pyod 0.1.2

A Python Outlier Detection (Anomaly Detection) Toolbox

Python Outlier Detection (PyOD)

Note: PyOD is still under development without full test coverages. However, it has been successfully used in the various academic research projects [8, 9].

More anomaly detection related resources, e.g., books, papers and videos, can be found at anomaly-detection-resources

---

• Python Outlier Detection (PyOD)
  • Quick Introduction
  • Installation
  • API Cheatsheet
  • Quick Start for Outlier Detection
  • Quick Start for Combining Outlier Scores from Various Base Detectors
  • Reference

---

Quick Introduction

PyOD is a Python-based toolkit to identify outliers in data with both unsupervised and supervised algorithms. It strives to provide unified APIs across for different anomaly detection algorithms. The toolkit consists of three major groups of functionalities:

• Individual Detection Algorithms:

  1. Local Outlier Factor, LOF (wrapped on sklearn implementation) [1]
  2. Isolation Forest, iForest (wrapped on sklearn implementation) [2]
  3. One-Class Support Vector Machines (wrapped on sklearn implementation) [3]
  4. kNN Outlier Detection (use the distance to the kth nearst neighbor as the outlier score)
  5. Average KNN Outlier Detection (use the average distance to k nearst neighbors as the outlier score)
  6. Median KNN Outlier Detection (use the median distance to k nearst neighbors as the outlier score)
  7. Global-Local Outlier Score From Hierarchies [4]
  8. Histogram-based Outlier Score, HBOS [5]
  9. Angle-Based Outlier Setection, ABOD [7]

• Ensemble Framework (Outlier Score Combination Frameworks)

  1. Feature bagging
  2. Average of Maximum (AOM) [6]
  3. Maximum of Average (MOA) [6]
  4. Threshold Sum (Thresh) [6]

• Utility functions:

  1. scores_to_lables(): converting raw outlier scores to binary labels
  2. precision_n_scores(): one of the popular evaluation metrics for outlier mining (precision @ rank n)

Please be advised the purpose of the toolkit is for quick exploration. Using it as the final output should be understood with cautions. Fine-tunning may be needed to generate meaningful results. It is recommended to be used for the first-step data exploration only. Due to the restriction of time, the unit tests are not supplied but have been planned to implement.

Installation

It is advised to install with pip to manage the package:

pip install pyod

Alternatively, downloading/cloning the Github repository also works.

---

API Cheatsheet

For all algorithms implemented/wrapped in PyOD, the similar API is forced for consistency.

• fit(): fitting the model with the training data
• decision_function(): return raw outlier scores for test data
• predict(): returning binary outlier labels of test data
• predict_proba(): returning outlier probability of test data (0 to 1)
• predict_rank(): returning outlier rank of test data (data outlyness rank in training data)

Import outlier detection models:

from pyod.models import Knn
from pyod.models import ABOD

Import utility functions:

from pyod.util.utility import precision_n_scores

---

Quick Start for Outlier Detection

"examples/knn_example.py" is an example to demo the basic API of PyOD with kNN detector. It is noted the APIs for other detectors are similar.

1. Generate sample data first; normal data is generated by a 2-d gaussian distribution, and outliers are generated by a 2-d uniform distribution.

   contamination = 0.1  # percentage of outliers
   n_train = 1000  # number of training points
   n_test = 500  # number of testing points
   
   X_train, y_train, c_train, X_test, y_test, c_test = generate_data(
       n=n_train, contamination=contamination, n_test=n_test)
   

2. Initialize a kNN detector, fit the model, and make the prediction.

   # train a k-NN detector (default parameters, k=10)
   clf = Knn()
   clf.fit(X_train)
   
   # get the prediction label and scores on the training data
   y_train_pred = clf.y_pred
   y_train_score = clf.decision_scores
   
   # get the prediction on the test data
   y_test_pred = clf.predict(X_test)  # outlier label (0 or 1)
   y_test_score = clf.decision_function(X_test)  # outlier scores
   

3. Evaluate the prediction by ROC and Precision@rank n (p@n):

   print('Train ROC:{roc}, precision@n:{prn}'.format(
       roc=roc_auc_score(y_train, y_train_score),
       prn=precision_n_scores(y_train, y_train_score)))
   
   print('Test ROC:{roc}, precision@n:{prn}'.format(
       roc=roc_auc_score(y_test, y_test_score),
       prn=precision_n_scores(y_test, y_test_score)))
   

   See a sample output:

   Train ROC:0.9473, precision@n:0.7857
   Test ROC:0.992, precision@n:0.9
   

To check the result of the classification visually (knn_figure):

---

Quick Start for Combining Outlier Scores from Various Base Detectors

"comb_example.py" is a quick demo for showing the API for combining multiple algorithms. Given we have n individual outlier detectors, each of them generates an individual score for all samples. The task is to combine the outputs from these detectors effectivelly.

Key Step: conducting Z-score normalization on raw scores before the combination. Four combination mechanisms are shown in this demo:

1. Mean: use the mean value of all scores as the final output.
2. Max: use the max value of all scores as the final output.
3. Average of Maximum (AOM): first randomly split n detectors in to p groups. For each group, use the maximum within the group as the group output. Use the average of all group outputs as the final output.
4. Maximum of Average (MOA): similarly to AOM, the same grouping is introduced. However, we use the average of a group as the group output, and use maximum of all group outputs as the final output. To better understand the merging techniques, refer to [6].

The walkthrough of the code example is provided:

1. First initialize 20 kNN outlier detectors with different k (10 to 200), and get the outlier scores:

   # initialize 20 base detectors for combination
   k_list = [10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140,
                       150, 160, 170, 180, 190, 200]
   
   train_scores = np.zeros([X_train.shape[0], n_clf])
   test_scores = np.zeros([X_test.shape[0], n_clf])
   
   for i in range(n_clf):
           k = k_list[i]
   
           clf = Knn(n_neighbors=k, method='largest')
           clf.fit(X_train_norm)
   
           train_scores[:, i] = clf.decision_scores.ravel()
           test_scores[:, i] = clf.decision_function(X_test_norm).ravel()
   

2. Then the output codes are standardized into zero mean and unit std before combination.

   # scores have to be normalized before combination
   train_scores_norm, test_scores_norm = standardizer(train_scores, test_scores)
   

3. Then four different combination algorithms are applied as described above:

   comb_by_mean = np.mean(test_scores_norm, axis=1)
   comb_by_max = np.max(test_scores_norm, axis=1)
   comb_by_aom = aom(test_scores_norm, 5, 20) # 5 groups
   comb_by_moa = moa(test_scores_norm, 5, 20)) # 5 groups
   

4. Finally, all four combination methods are evaluated with 20 iterations:

   Summary of 10 iterations
   comb by mean, ROC: 0.9196, precision@n: 0.5464
   comb by max, ROC: 0.9198, precision@n: 0.5532
   comb by aom, ROC: 0.9260, precision@n: 0.5630
   comb by moa, ROC: 0.9244, precision@n: 0.5523
   

Reference

[1] Breunig, M.M., Kriegel, H.P., Ng, R.T. and Sander, J., 2000, May. LOF: identifying density-based local outliers. In ACM SIGMOD Record, pp. 93-104. ACM.

[2] Liu, F.T., Ting, K.M. and Zhou, Z.H., 2008, December. Isolation forest. In ICDM '08, pp. 413-422. IEEE.

[3] Ma, J. and Perkins, S., 2003, July. Time-series novelty detection using one-class support vector machines. In IJCNN' 03, pp. 1741-1745. IEEE.

[4] Campello, R.J., Moulavi, D., Zimek, A. and Sander, J., 2015. Hierarchical density estimates for data clustering, visualization, and outlier detection. TKDD, 10(1), pp.5.

[5] Goldstein, M. and Dengel, A., 2012. Histogram-based outlier score (hbos): A fast unsupervised anomaly detection algorithm. In KI-2012: Poster and Demo Track, pp.59-63.

[6] Aggarwal, C.C. and Sathe, S., 2015. Theoretical foundations and algorithms for outlier ensembles.ACM SIGKDD Explorations Newsletter, 17(1), pp.24-47.

[7] Kriegel, H.P. and Zimek, A., 2008, August. Angle-based outlier detection in high-dimensional data. In KDD '08, pp. 444-452. ACM.

[8] Y. Zhao and M.K. Hryniewicki, "XGBOD: Improving Supervised Outlier Detection with Unsupervised Representation Learning," IEEE International Joint Conference on Neural Networks, 2018.

[9] Y. Zhao and M.K. Hryniewicki, "DCSO: Dynamic Combination of Detector Scores for Outlier Ensembles," ACM SIGKDD Workshop on Outlier Detection De-constructed, 2018. Submitted, under review.