffd@0: 
wolffd@0: % ---
wolffd@0: % analyse the number of comparisons for each pair of songs
wolffd@0: % ---
wolffd@0: [num_compares] = get_comparison_stats(comparison, comparison_ids);
wolffd@0: 
wolffd@0: % ---
wolffd@0: % in comparison, the outlying piece is highlighted. 
wolffd@0: % thus, we naively consider that
wolffd@0: % a. both of the remaining pieces are more similar to each other.
wolffd@0: % b. the outlier is dissimilar to both of the other pieces
wolffd@0: % ---
wolffd@0: [outsort, outidx] = sort(comparison(:,4:6),2,'ascend');
wolffd@0: 
wolffd@0: % ---
wolffd@0: % similarity of the two non-outliers a, b 
wolffd@0: % they are similar if both of them have scores way smaller 
wolffd@0: % than the outlier c:
wolffd@0: % score (a,b) = 1 - (max(a,b)/c)
wolffd@0: %
wolffd@0: % dissimilarity: clip b is considered more different to clip c than
wolffd@0: % a, as clip a seems to share some properties with both songs
wolffd@0: % dissim(b,c) = 0.5 + b/(2c)
wolffd@0: % ---
wolffd@0: 
wolffd@0: sim = sparse(numel(comparison_ids),numel(comparison_ids));
wolffd@0: dissim = sparse(numel(comparison_ids),numel(comparison_ids));
wolffd@0: for i = 1:size(comparison,1)
wolffd@0:     
wolffd@0:     % get the outlier votes
wolffd@0:     simpair = comparison(i,outidx(i,1:2));
wolffd@0:     c = comparison(i,outidx(i,3));
wolffd@0:     
wolffd@0:     % we want a triangular similarity matrix
wolffd@0:     [simpair, simidx] = sort(simpair);
wolffd@0:     outsort(i,1:2) = outsort(i,simidx);
wolffd@0:     
wolffd@0:     % ---
wolffd@0:     % save the distance between the second biggest vote and the max vote.
wolffd@0:     % NOTE: we bias the vote by dividing through the number of total
wolffd@0:     % comparisons for the particular pair of clips
wolffd@0:     % ---
wolffd@0:     sim(simpair(1), simpair(2)) = sim(simpair(1), simpair(2)) +  ...
wolffd@0:         (1 - outsort(i,2) / outsort(i,3)) * (1 / num_compares(simpair(1),simpair(2)));
wolffd@0:     
wolffd@0:     dissim(simpair(1:2), c) = 0.5 + (outsort(i,1:2) ./ (2 * outsort(i,3)));
wolffd@0: end
wolffd@0: 
wolffd@0: % ---
wolffd@0: % mirror to make matrix symmetrical 
wolffd@0: % ---
wolffd@0: if nargin == 3 && symmetrical
wolffd@0:     sim = sim + sim';
wolffd@0:     dissim = dissim + dissim';
wolffd@0: end
wolffd@0: 
wolffd@0: % ---
wolffd@0: % TODO: use number of votes and std or similar to 
wolffd@0: % rate the confidence for each similarity mesurement
wolffd@0: % ---
wolffd@0: confidence = [];
wolffd@0: