Frequency-encode a CSV categorical column
Paste CSV, pick a column, and turn every category into how often it occurs — as a raw count, a 0–1 share, a percent, or a log-count.
About this tool
Frequency Encoder replaces a categorical CSV column with how often each value occurs — the technique usually called frequency encoding or count encoding. A high-cardinality text column (product ID, merchant, ZIP, user agent) becomes one numeric feature instead of the hundreds or thousands of columns one-hot encoding would add.
Pick the column and an encoding: the raw count, the frequency share of
rows (0–1), a percent, or a log-count (ln(1 + count)) that compresses
very skewed distributions. Output can overwrite the column or land in a new
appended column. Rare values can be pooled into a single group so one-off
categories do not each get their own noisy level.
Worked example
CSV:
product_id,price
X,10
Y,20
X,30
Z,40
X,50
Y,60
With column product_id and Count encoding, each ID becomes the number of
rows it appears in — X = 3, Y = 2, Z = 1:
product_id,price
3,10
2,20
3,30
1,40
3,50
2,60
Switch the encoding to Frequency and the output to append, and the
original column is kept while a product_id_freq column holds each value's
share of the 6 rows (X = 3/6 = 0.5000, Y = 0.3333, Z = 0.1667):
product_id,price,product_id_freq
X,10,0.5000
Y,20,0.3333
X,30,0.5000
Z,40,0.1667
X,50,0.5000
Y,60,0.3333
Limits & edge cases
- Counts come from this one input only. There is no separate fit/transform split, so "unseen category" handling does not apply — every value in the data is counted as it is read.
- Frequency and percent divide by the number of counted rows. When blank cells are set to NaN or Zero those rows are excluded from both the counts and the denominator; when blanks are counted they form their own category.
- Decimal places apply to frequency, percent, and log-count. Raw counts are always whole numbers.
- Frequency encoding cannot separate two categories that occur equally often — they collapse onto the same number. That is inherent to the method, not a limit of this tool.
- Pooling below a minimum count gives every rare value the combined count of all pooled values, so they share one level. A value of 0 or 1 disables it.
- Values are trimmed before counting; with case-sensitivity off they are also lower-cased for grouping. The original cell text is untouched in append mode.
- One column per run. Columns are chosen by header name, or by 1-based number when the header checkbox is off.
FAQ
What is frequency (count) encoding and when should I use it?
Frequency encoding replaces each category with how often it occurs in the data. It is a good fit for high-cardinality columns where one-hot encoding would add too many columns, and it works well when rarity itself carries signal — for example when one-off merchants behave differently from frequent ones.
What is the difference between count, frequency, percent, and log-count?
They are the same quantity with different scaling. Count is the raw number
of rows. Frequency divides that by the number of counted rows, giving a 0–1
share. Percent is that share times 100. Log-count is ln(1 + count),
which pulls in a long tail so a value seen 5,000 times does not dwarf one seen
5 times.
Two categories occur the same number of times — is that a problem?
They will receive exactly the same encoded value, because frequency encoding only knows how often a value occurs, not which value it was. If that collision matters for your model, keep the original column by using append output, or combine this feature with another encoding.
What does pooling rare categories do?
Set a minimum count and every value occurring fewer times than that is treated as one combined group: the pooled total replaces each of their individual counts. This keeps a long tail of one-off values from becoming many distinct noisy levels. Values at or above the minimum keep their own count.
How are blank cells handled?
The Blank cells setting decides. Count as their own category (the default) treats blank as a real value and counts it like any other, which mirrors how common encoder libraries treat missing values by default. NaN and Zero write that marker instead and leave those rows out of the counts and the frequency denominator.
Developer & Automation Access
Run it from the terminal
Same engine as this page, headless — via the gizza CLI:
gizza tool frequency-encoder "product_id,price
X,10
Y,20
X,30
Z,40
X,50
Y,60" 'column=product_id'New to the CLI? Get gizza →
Open it by URL
Pre-fill and auto-run this tool with query parameters — the names match the API/CLI:
https://gizza.ai/tools/frequency-encoder/?data=product_id%2Cprice%0AX%2C10%0AY%2C20%0AX%2C30%0AZ%2C40%0AX%2C50%0AY%2C60&column=product_id&mode=count&output=replace&blank=count&min_count=0&case_sensitive=true&decimals=4&has_header=true&delimiter=commaMachine-readable descriptor: tool.json — title + parameters JSON Schema for agents.