Can this silly accumarray for-loop be removed by a vectorization?

I have a matrix A where some of the values in column 1 are duplicates. Where a duplicate exists in column 1 (representing a timestamp) I want to replace the corresponding values in the other columns by a mean value of the duplicates.

Right now Im doing it by identifying the unique values and then loop through the columns one by one via a temporary variable since accumarray only accepts one column at a time:

[UA,~,idx] = unique(A(:,1));  % Finds duplicates in the first column of matrix A

for i = 28:-1:2
Temp = [UA,accumarray(idx,A(:,i),[],@mean)];  % Replaces duplicates with the mean of the duplicate values in column i
B(:,i) = Temp(:,2);  % Transfers the results and builds the final matrix B containing the same as A but with no duplicates
end

It is doing what I want but it is really slow, is there any simple way of rewriting it so I don’t have to deal with the for-loop?

Answer 1

[Sean de Wolski]

A = randi(32280,28);
[UA,~,idx] = unique(A(:,1));
[X,Y] = ndgrid(idx, 2:size(A,2));
newcolumns = accumarray([X(:), Y(:)], reshape(A(:,2:end),[],1), [], @mean);

The ndgrid needs to start in the second column and only extract the 2nd through end columns of A for vals.

[Peta]

Thank you that worked as it should. But sadly it turned out to be slower compared to the original for-loop method.

Answer 2

[Kelly Kearney]

I'm not sure if it will be faster, but you may try my aggregate function. It only calls accumarray once under the hood, though, using indices to make the application to multiple columns a little more efficient.

[~, B] = aggregate(A(:,1), A, @(x) mean(x,1));
B = cat(1, B{:});

[Peta]

Yes, actually that was substantially faster! Thank you.

Answer 3

[Walter Roberson]

colidx = repmat(1:size(A,2), size(A,1), 1);
newcolumns = accumarray([idx, colidx(:)], reshape(A(2:end,:),[],1), [], @mean);
output = [UA, newcolumns];

[Peta]

That’s not working for me, I’m getting ” Error using horzcat Dimensions of matrices being concatenated are not consistent.”. Because my A is 32280x28, which makes my colidx(:) size 903840x1 while idx is just 32280x1.

So I think the error is in [idx, colidx(:)]

[Walter Roberson]

Hmmm... Try

[X,Y] = ndgrid(idx, 1:size(A,2));
newcolumns = accumarray([X(:), Y(:)], reshape(A(2:end,:),[],1), [], @mean);

[Peta]

No, it’s still not quite right I think.

I’m getting an error saying that VAL must have one element for each row in SUBS. Which is because [X(:), Y(:)] which corresponds to SUBS has a slightly different dimension than reshape(A(2:end,:),[],1) which would be VAL.

A = randi(32280,28);
[UA,~,idx] = unique(A(:,1));
[X,Y] = ndgrid(idx, 1:size(A,2));
newcolumns = accumarray([X(:), Y(:)], reshape(A(2:end,:),[],1), [], @mean);

Error using accumarray
Second input VAL must be a vector with one element for each row in SUBS, or a scalar.