STUDY / 03

Explore Anomaly Detection

Find points that stand apart.

Algorithm

Compare density with nearby points.

9 / 90flagged

3.519highest score

1.344cutoff score

× Coral: flagged · cyan: unflagged

Compare with

Score how quickly splits isolate.

9 / 90flagged

0.737highest score

0.563cutoff score

× Coral: flagged · cyan: unflagged

Edit points with coordinates

Coordinates range from 0 to 1. You can also click the plot to add a point or remove a nearby point.

Compare ranked scores

Ranked by the left method. Higher is more unusual. Scores have different units across methods.

x	y	Local outlier factor	Isolation forest	Flagged by
0.148	0.901	3.5194	0.7374	Both
0.577	0.083	3.4938	0.6753	Both
0.115	0.545	2.0529	0.5891	Both
0.198	0.491	1.4989	0.5668	Both
0.469	0.313	1.4363	0.5079	Left
0.298	0.679	1.4249	0.5601	Left
0.353	0.694	1.4116	0.5347	Left
0.864	0.236	1.3843	0.6040	Both
0.565	0.680	1.3437	0.5650	Both
0.335	0.598	1.3318	0.4509	Neither
0.396	0.319	1.3197	0.5042	Neither
0.438	0.630	1.2876	0.4464	Neither
0.542	0.583	1.2487	0.4588	Neither
0.432	0.644	1.2470	0.4574	Neither
0.527	0.575	1.2448	0.4389	Neither
0.418	0.608	1.2447	0.4344	Neither
0.507	0.696	1.2278	0.5230	Neither
0.312	0.648	1.2247	0.5208	Neither
0.353	0.577	1.2095	0.4332	Neither
0.310	0.555	1.2072	0.4317	Neither
0.834	0.267	1.1883	0.5724	Right
0.493	0.561	1.1680	0.4154	Neither
0.278	0.536	1.1535	0.4888	Neither
0.308	0.561	1.1513	0.4324	Neither
0.524	0.650	1.1411	0.4905	Neither
0.452	0.611	1.1345	0.4262	Neither
0.516	0.585	1.1273	0.4369	Neither
0.338	0.398	1.1227	0.4739	Neither
0.464	0.661	1.1199	0.4744	Neither
0.480	0.645	1.1189	0.4614	Neither
0.518	0.604	1.1170	0.4496	Neither
0.344	0.500	1.1153	0.4314	Neither
0.431	0.382	1.1119	0.4441	Neither
0.311	0.498	1.1049	0.4494	Neither
0.823	0.251	1.0991	0.5191	Neither
0.293	0.554	1.0960	0.4594	Neither
0.319	0.432	1.0918	0.4682	Neither
0.821	0.251	1.0889	0.5176	Neither
0.308	0.444	1.0797	0.4845	Neither
0.814	0.276	1.0787	0.5635	Right
0.412	0.387	1.0721	0.4388	Neither
0.372	0.519	1.0695	0.4179	Neither
0.441	0.489	1.0678	0.3950	Neither
0.701	0.212	1.0661	0.5516	Neither
0.518	0.465	1.0651	0.4239	Neither
0.386	0.399	1.0647	0.4331	Neither
0.489	0.527	1.0625	0.4044	Neither
0.391	0.478	1.0577	0.4095	Neither
0.747	0.182	1.0576	0.5625	Right
0.367	0.480	1.0573	0.4247	Neither
0.436	0.586	1.0525	0.4076	Neither
0.516	0.469	1.0487	0.4252	Neither
0.514	0.457	1.0472	0.4216	Neither
0.427	0.456	1.0471	0.3944	Neither
0.739	0.224	1.0470	0.5100	Neither
0.491	0.393	1.0458	0.4390	Neither
0.366	0.432	1.0433	0.4375	Neither
0.401	0.524	1.0418	0.3944	Neither
0.490	0.445	1.0409	0.4150	Neither
0.460	0.494	1.0388	0.4111	Neither
0.471	0.501	1.0366	0.4071	Neither
0.494	0.464	1.0352	0.4127	Neither
0.533	0.440	1.0308	0.4453	Neither
0.716	0.221	1.0294	0.5332	Neither
0.796	0.241	1.0217	0.5119	Neither
0.744	0.193	1.0214	0.5464	Neither
0.508	0.411	1.0169	0.4522	Neither
0.535	0.430	1.0151	0.4585	Neither
0.387	0.424	1.0077	0.4272	Neither
0.400	0.542	1.0033	0.3909	Neither
0.747	0.215	1.0019	0.5150	Neither
0.460	0.533	1.0015	0.3935	Neither
0.400	0.548	0.9976	0.3918	Neither
0.462	0.401	0.9831	0.4240	Neither
0.438	0.541	0.9825	0.3852	Neither
0.403	0.560	0.9786	0.4007	Neither
0.725	0.259	0.9785	0.5016	Neither
0.428	0.520	0.9760	0.3861	Neither
0.471	0.559	0.9752	0.4059	Neither
0.768	0.257	0.9742	0.4921	Neither
0.768	0.250	0.9732	0.4924	Neither
0.448	0.449	0.9711	0.3991	Neither
0.448	0.420	0.9664	0.4214	Neither
0.438	0.555	0.9643	0.3846	Neither
0.448	0.546	0.9590	0.3783	Neither
0.443	0.544	0.9515	0.3774	Neither
0.443	0.552	0.9512	0.3829	Neither
0.752	0.263	0.9462	0.5023	Neither
0.773	0.206	0.9396	0.5328	Neither
0.750	0.259	0.9296	0.4839	Neither

Anomaly detection: methods and experiments ↗

Correlated cloud: compare axis-based Z-scores with covariance-aware Mahalanobis distance. Unequal density: compare local LOF with global distance. Clusters + outliers: compare robust median/MAD scores with ordinary mean/std scores.

kNN uses mean distance to k neighbors. Z-score and robust Z-score take the largest absolute coordinate score. Mahalanobis uses regularized empirical covariance, not a robust estimator. GMM scores negative log density, which can be negative; they are not probabilities. Flags always select the requested top fraction, even on uniform data. Ties break by point index. Original walkthrough ↗

Source / Rajiv’s original outlier app ↗