Loda now performs density estimation on the fly

Author: andrewdalpino

CHANGELOG.md

@@ -1,14 +1,15 @@
-- Unreleased
+- 0.0.13-beta
     - Added documentation site
     - Added Regression and Classification Loss interfaces
     - Robust Z Score is now a Ranking anomaly detector
-    - LODA now defaults to auto detect bin count
+    - Loda now defaults to auto detect bin count
     - Removed tolerance param from Gradient Boost and AdaBoost
     - Screen logger timestamp format now configurable
     - Dropped Persistable contract between SVM-based learners
     - Random Forest feature importances now serial
     - Removed Robust Z Score tolerance parameter
     - Added slice method to Dataset API
+    - Loda now performs density estimation on the fly
 
 - 0.0.12-beta
     - Added AdaMax neural network Optimizer
@@ -33,7 +34,7 @@
     - Removed Model Orchestra meta-estimator
     - Grid Search automatically retrains base estimator
     - Added neural net Parameter namespace and interface
-    - Changed order of LODA hyperparameters
+    - Changed order of Loda hyperparameters
     - Replaced F1 Score with F Beta metric
     - Removed ISRU and Gaussian activation functions
     - Fixed SELU derivitive computation
@@ -69,7 +70,7 @@
     - Pipeline is now more verbose
 
 - 0.0.10-beta
-    - Added LODA online anomaly detector
+    - Added Loda online anomaly detector
     - Added Radius Neighbors classifier and regressor
     - Added fast k-d LOF anomaly detector
     - Added base Ball Tree implementation
@@ -84,7 +85,7 @@
     - K Means is no longer Online
     - Removed Sign function
     - Added Binary Tree interface
-    - Added bin count heuristic to LODA
+    - Added bin count heuristic to Loda
     - Changed order of k-d neighbors hyperparameters
     - Removed Hamming distance kernel
 

docs/anomaly-detectors/loda.md

@@ -1,6 +1,6 @@
-<span style="float:right;"><a href="https://github.com/RubixML/RubixML/blob/master/src/AnomalyDetectors/LODA.php">Source</a></span>
+<span style="float:right;"><a href="https://github.com/RubixML/RubixML/blob/master/src/AnomalyDetectors/Loda.php">Source</a></span>
 
-# LODA
+# Loda
 *Lightweight Online Detector of Anomalies* uses a sparse random projection matrix to produce an ensemble of unique one dimensional equi-width histograms able to estimate the probability density of an unknown sample. The decision function takes the anomaly score given by the negative log likelihood whose upper threshold can be set by the user.
 
 **Interfaces:** [Estimator](../estimator.md), [Learner](../learner.md), [Online](../online.md), [Ranking](api.md#ranking), [Persistable](../persistable.md)
@@ -22,13 +22,13 @@ public static estimateBins($dataset) : int
 
 ### Example
 ```php
-use Rubix\ML\AnomalyDetection\LODA;
+use Rubix\ML\AnomalyDetection\Loda;
 
-$bins = LODA::estimateBins(1000);
+$bins = Loda::estimateBins(1000);
 
-$estimator = new LODA($bins); // Automatically choose bin count
+$estimator = new Loda($bins); // Automatically choose bin count
 
-$estimator = new LODA(5, 250, 3.5); // Specifying bins
+$estimator = new Loda(5, 250, 3.5); // Specifying bins
 ```
 
 ### References

mkdocs.yml

@@ -35,7 +35,7 @@ nav:
   - API Reference: anomaly-detectors/api.md
   - Isolation Forest: anomaly-detectors/isolation-forest.md
   - K-d LOF: anomaly-detectors/kd-lof.md
-  - LODA: anomaly-detectors/loda.md
+  - Loda: anomaly-detectors/loda.md
   - Local Outlier Factor: anomaly-detectors/local-outlier-factor.md
   - One Class SVM: anomaly-detectors/one-class-svm.md
   - Robust Z Score: anomaly-detectors/robust-z-score.md

src/AnomalyDetectors/LODA.php renamed to src/AnomalyDetectors/Loda.php

@@ -18,13 +18,13 @@
 use const Rubix\ML\LOG_EPSILON;
 
 /**
- * LODA
+ * Loda
  *
- * Lightweight Online Detector of Anomalies uses sparse random projection vectors
- * to generate an ensemble of unique one dimensional equi-width histograms able
- * to estimate the probability density of an unknown sample. The anomaly score is
- * given by the negative log likelihood whose upper threshold can be set by the
- * user through the *contamination* hyper-parameter.
+ * Lightweight Online Detector of Anomalies a.k.a. Loda uses sparse random
+ * projection vectors to generate an ensemble of unique one dimensional
+ * equi-width histograms able to estimate the probability density of an unknown
+ * sample. The anomaly score is given by the negative log likelihood whose upper
+ * threshold can be set by the user through the *contamination* hyper-parameter.
  *
  * References:
  * [1] T. Pevný. (2015). Loda: Lightweight on-line detector of anomalies.
@@ -34,7 +34,7 @@
  * @package     Rubix/ML
  * @author      Andrew DalPino
  */
-class LODA implements Estimator, Learner, Online, Ranking, Persistable
+class Loda implements Estimator, Learner, Online, Ranking, Persistable
 {
     protected const MIN_SPARSE_DIMENSIONS = 3;
 
@@ -74,7 +74,7 @@ class LODA implements Estimator, Learner, Online, Ranking, Persistable
     protected $r;
 
     /**
-     * The edges, counts, and precomputed probability densities of each histogram.
+     * The edges, and bin counts of each histogram.
      *
      * @var array[]|null
      */
@@ -83,9 +83,9 @@ class LODA implements Estimator, Learner, Online, Ranking, Persistable
     /**
      * The number of samples that have been learned so far.
      *
-     * @var int|null
+     * @var int
      */
-    protected $n;
+    protected $n = 0;
 
     /**
      * Estimate the number of bins from the number of samples in a dataset.
@@ -212,15 +212,7 @@ public function train(Dataset $dataset) : void
                 $counts[$this->bins]++;
             }
 
-            $densities = [];
-
-            foreach ($counts as $count) {
-                $densities[] = $count > 0
-                    ? -log($count / $m)
-                    : -LOG_EPSILON;
-            }
-
-            $this->histograms[] = [$edges, $counts, $densities];
+            $this->histograms[] = [$edges, $counts];
         }
 
         $this->n = $m;
@@ -234,22 +226,20 @@ public function train(Dataset $dataset) : void
      */
     public function partial(Dataset $dataset) : void
     {
-        if (!$this->r or !$this->histograms or !$this->n) {
+        if (!$this->r or !$this->histograms) {
             $this->train($dataset);
 
             return;
         }
 
         DatasetIsCompatibleWithEstimator::check($dataset, $this);
 
-        $this->n += $dataset->numRows();
-
         $projections = Matrix::quick($dataset->samples())
             ->matmul($this->r)
             ->transpose();
 
         foreach ($projections as $i => $values) {
-            [$edges, $counts, $densities] = $this->histograms[$i];
+            [$edges, $counts] = $this->histograms[$i];
 
             $interior = array_slice($edges, 1, $this->bins, true);
 
@@ -265,14 +255,10 @@ public function partial(Dataset $dataset) : void
                 $counts[$this->bins]++;
             }
 
-            foreach ($counts as $j => $count) {
-                $densities[$j] = $count > 0
-                    ? -log($count / $this->n)
-                    : -LOG_EPSILON;
-            }
-
-            $this->histograms[$i] = [$edges, $counts, $densities];
+            $this->histograms[$i] = [$edges, $counts];
         }
+
+        $this->n += $dataset->numRows();
     }
 
     /**
@@ -308,27 +294,31 @@ public function rank(Dataset $dataset) : array
             ->matmul($this->r)
             ->transpose();
 
-        $scores = array_fill(0, $projections->n(), 0.);
+        $densities = array_fill(0, $projections->n(), 0.);
 
         foreach ($projections as $i => $values) {
-            [$edges, $counts, $densities] = $this->histograms[$i];
+            [$edges, $counts] = $this->histograms[$i];
 
             foreach ($values as $j => $value) {
                 foreach ($edges as $k => $edge) {
                     if ($value < $edge) {
-                        $scores[$j] += $densities[$k];
+                        $count = $counts[$k];
+
+                        $densities[$j] += $count > 0
+                            ? -log($counts[$k] / $this->n)
+                            : -LOG_EPSILON;
 
                         break 1;
                     }
                 }
             }
         }
 
-        foreach ($scores as &$score) {
-            $score /= $this->estimators;
+        foreach ($densities as &$density) {
+            $density /= $this->estimators;
         }
 
-        return $scores;
+        return $densities;
     }
 
     /**

tests/AnomalyDetectors/LODATest.php renamed to tests/AnomalyDetectors/LodaTest.php

@@ -7,7 +7,7 @@
 use Rubix\ML\Estimator;
 use Rubix\ML\Persistable;
 use Rubix\ML\Datasets\Unlabeled;
-use Rubix\ML\AnomalyDetectors\LODA;
+use Rubix\ML\AnomalyDetectors\Loda;
 use Rubix\ML\Other\Helpers\DataType;
 use Rubix\ML\AnomalyDetectors\Ranking;
 use Rubix\ML\Datasets\Generators\Blob;
@@ -18,7 +18,7 @@
 use InvalidArgumentException;
 use RuntimeException;
 
-class LODATest extends TestCase
+class LodaTest extends TestCase
 {
     protected const TRAIN_SIZE = 400;
     protected const TEST_SIZE = 10;
@@ -39,7 +39,7 @@ public function setUp()
             1 => new Circle(0., 0., 8., 0.1),
         ], [0.9, 0.1]);
 
-        $this->estimator = new LODA(null, 100, 10.);
+        $this->estimator = new Loda(null, 100, 10.);
 
         $this->metric = new FBeta();
 
@@ -48,7 +48,7 @@ public function setUp()
 
     public function test_build_detector()
     {
-        $this->assertInstanceOf(LODA::class, $this->estimator);
+        $this->assertInstanceOf(Loda::class, $this->estimator);
         $this->assertInstanceOf(Learner::class, $this->estimator);
         $this->assertInstanceOf(Online::class, $this->estimator);
         $this->assertInstanceOf(Ranking::class, $this->estimator);
@@ -65,11 +65,11 @@ public function test_build_detector()
 
     public function test_estimate_bins()
     {
-        $this->assertSame(4, LODA::estimateBins(10));
-        $this->assertSame(8, LODA::estimateBins(100));
-        $this->assertSame(11, LODA::estimateBins(1000));
-        $this->assertSame(14, LODA::estimateBins(10000));
-        $this->assertSame(18, LODA::estimateBins(100000));
+        $this->assertSame(4, Loda::estimateBins(10));
+        $this->assertSame(8, Loda::estimateBins(100));
+        $this->assertSame(11, Loda::estimateBins(1000));
+        $this->assertSame(14, Loda::estimateBins(10000));
+        $this->assertSame(18, Loda::estimateBins(100000));
     }
 
     public function test_train_partial_predict()