Confusion Matrix | Zyeon UI

Preview in your theme

status="ready" · normalize="none" · click a cell or press Enter to pin it

Species predictionsCamera-trap classifier v7 · held-out night frames · 5 classes · 1,520 frames · accuracy 85.8% · macro F1 0.79 · worst confusion Grey wolf → Feral dog, 38

Predicted →Actual ↓	Wildcat	Feral dog	Red fox	Grey wolf	Eurasian lynx	Recall
Wildcatn 480	441	6	5	2	26	91.9%
Feral dogn 512	4	470	9	27	2	91.8%
Red foxn 246	3	12	205	24	2	83.3%
Grey wolfn 198	2	38	17	141	0	71.2%
Eurasian lynxn 84	32	0	3	2	47	56%
Precision	91.5%	89.4%	85.8%	71.9%	61%	85.8%

0470correct — the diagonalconfused — hatched, so the two never rely on colour alone

Shading runs from empty to the biggest cell on the grid (470). With uneven class sizes one diagonal tile owns that ceiling — normalize by row or column to compare classes instead. An exact zero gets no ink at all. Click or press Enter to pin a cell; Escape releases it.

normalize="row" — each row divided by its own support, so the diagonal reads as recall and the 84-frame Lynx class stops being invisible

Species predictionsSame 1,520 frames, read as recall · 5 classes · 1,520 frames · accuracy 85.8% · macro F1 0.79 · worst confusion Grey wolf → Feral dog, 38

Predicted →Actual ↓	Wildcat	Feral dog	Red fox	Grey wolf	Eurasian lynx	Recall
Wildcatn 480	91.9%	1.3%	1%	0.4%	5.4%	91.9%
Feral dogn 512	0.8%	91.8%	1.8%	5.3%	0.4%	91.8%
Red foxn 246	1.2%	4.9%	83.3%	9.8%	0.8%	83.3%
Grey wolfn 198	1%	19.2%	8.6%	71.2%	0%	71.2%
Eurasian lynxn 84	38.1%	0%	3.6%	2.4%	56%	56%
Precision	91.5%	89.4%	85.8%	71.9%	61%	85.8%

0%100%correct — the diagonalconfused — hatched, so the two never rely on colour alone

Every row is divided by its own support, so a row sums to 100% and the diagonal reads as recall. An exact zero gets no ink at all. Click or press Enter to pin a cell; Escape releases it.

normalize="column" — each column divided by everything called that class, so the diagonal reads as precision

Species predictionsSame 1,520 frames, read as precision · 5 classes · 1,520 frames · accuracy 85.8% · macro F1 0.79 · worst confusion Grey wolf → Feral dog, 38

Predicted →Actual ↓	Wildcat	Feral dog	Red fox	Grey wolf	Eurasian lynx	Recall
Wildcatn 480	91.5%	1.1%	2.1%	1%	33.8%	91.9%
Feral dogn 512	0.8%	89.4%	3.8%	13.8%	2.6%	91.8%
Red foxn 246	0.6%	2.3%	85.8%	12.2%	2.6%	83.3%
Grey wolfn 198	0.4%	7.2%	7.1%	71.9%	0%	71.2%
Eurasian lynxn 84	6.6%	0%	1.3%	1%	61%	56%
Precision	91.5%	89.4%	85.8%	71.9%	61%	85.8%

0%100%correct — the diagonalconfused — hatched, so the two never rely on colour alone

Every column is divided by everything called that class, so a column sums to 100% and the diagonal reads as precision. An exact zero gets no ink at all. Click or press Enter to pin a cell; Escape releases it.

2 × 2 at real imbalance — 9,400 legitimate against 210 fraudulent, so raw counts are one tile and three specks

Fraud screeningPayment risk model · one day of authorisations · 2 classes · 9,610 transactions · accuracy 97% · macro F1 0.74 · worst confusion Legitimate → Fraudulent, 213

Predicted →Actual ↓	Legitimate	Fraudulent	Recall
Legitimaten 9,400	9,187	213	97.7%
Fraudulentn 210	74	136	64.8%
Precision	99.2%	39%	97%

09,187correct — the diagonalconfused — hatched, so the two never rely on colour alone

Shading runs from empty to the biggest cell on the grid (9,187). With uneven class sizes one diagonal tile owns that ceiling — normalize by row or column to compare classes instead. An exact zero gets no ink at all. Click or press Enter to pin a cell; Escape releases it.

14 classes · showValues="auto" gives up printing past 12, so the grid becomes a texture map and the readout carries the numbers

Document typeDocument router · confusions gather beside the diagonal · 14 classes · 3,104 documents · accuracy 73.4% · macro F1 0.72 · worst confusion Tax return → Insurance claim, 69

Predicted →Actual ↓	Invoice	Purchase order	Delivery note	Credit note	Bank statement	Payslip	Tax return	Insurance claim	Rental agreement	Utility bill	Medical report	Court filing	Shipping manifest	Certificate of origin	Recall
Invoicen 283															72.8%
Purchase ordern 201															65.2%
Delivery noten 140															88.6%
Credit noten 247															75.3%
Bank statementn 224															70.1%
Payslipn 118															61%
Tax returnn 339															64%
Insurance claimn 182															83%
Rental agreementn 140															86.4%
Utility billn 320															68.4%
Medical reportn 308															59.1%
Court filingn 79															74.7%
Shipping manifestn 328															91.2%
Certificate of originn 195															79.5%
Precision	88%	71.2%	57.9%	74.4%	79.7%	46.8%	81.3%	61.9%	68.8%	79.6%	72.8%	36.2%	89.3%	96.3%	73.4%

0%100%correct — the diagonalconfused — hatched, so the two never rely on colour alone

Every row is divided by its own support, so a row sums to 100% and the diagonal reads as recall. An exact zero gets no ink at all. Point at a cell, or arrow into the grid, to read its count. Click or press Enter to pin a cell; Escape releases it.

margins={false} · showValues={false} · cellSize={24} — the dashboard-tile density, where the grid is a shape and nothing else

Species predictionsCompact variant · 5 classes · 1,520 frames · accuracy 85.8% · macro F1 0.79 · worst confusion Grey wolf → Feral dog, 38

Predicted →Actual ↓	Wildcat	Feral dog	Red fox	Grey wolf	Eurasian lynx
Wildcatn 480
Feral dogn 512
Red foxn 246
Grey wolfn 198
Eurasian lynxn 84

0%100%correct — the diagonalconfused — hatched, so the two never rely on colour alone

Degenerate data — a class with no samples, a class never predicted, and three tallies that could not be counted

Support ticket intentsIntent router · evaluation slice · 4 classes · 140 tickets · accuracy 79.3% · macro F1 0.82 · worst confusion Refund request → Billing question, 11

Predicted →Actual ↓	Refund request	Billing question	Escalate to human	Spam	Recall
Refund requestn 77	81.8%	14.3%	3.9%	0%	81.8%
Billing questionn 57	15.8%	84.2%	0%	0%	84.2%
Escalate to humann 0	—	—	—	—	—
Spamn 6	66.7%	33.3%	0%	0%	0%
Precision	82.9%	78.7%	0%	—	79.3%

0%100%correct — the diagonalconfused — hatched, so the two never rely on colour alone

Every row is divided by its own support, so a row sums to 100% and the diagonal reads as recall. An exact zero gets no ink at all. Click or press Enter to pin a cell; Escape releases it.

2 tallies carried a count that is not a count (negative, NaN or Infinity) and 1 named a class outside the label set. They are not in any figure on this card.

A one-class label set — the smallest matrix there is, and the layout still has to hold

Surface defectsSingle-class detector · line 4 · 1 class · 128 panels · accuracy 100% · macro F1 1.00 · every prediction landed on the diagonal

Predicted →Actual ↓	Surface defect	Recall
Surface defectn 128	128	100%
Precision	100%	100%

0128correct — the diagonalconfused — hatched, so the two never rely on colour alone

Shading runs from empty to the biggest cell on the grid (128). With uneven class sizes one diagonal tile owns that ceiling — normalize by row or column to compare classes instead. An exact zero gets no ink at all. Click or press Enter to pin a cell; Escape releases it.

status="empty" — the run finished and scored nothing

Species predictionsRows are what the samples are, columns are what the model called them

Nothing scored yet

A confusion matrix needs a label set and at least one scored frame — one row per prediction, each carrying the class it really was and the class the model called it.

status="loading" — skeleton only, with its own diagonal, and no table anywhere near the sr-only line

Species predictionsRows are what the samples are, columns are what the model called them

status="error" with onRetry — retried 0 times

Species predictionsRows are what the samples are, columns are what the model called them

Couldn't load the matrix

The evaluation run didn't report back.

Loading preview…

This component is part of Pro

The install command, prompt and concept map are included with a Pro license — one purchase unlocks every Pro component, forever.

Unlock with Pro

Predicted →Actual ↓

Wildcat

Feral dog

Red fox

Grey wolf

Eurasian lynx

Recall

Wildcatn 480

441

91.9%

Feral dogn 512

470

91.8%

Red foxn 246

205

83.3%

Grey wolfn 198

141

71.2%

Eurasian lynxn 84

56%

Precision

91.5%

89.4%

85.8%

71.9%

61%

85.8%

Predicted →Actual ↓

Wildcat

Feral dog

Red fox

Grey wolf

Eurasian lynx

Recall

Wildcatn 480

91.9%

1.3%

0.4%

5.4%

91.9%

Feral dogn 512

0.8%

91.8%

1.8%

5.3%

0.4%

91.8%

Red foxn 246

1.2%

4.9%

83.3%

9.8%

0.8%

83.3%

Grey wolfn 198

19.2%

8.6%

71.2%

Eurasian lynxn 84

38.1%

3.6%

2.4%

56%

Precision

91.5%

89.4%

85.8%

71.9%

61%

85.8%

Predicted →Actual ↓

Wildcat

Feral dog

Red fox

Grey wolf

Eurasian lynx

Recall

Wildcatn 480

91.5%

1.1%

2.1%

33.8%

91.9%

Feral dogn 512

0.8%

89.4%

3.8%

13.8%

2.6%

91.8%

Red foxn 246

0.6%

2.3%

85.8%

12.2%

2.6%

83.3%

Grey wolfn 198

0.4%

7.2%

7.1%

71.9%

71.2%

Eurasian lynxn 84

6.6%

1.3%

61%

56%

Precision

91.5%

89.4%

85.8%

71.9%

61%

85.8%

Predicted →Actual ↓

Legitimate

Fraudulent

Recall

Legitimaten 9,400

9,187

213

97.7%

Fraudulentn 210

136

64.8%

Precision

99.2%

39%

97%

Predicted →Actual ↓

Invoice

Purchase order

Delivery note

Credit note

Bank statement

Payslip

Tax return

Insurance claim

Rental agreement

Utility bill

Medical report

Court filing

Shipping manifest

Certificate of origin

Recall

Invoicen 283

72.8%

Purchase ordern 201

65.2%

Delivery noten 140

88.6%

Credit noten 247

75.3%

Bank statementn 224

70.1%

Payslipn 118

61%

Tax returnn 339

64%

Insurance claimn 182

83%

Rental agreementn 140

86.4%

Utility billn 320

68.4%

Medical reportn 308

59.1%

Court filingn 79

74.7%

Shipping manifestn 328

91.2%

Certificate of originn 195

79.5%

Precision

88%

71.2%

57.9%

74.4%

79.7%

46.8%

81.3%

61.9%

68.8%

79.6%

72.8%

36.2%

89.3%

96.3%

73.4%

Predicted →Actual ↓

Wildcat

Feral dog

Red fox

Grey wolf

Eurasian lynx

Wildcatn 480

Feral dogn 512

Red foxn 246

Grey wolfn 198

Eurasian lynxn 84

Predicted →Actual ↓

Refund request

Billing question

Escalate to human

Spam

Recall

Refund requestn 77

81.8%

14.3%

3.9%

81.8%

Billing questionn 57

15.8%

84.2%

Escalate to humann 0

—

Spamn 6

66.7%

33.3%

Precision

82.9%

78.7%

—

79.3%

Predicted →Actual ↓

Surface defect

Recall

Surface defectn 128

128

100%

Precision

100%