Problem Description #

I wrote a MatLab function to calculate the cosine similarity of 60 vectors in a dataset. The dataset has 60 rows (corresponding to vector IDs) and 2 columns (x and y components of the vectors). Here's my code:

Cosine Similarity Function #

function [cosSim] = cosineSimilarity(data)
    [n_row n_col] = size(data);
    norm_r = sqrt(sum(abs(data).^2,2));
    for i = 1:n_row
        for j = i:n_row
            cosSim(i,j) = dot(data(i,:), data(j,:)) / (norm_r(i) * norm_r(j));
            cosSim(j,i) = cosSim(i,j);
        end
    end
end

Main Script #

cd(matlabroot)
cd('help/toolbox/stats/examples')
ds = dataset('XLSFile','TestCosSim.xlsx');
c = cosineSimilarity(ds);

When running the script, I get the error: Undefined operator '.^' for input arguments of type 'dataset', pointing to line 8 of the cosineSimilarity function and line 6 of the main script. What's causing this issue?

Answer #

Let's break this down clearly:

Root Cause #

The error happens because you're passing a dataset object to your cosineSimilarity function, not a standard numeric matrix.

Matlab's dataset type is designed to store labeled tabular data (with column names, row labels, etc.)—it's not a raw numeric array. Operations like .^ (element-wise exponentiation), sum(...,2), and dot() are built to work with numeric matrices, not dataset objects. When you try to run abs(data).^2 on a dataset, Matlab doesn't know how to handle that, hence the error.

Fixes #

You have two straightforward ways to resolve this:

Option 1: Convert the Dataset to a Numeric Matrix in the Main Script

Modify your main script to convert the dataset object to a numeric matrix before calling the function. Use ds{:,:} to extract all numeric columns as a matrix:

cd(matlabroot)
cd('help/toolbox/stats/examples')
ds = dataset('XLSFile','TestCosSim.xlsx');
% Convert dataset to numeric matrix
data_matrix = ds{:,:}; 
c = cosineSimilarity(data_matrix);

Option 2: Make the Function Handle Dataset Inputs Directly

Update your cosineSimilarity function to check for dataset inputs and convert them automatically. This makes the function more flexible:

function [cosSim] = cosineSimilarity(data)
    % Convert dataset to numeric matrix if needed
    if isa(data, 'dataset')
        data = data{:,:};
    end
    
    [n_row n_col] = size(data);
    norm_r = sqrt(sum(abs(data).^2,2));
    cosSim = zeros(n_row);  % Preallocate matrix for efficiency
    for i = 1:n_row
        for j = i:n_row
            cosSim(i,j) = dot(data(i,:), data(j,:)) / (norm_r(i) * norm_r(j));
            cosSim(j,i) = cosSim(i,j);
        end
    end
end

(I also added cosSim = zeros(n_row); to preallocate the result matrix—this is a good practice in Matlab to speed up loops with large datasets.)

Why This Works #

ds{:,:} extracts all the numeric data from the dataset into a standard Matlab numeric matrix. Once you're working with a numeric matrix, all the operations in your function (.^, sum, dot) will behave exactly as you expect.

内容的提问来源于stack exchange，提问作者Joana Leite