Make LOCI honor contamination and keep k as a parameter

[Mohit-Ak]

LOCI accepts a contamination argument and documents threshold_ as "based on
contamination ... the n_samples * contamination most abnormal samples", but
the value never reaches the thresholding logic. In __init__ the other parameter,
k, is written straight into self.threshold_:

def __init__(self, contamination=0.1, alpha=0.5, k=3):
    super(LOCI, self).__init__(contamination=contamination)
    self.alpha = alpha
    self.threshold_ = k        # <- k, not contamination

and fit() then labels against that same k, so contamination is dead weight.
Sweeping it changes nothing:

contamination=0.05  -> outlier frac=0.1583   threshold_=3
contamination=0.1   -> outlier frac=0.1583   threshold_=3
contamination=0.2   -> outlier frac=0.1583   threshold_=3
contamination=0.3   -> outlier frac=0.1583   threshold_=3
contamination=0.45  -> outlier frac=0.1583   threshold_=3

There's a second consequence of the same line. Because k is consumed instead of
stored, LOCI has no k attribute, so scikit-learn's get_params falls back to
None and a cloned estimator is silently broken:

>>> clf = LOCI(k=5)
>>> hasattr(clf, "k")
False
>>> clone(clf).fit(X)
TypeError: unsupported operand type(s) for *: 'NoneType' and 'float'

The existing test_model_clone doesn't catch this because it only clones, it never
fits the clone.

What this changes

Three small edits, all in loci.py:

• __init__ stores self.k = k instead of overwriting threshold_.
• The score loop's early-break test uses self.k — that's the outlier cutoff k
  is documented to be, and it's the only place it was ever really used.
• fit() calls self._process_decision_scores() instead of hand-rolling
  labels_/_mu/_sigma. That's the base-class helper every other detector in
  the package uses (lof, cof, sod, ...), and it derives threshold_ from
  contamination via the documented percentile rule.

The stale k=None in the class docstring's example repr is updated to k=3 to
match the actual default.

Scores are untouched — decision_function and _calculate_decision_score compute
exactly what they did before for a given k. Only the threshold used to turn those
scores into binary labels changes, which is the reported bug.

Testing

Ran the module's suite in a clean venv (numpy/scipy/scikit-learn/numba, pip install -e .):

$ python -m pytest pyod/test/test_loci.py -q
20 passed, 7 warnings in 138.83s

That includes three new cases added to TestLOCI, each of which fails on master
before the change:

• test_k_is_stored_as_parameter — AttributeError: 'LOCI' object has no attribute 'k'
• test_contamination_controls_labels — all three contamination values produced the
  identical outlier fraction
• test_threshold_matches_contamination — ACTUAL: 0.14 DESIRED: 0.25

After the change the same sweep tracks the requested rate:

contamination=0.05  -> outlier frac=0.0500
contamination=0.1   -> outlier frac=0.1000
contamination=0.2   -> outlier frac=0.2000
contamination=0.3   -> outlier frac=0.3000
contamination=0.45  -> outlier frac=0.4500

Detection quality is unchanged on the standard synthetic benchmark (train ROC 0.959,
test ROC 0.963 at contamination=0.1), and k still visibly drives the scores
(k=1/3/10 give mean scores 0.519/0.506/0.240), confirming it kept its real role.

Also ran pyod/test/test_base.py, test_data.py and test_utility.py (39 passed)
since fit() now routes through the shared base-class path. flake8 --max-line-length=127
on both touched files reports the same 25 pre-existing violations as on master —
no new ones.

Fixes #194

[yzhao062]

I reviewed this and the fix looks correct. CI is green, and I also verified it locally with the fixture in pyod/test/test_loci.py (n_train=200, contamination=0.1, random_state=42) on the current development branch:

contamination=0.05 -> outlier fraction 0.0500
contamination=0.15 -> 0.1500
contamination=0.30 -> 0.3000
contamination=0.45 -> 0.4500     (previously 0.1400 for every value)

LOCI(k=5).k          -> 5   (previously AttributeError)
clone(LOCI(k=5)).k   -> 5   (previously AttributeError; get_params()["k"] was None)
ROC on this fixture  -> 0.9633, unchanged before and after (floor is 0.8)

The ROC being identical is expected and worth stating explicitly: this changes how labels_ are thresholded, not the ranking produced by decision_function, so there's no accuracy regression to worry about either way.

Your read of the root cause is right: self.threshold_ = k in __init__ consumed k in place of contamination, so the documented "n_samples * contamination most abnormal samples" behavior never applied, and k was never stored as an attribute for get_params / clone. Routing through _process_decision_scores() is the right fix — it's the same path the other detectors use to derive threshold_ and labels_ from contamination. This addresses the contamination half of #194.

Two things before it can be merged:

1. The PR is still marked as a draft. Is it ready for review, or were you planning further changes?
2. Please retarget the base branch from master to development — that's our integration branch, and master is currently behind it.

Thanks for this — it's been a latent bug since LOCI landed.

[Mohit-Ak]

Thanks for the thorough check — good to see the contamination sweep and the clone path reproduce on your side too, and agreed that the unchanged ROC is the right thing to call out: decision_scores_ and the ranking are untouched, only the threshold_/labels_ derivation moves onto _process_decision_scores().

I've retargeted the base branch to development.

On the draft status: I'm keeping it as a draft until Mohit signs off on it — that's my standing rule for anything opened under this account, not a sign of pending work. The change is complete from my side and I'm not planning further commits. I'll flip it to ready-for-review as soon as he confirms, which should be within the day.

One note on scope while it's in front of you: this covers the contamination half of #194 plus the k/get_params breakage I hit while testing it. If there's a second part of #194 you'd want addressed in the same PR rather than a follow-up, say so and I'll fold it in before it goes up for review.

[Mohit-Ak]

Marked ready for review — thanks for waiting, and for taking the time to reproduce the sweep independently.

Nothing changed since your review: the base is on development, the diff is still the same two files, and Testing + CodeQL are green on 910869e. Happy to rebase or split anything out if it makes the review easier.