Why Numpy’s Array vs Matrix Can Trip Up Your Machine Learning Projects

Trap 1: Chaotic Data Structures

Both array and matrix can represent multi‑dimensional data, but selecting rows and columns often yields unexpected shapes. Using a[:, 0] on an array returns a one‑dimensional shape (3,) instead of a 3 × 1 column vector, requiring an explicit reshape. In contrast, matrix preserves two‑dimensional results, though it can still produce a transposed 1 × 2 vector when a 2 × 1 column vector is expected.

Trap 2: Insufficient Data‑Processing Capability and Low Language Efficiency

When filtering a 5 × 2 matrix X with a 5 × 1 boolean matrix Y, NumPy’s boolean indexing keeps only the first column and collapses the result into a 1 × 3 row vector, contrary to the expected 3 × 2 matrix. Achieving the correct shape requires multiple complex indexing steps, whereas MATLAB/Octave accomplishes the same task with a single expression X(Y==1, :).

Trap 3: Confusing Numerical Operation Syntax

NumPy overloads the * operator: it performs matrix multiplication for matrix objects but element‑wise multiplication for array objects. Consequently, a naïve x * y either raises an error or yields an unintended matrix product. The correct workflow is to convert operands to array for element‑wise multiplication, then use dot for true matrix products, which adds cognitive overhead compared to MATLAB’s concise x .* y * theta syntax.

Trap 4: Complex and Unnatural Syntax

Appending a column of ones to a 5 × 2 matrix requires a verbose expression with many parentheses in NumPy, whereas MATLAB accomplishes the same task with the simple and readable [ones(5,1) x] syntax.

Conclusion

Python has become popular for machine learning and data analysis, but compared with domain‑specific languages like MATLAB/Octave it feels cumbersome. Limitations such as the lack of custom operators and NumPy’s design choices (e.g., automatic conversion of column vectors to row vectors) make the code less elegant and intuitive, which explains why many classic courses still prefer MATLAB.