numpy-set-ops-cuba6112-skillfactory-2

Overview

NumPy provides vectorized set operations for 1D arrays and multidimensional subarrays. These tools allow for deduplication, membership testing, and finding differences/intersections between datasets.

When to Use

• Deduplicating rows in a large feature matrix.
• Filtering a dataset to exclude a list of forbidden values.
• Synchronizing two datasets by finding their intersection.
• Compressing data by storing unique values and their index mappings.

Decision Tree

1. Need to find non-duplicate elements?
   • Use np.unique.
2. Need to reconstruct the original array from unique values?
   • Set return_inverse=True in np.unique.
3. Checking if elements exist in another list?
   • Use np.isin(data, target_list).

Workflows

1. Finding Unique Rows in a Dataset

   • Create a 2D array.
   • Call np.unique(arr, axis=0).
   • Inspect the result to see the deduplicated records.

2. Reconstructing an Array from Sets

   • Call u, inv = np.unique(arr, return_inverse=True).
   • Store 'u' and 'inv' separately (useful for data compression).
   • Rebuild the original array using u[inv].

3. Filtering by Membership

   • Define a 'forbidden' set of values.
   • Generate a boolean mask using ~np.isin(data, forbidden).
   • Filter the data: clean_data = data[mask].

Non-Obvious Insights

• Flattening by Default: Set operations work on flattened 1D versions of input arrays unless an axis is explicitly specified.
• NaN Handling: Like sorting, unique treats NaN as a value and sorts it to the end of the unique output.
• Lexicographic Row Sort: When axis=0 is used in unique, the resulting unique rows are sorted lexicographically.

Evidence

• "Returns the sorted unique elements of an array." Source
• "isin(element, test_elements...)... broadcasting over element only." Source

Scripts

• scripts/numpy-set-ops_tool.py: Routines for unique row detection and inverse reconstruction.
• scripts/numpy-set-ops_tool.js: Simulated set intersection logic.

Dependencies

• numpy (Python)

References

• references/README.md