Description
Derive full one-versus-rest TP/FP/FN/TN records for every label; no division, threshold choice or prevalence adjustment.
A confusion table uses actual labels as rows and predicted labels as columns. Cells can contain occurrence counts or supplied nonnegative weights; the axis label order must stay with the matrix. Row, column and joint normalization have different denominators, and a zero denominator needs an explicit undefined result.
Derive full one-versus-rest TP/FP/FN/TN records for every label; no division, threshold choice or prevalence adjustment.
When to use this tool
Use this tool to derive full one-versus-rest TP/FP/FN/TN records for every label; no division, threshold choice or prevalence adjustment.
How to use the tool and how it works
- Enter matrix, labels using the input definitions below.
- Run the operation. Resolve malformed-input errors before interpreting the result.
- Read result.classes. Interpret signs, units, ordering and null values using the rules below.
Sum the entire table. For each label set tp to its diagonal cell, fn to its remaining row sum, fp to its remaining column sum and tn to total−tp−fn−fp.1
| Input field | Meaning and units | Accepted input |
|---|---|---|
| matrix | Square nonnegative finite table: matrix[actualIndex][predictedIndex], both axes in labels order. | Dense list of at most 128 rows of at most 128 finite numbers; the operation’s shape constraints also apply. |
| labels | Distinct label strings defining actual row and predicted column order; score tools require at least one label. | Dense list of at most 128 strings, each at most 8,192 UTF-16 code units. |
| Output field | Interpretation |
|---|---|
| result | Object containing the fields below. |
| result.classes | Input-label-order records {label,tp,fp,fn,tn}, retaining count/weight units. |
{
"input": {
"matrix": [
[
2,
1
],
[
0,
1
]
],
"labels": [
"a",
"b"
]
},
"output": {
"result": {
"classes": [
{
"label": "a",
"tp": 2,
"fp": 0,
"fn": 1,
"tn": 1
},
{
"label": "b",
"tp": 1,
"fp": 1,
"fn": 0,
"tn": 2
}
]
}
}
}Limitations and assumptions
- One-versus-rest records retain support units and do not select thresholds, divide by denominators or adjust for prevalence.
- Label strings are compared exactly. Confusion values may be nonnegative fractional weights; a count interpretation requires count data. Probability-like score rows use [0,1] entries with sum tolerance 10⁻⁸, without establishing calibration.
- Every declared top-level input is required and unknown fields are rejected. Ordinary strings are limited to 8,192 UTF-16 code units and lists to 128 entries, with tighter operation-specific bounds where stated. JSON text has bounded nesting and node counts.
- Numeric outputs must be finite and within the absolute safe-integer magnitude; exact larger integers use decimal text. Source positions are zero-based UTF-16 offsets and span ends are exclusive unless a different convention is explicitly stated.
Alternative or Complementary methods
Check label coverage and axis order before comparing tables. Use a separate evaluation design for uncertainty, calibration, operating thresholds and the consequences of real-world decisions.
References
Similar or alternative tools
- Confusion Actual Support Table
Report actual row support totals without calculating per-class rates.
- Confusion Column Rate Table
Normalize each predicted column independently with null for zero-support columns; no calibrated probability claim.