Avoid These Common NumPy Pitfalls When Doing Machine Learning

Trap 1: Chaotic Data Structures

Both array and matrix can represent multidimensional data, but using array to select the first column may return a 1‑D array (shape (3,)) instead of a 2‑D column vector, causing downstream errors unless reshaped.

In contrast, matrix preserves two‑dimensional results, though it still has quirks that will be discussed later.

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 returns only the first column as a row vector (1×3) instead of the expected 3×2 matrix, requiring cumbersome work‑arounds that are trivial in MATLAB/Octave ( X(Y==1,:)).

Trap 3: Confusing Numerical‑Operation Syntax

Machine‑learning code often needs both dot products and cross products. NumPy’s * operator performs matrix multiplication for matrix objects and element‑wise multiplication for array objects, forcing the user to switch between types or use dot for true dot products, unlike the concise syntax of MATLAB/Octave.

Therefore, we must convert each column to an array before performing element‑wise multiplication with y, then reshape the result back to a 5×3 matrix. After successfully computing the dot product of x and y, we can finally multiply by theta.

In array objects the * operator denotes element‑wise multiplication, while in matrix it denotes matrix multiplication; to perform a true dot product in an array you must use the dot method, which adds cognitive load compared to MATLAB/Octave’s simple x .* y * theta syntax.

Trap 4: Unnatural Syntax

Appending a column of ones to a 5×2 matrix to create a 5×3 matrix requires a verbose expression with many parentheses in NumPy, whereas MATLAB/Octave can achieve the same with a simple [ones(5,1) x], highlighting the syntactic overhead in Python.

Conclusion

Python, especially NumPy, is popular for machine learning and data analysis, but its design choices—such as lack of custom operators and the way it handles vector shapes—make it less ergonomic than domain‑specific languages like MATLAB/Octave, which is why many classic courses still use the latter.