Fix SR-CNN missing anomalies when Period > 0

[FanBaoMS]

We are excited to review your PR.

So we can do the best job, please check:

• There's a descriptive title that will make sense to other developers some time from now.
• There's associated issues. All PR's should have issue(s) associated - unless a trivial self-evident change such as fixing a typo. You can use the format Fixes #nnnn in your description to cause GitHub to automatically close the issue(s) when your PR is merged.
• Your change description explains what the change does, why you chose your approach, and anything else that reviewers should know.
• You have included any necessary tests in the same PR.

Fixes #5891 - DetectEntireAnomalyBySrCnn - No anomalies detected as of version 1.5.4 or greater

Root cause

Two domain-mismatch bugs in the Period > 0 code path of SrCnnEntireModeler:

1. Train() — z-score statistics computed in the wrong domain.
_mean / _std were computed from the raw input series before deseasonality ran. The z-score "false anomaly" filter inside SpectralResidual then compared _seriesToDetect[i] (the deseasonalized residual) against those raw-domain stats, producing meaningless z-scores that suppressed real anomalies.

2. GetMarginPeriod() — CalculateAnomalyScore called with exp in the wrong domain.
_ifftRe[i] was passed as the expected-value argument. CalculateAnomalyScore computes |exp - value| / unit, which is only meaningful when both arguments share a domain.

In the no-period sibling GetMargin(), this call is correct: CalculateExpectedValueByFft(_deAnomalyData) is invoked immediately beforehand and overwrites _ifftRe with the raw-domain expected value of the de-anomalized series — so _ifftRe[i] and values[i] share a domain there.

GetMarginPeriod() was added later (#5202) and reused the same call shape without that prior overwrite. In the period > 0 path _ifftRe holds the SR saliency map's real component computed on the deseasonalized residual series, not a raw-domain expected value. The resulting score collapses to roughly |raw_value| / unit — proportional to raw magnitude rather than deviation from expectation — so real anomalies often score near zero. This looks like a copy-paste of the no-period call site without re-deriving what _ifftRe actually holds in the new path.

Fix

1. Move the sum / _mean / _std computation in Train() to after the deseasonality call and compute the statistics against _seriesToDetect (the same series SR will actually score).
2. In GetMarginPeriod() pass results[i][3] (the raw-domain expected value just produced by GetExpectedValuePeriod) instead of _ifftRe[i].

The no-period path (Train with _period == 0 and GetMargin) is correct as written and is unchanged.

Test

Adds TestSrCnnAnomalyDetectorPhoneCalls, ported from the dotnet/samples PhoneCallsAnomalyDetection tutorial. Data file is taken verbatim from dotnet/samples.

The test asserts DetectSeasonality returns 7 and that DetectEntireAnomalyBySrCnn flags exactly indices {28, 44, 56, 70}. Prior to this fix the same call detects zero anomalies on the same input.

Sensitivity = 87.0 is used to obtain the boundary width the tutorial was originally calibrated for under the v1.5.2 _factors table; the table was rewritten in #5436, so the tutorial's literal Sensitivity = 64.0 now produces a much wider boundary that masks all but the strongest spike at index 28.

[FanBaoMS]

@dotnet-policy-service agree company="Microsoft"

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 69.90%. Comparing base (2fc13ba) to head (0aa8afe).
⚠️ Report is 7 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #7610      +/-   ##
==========================================
+ Coverage   69.88%   69.90%   +0.02%     
==========================================
  Files        1487     1487              
  Lines      276253   276294      +41     
  Branches    28291    28294       +3     
==========================================
+ Hits       193053   193138      +85     
+ Misses      75710    75671      -39     
+ Partials     7490     7485       -5     

Flag	Coverage Δ
Debug	69.90% <100.00%> (+0.02%)	⬆️
production	64.08% <100.00%> (+0.02%)	⬆️
test	89.83% <100.00%> (+<0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
...rosoft.ML.TimeSeries/SrCnnEntireAnomalyDetector.cs	93.51% <100.00%> (+1.14%)	⬆️
...crosoft.ML.TimeSeries.Tests/TimeSeriesDirectApi.cs	99.40% <100.00%> (+0.03%)	⬆️

... and 9 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[matouskozak]

LGTM! I updated the test a bit to test the fixed behavior.

[Copilot]

Copilot review overview

🟢 Approval recommended

The domain mismatches are corrected consistently and covered by a targeted end-to-end regression test.

Review effort: Balanced
Findings: None

What changed in this PR

Fixes SR-CNN anomaly suppression for seasonal data by keeping statistical comparisons and anomaly scoring within consistent domains.

Changes:

• Computes z-score statistics from deseasonalized data.
• Scores anomalies against the raw-domain expected value.
• Adds a phone-call dataset regression test.

File	Description
src/​Microsoft.ML.TimeSeries/​SrCnnEntireAnomalyDetector.cs	Corrects seasonal anomaly filtering and scoring.
test/​Microsoft.ML.TimeSeries.Tests/​TimeSeriesDirectApi.cs	Adds regression coverage for detected anomalies and scores.
test/​data/​Timeseries/​phone-calls.csv	Provides seasonal test data.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.