wolffd@0: function [lhs, rhs, dim, factors, invalid] = get_svmlight_inequalities_from_ranking(r, X, delta_fun,delta_params)
wolffd@0: % [lhs, rhs] = get_svmlight_inequalities_from_ranking(r, features)
wolffd@0: % 
wolffd@0: % delta_fun.fun function pointer returning distance / delta(a, b) 
wolffd@0: %          for feature vectors a,b, default: delta(a, b) = (a - b).^2;
wolffd@0: % delta_fun.args extra arguments
wolffd@0: %
wolffd@0: % prepares ranking data to be used with the svmlight implementation
wolffd@0: % of Schultz & Joachims 2003
wolffd@0: 
wolffd@0: % ---
wolffd@0: % CAVE: this file still allows arbitrary rankings
wolffd@0: %  this could be good but also introducing 
wolffd@0: %  confusion in the future.  Especially the 
wolffd@0: %  weightings are only defined per query 
wolffd@0: % ---
wolffd@0: 
wolffd@0: % ---
wolffd@0: % variable function to get vector deltas
wolffd@0: % ---
wolffd@0: if nargin < 3
wolffd@0:     delta_fun = @get_delta;
wolffd@0:     delta_params = {[]};
wolffd@0: end
wolffd@0: 
wolffd@0: % ---
wolffd@0: % fix right hand side to one
wolffd@0: % (after Schultz & Joachims 2003)
wolffd@0: % ---
wolffd@0: fix_rhs = 1;
wolffd@0: fix_weight = 1;
wolffd@0: 
wolffd@0: % ---
wolffd@0: % NOTE: this preallocation is not complete
wolffd@0: % ---
wolffd@0: lhs = cell(0,2);
wolffd@0: factors = [];
wolffd@0: invalid = [];
wolffd@0: for i = 1:size(r,1)
wolffd@0:     
wolffd@0:     % feature index
wolffd@0:     a = i;
wolffd@0:     
wolffd@0:     % check if ranking is valid
wolffd@0:     if ~isempty(r{i,1}) && ~isempty(r{i,2})&& ...
wolffd@0:         isempty(intersect(r{i,1}, r{i,2}));
wolffd@0:     
wolffd@0:         % ---
wolffd@0:         % NOTE / TODO: the follwing is intended for compability
wolffd@0:         %  both sides of the ranking may have more than one entry.
wolffd@0:         %  for the MTT database, the ranking may be correct, but the 
wolffd@0:         %  inequalities build from non-singular rankings are not
wolffd@0:         %  based on the actual data
wolffd@0:         % ---
wolffd@0:         for j = 1:numel(r{i,1})
wolffd@0:             b = r{i,1}(j);
wolffd@0:             
wolffd@0:             for k = 1:numel(r{i,2})
wolffd@0:                 c = r{i,2}(k);
wolffd@0: 
wolffd@0:                 % ---
wolffd@0:                 % get vector deltas
wolffd@0:                 % ---
wolffd@0:                 [dab] = delta_fun(X(:,a), X(:,b), delta_params{:});
wolffd@0:                 [dac] = delta_fun(X(:,a), X(:,c), delta_params{:});
wolffd@0:                 
wolffd@0:                 % get the delta difference vector
wolffd@0:                 ddiff = dac - dab;
wolffd@0:                 if i == 1
wolffd@0:                     dim = numel(ddiff);
wolffd@0:                 end
wolffd@0:                 
wolffd@0:                 % ---
wolffd@0:                 % save the non-empty differences row by row 
wolffd@0:                 % NOTE: it is not clear  whether the indexing for 
wolffd@0:                 % \omega starts a 0 or 1.
wolffd@0:                 % ----
wolffd@0:                 xgzero = find( ddiff ~= 0);
wolffd@0:                 
wolffd@0:                 if ~isempty(xgzero)
wolffd@0:                     lhs = cat(1,lhs,{xgzero, ddiff(xgzero)});
wolffd@0:                 else
wolffd@0:                     invalid = [invalid i];
wolffd@0:                 end
wolffd@0:                 
wolffd@0:                 % save factors
wolffd@0:                 if (nargout > 2)
wolffd@0:                     if (size(r,2) > 2)
wolffd@0:                         factors = [factors, r{i,3}]; % / max_weight
wolffd@0:                     else
wolffd@0:                         factors = [factors, fix_weight];
wolffd@0:                     end
wolffd@0:                 end
wolffd@0:                 
wolffd@0:             end
wolffd@0:         end
wolffd@0:     end
wolffd@0: end
wolffd@0: 
wolffd@0: % ---
wolffd@0: % determine right hand side
wolffd@0: % ---
wolffd@0: rhs = ones(size(lhs,1), 1) .* fix_rhs;
wolffd@0: cprint(2, 'SVMLight data: %d invalid rankings excluded from training set', numel(invalid));
wolffd@0: end
wolffd@0: 
wolffd@0: 
wolffd@0: function out = get_delta(a, b, A)
wolffd@0: % returns the pointwise (transformed) feature vector differences
wolffd@0: 
wolffd@0: if nargin == 2 || isempty(A)
wolffd@0:     
wolffd@0:     % ---
wolffd@0:     % return squared factors as in euclidean distance
wolffd@0:     % ---
wolffd@0:     out = (a - b).^2;
wolffd@0:     
wolffd@0: else
wolffd@0:     
wolffd@0:     % transform vectors before accessing difference
wolffd@0:     out = (A' * a - A' * b).^2;
wolffd@0: end
wolffd@0: 
wolffd@0: end
wolffd@0: 
wolffd@0: