Simon Funk vs. Matlabs SVDS - matlab

Question

I want to build an recommender system using Simon Funks' algorithm.
The idea is to first construct the model offline in Matlab to perform some evaluation on the results to what number of features (or dimensions) gives the best performance.
I have a matrix with users x items, where the rating '0' means that the user has not given a rating for that item.
So far I have tried to implement Simon Funks' algorithm in Matlab (see source-code) below, but it performs really (!) bad. By accident I found Matlab's SVDS function to fill the empty ratings and it actually performs better and faster then my Simon Funks' function
Could someone tell me what could be cause of this? Or tell me what stupid mistake I have made in Matlab function? :P
function ratings = simon_funk(original_ratings, dimensions) % To construct a complete rating matrix
% See: http://www.timelydevelopment.com/demos/NetflixPrize.aspx
% Variables
global max_features; max_features = dimensions;
global init; init = 0.1;
min_epochs = 2;
max_epochs = 16;
min_improvement = 0.001;
lrate = 0.01;
k = 0.015;
respondent_count = size(original_ratings, 1);
item_count = size(original_ratings, 2);
rating_count = 0;
data = [];
for r=1:respondent_count
for i=1:item_count
rating = original_ratings(r,i);
if(rating > 0)
data = [data; [r, i, rating, 0]];
rating_count = rating_count + 1;
end
end
end
% Now data contain all ratings in the form [user_id, item_id, rating, cache (default 0)]
ratings = zeros(respondent_count, item_count);
global item_features; item_features = zeros(max_features, item_count);
global respondent_features; respondent_features = zeros(max_features, respondent_count);
% Init
item_features(:,:) = init;
respondent_features(:,:) = init;
% CalcFeatures
rmse_last = 2.0;
rmse = 2.0;
for f=1:max_features
for e=1:max_epochs
sq = 0;
rmse_last = rmse;
for r=1:rating_count
% for i=1:item_count
respondent = data(r,1);
item = data(r,2);
rating = data(r,3);
% Predict rating
p = simon_funk_predict_rating(data(r,:), f, 1);
err = (1.0 * rating - p);
sq = sq + err*err;
rf = respondent_features(f,respondent);
mf = item_features(f, item);
% Cross-train the features
respondent_features(f, respondent) = respondent_features(f,respondent) + (lrate * (err * mf - k * rf));
item_features(f, item) = item_features(f, item) + (lrate * (err * rf - k * mf));
% end
end
rmse = sqrt(sq/rating_count);
%if (e >= min_epochs && rmse > (rmse_last - min_improvement))
if ((e >= min_epochs) && ((rmse_last - rmse) < min_improvement))
break;
end
end
% Caching
for r=1:rating_count
data(r, 4) = simon_funk_predict_rating(data(r,:), f, 0);
end
end
% return new ratings set
for r=1:respondent_count
for i=1:item_count
sum = 1;
for f=1:max_features
sum = sum + item_features(f,i) * respondent_features(f,r);
if(sum > 10)
% sum = 10;
end
if(sum < 1)
% sum = 1;
end
end
ratings(r,i) = sum;
end
end
And the function to predict an individual rating:
function sum = simon_funk_predict_rating(rating, f, bTrailing)
global item_features;
global respondent_features;
global max_features;
global init;
respondent = rating(1,1);
item = rating(1,2);
cache_value = rating(1,4);
sum = 1;
if(cache_value > 0)
sum = cache_value;
end
sum = sum + (item_features(f,item) * respondent_features(f,respondent));
if (sum > 10)
%sum = 10;
end
if (sum < 1)
%sum = 1;
end
if (bTrailing == 1)
sum = sum + (max_features - f - 1) * (init * init);
if (sum > 10)
sum = 10;
end
if (sum < 1)
sum = 1;
end
end
Thank you very much!