Find elements meeting any of a number of criteria

Question

I am trying to find the indices of elements in a vector that correspond to another vector, preferably without using loops. For example, my input might be:

DJ         


        

Answer 1

For small datasets, it seems my original approach was faster than both the ismember solution proposed by Divakar and the intersect solution proposed by qmeeeeeee, but all three get beaten by Luis Mendo's solution using good old bsxfun. See below code, which times each approach:

function somescript()

IsmemberTime = timeit(@membersol)
IntersectTime = timeit(@intersectsol)
FindTime = timeit(@naivesol)
BsxTime = timeit(@bsxfunsol)

    function membersol()
        rng(1)
        set = randi(30,[1000 15]);             % generate 1000 vectors of length 15, containing random integers
        for i=1:1000
            [~,out] = ismember(set(i,1:5),set(i,6:end));      % first 5 random integers are the values to be found in the remainder of the vector
        end

    end

    function intersectsol()
        rng(1)
        set = randi(30,[1000 15]);
        for i=1:1000
            [~,~,Output] = intersect(set(i,1:5),set(i,6:end));
        end
    end

    function naivesol()
        rng(1)
        set = randi(30,[1000 15]);
        for i=1:1000
            Output = find(ismember(set(i,6:end),set(i,1:5)));
        end
    end

    function bsxfunsol()
        rng(1)
        set = randi(30,[1000 15]);
        for i=1:1000
            [~, Output] = max(bsxfun(@eq, set(i,1:5).', set(i,6:end)), [], 1);
        end
    end
end

Which on my machine (running R2014b) returns the following timings:

IsmemberTime =

    0.1101


IntersectTime =

    0.2008


FindTime =

    0.0698


BsxTime =

    0.0218

This suggests that, for small data sets at least, using find and ismember on the inverted order of vectors is actually faster than using ismember alone. Since there is also some fixed overhead for all methods from the generation of the datasets set that are used to test with, the difference seems to be pretty big. More thorough tests can ben found in the comments below.