Frequency Encoding

Frequency encoding replaces each category with how common it is — one column, no target involved, and no leakage to worry about.

Task: write frequency_encode(values) returning the encoding for each row, rounded to 4 decimal places.

• Return one value per row, in the original order. Rows sharing a category share an encoding.
• Divide by the total number of rows, so the distinct encodings don't sum to 1 — but every row's encoding does sum to 1 across the dataset.
• Categories may be numbers or strings.
• If every row is the same category, every encoding is 1.0.

The bet this encoding makes is that rarity itself is informative, and surprisingly often it is. A rare browser version correlates with fraud; a rare product category behaves differently from a bestseller; a rare job title in a salary model genuinely marks an unusual row. None of that needs the target to discover.

Its obvious failure mode is collisions: two categories appearing exactly the same number of times get exactly the same encoding, and the model can no longer tell them apart. In the long tail of a high-cardinality column, where thousands of categories appear once each, that's nearly everything — which is why frequency encoding is usually one signal among several rather than a replacement for the category.