Anomaly detection library
This anomaly detection library spins around gaussian/normal distribution. In this distribution, most examples revolves around some mean value and probability of some value drops the futher away it goes from mean. Therefore, it is possible to compute probabilty of each example in some dataset and label examples, which probabilities are less then some threshold as anomalies.
Basic structure of working with such algorithm is:
- Collect training data (it should mostly be negative, non-anomaly examples)
- Learn mu and sigma out of each feature in training set, provided it is normally distributed
- (optional, but recommended) Optimize threshold on cross-validation set, which should contain anomalies (around 50%). Currently, only F1 score is optimized.
- With learned threshold, it is now possible to make predictions about new example
# training set should an MxN matrix, where M is a number of examples, and N is a number of features. #It should all be real-valued numbers, and each feature should be normally distributed in order for algorithm to work well. #It should also consist of only non-anomalies, positive examples. training_set = [[1,2,3,..], [3,2,4,..], [4,2,1,..], ...] @magnifier = Magnifier.new(@training_set) @magnifier.train @magnifier.anomaly?([44, 55,..]) # => true # default threshold is 0.01, but it is possible to train it automatically. It will require 2 arrays: # first array is set of positive and negative examples (could be around 40-60% of training set) cross_validation_set = [[1,2,3,..], [5,1,6,..], ...] # second array consists of labels for the first one (true of 1 means anomaly, false or 0 means non-anomaly), and indexes should match. # E.g., if first example is anomaly, and second is not, then: # cross_validation_labels == true and cross_validation_labels == false cross_validation_labels = [true, false, ...] @magnifier.optimize_threshold(cross_validation_set, cross_validation_labels) @magnifier.threshold # => something with best f1 score
# you can easily export existing Magnifier object into a .yaml file. # in order to do this, you have to specify a path(or file) to save content to @magnifier.export("export.yaml") # => #<File> # method returs a file, so you can modify it later # it is also possible to do an export using separate class: Magnifier::Exporter.export("export.yaml", @magnifier) #import is done in the same manner: Magnifier::Importer.import("export.yaml") # => magnifier object #it is also possible to import inside current object, hovewer, all content will rewritten from file @magnifier.import("export.yaml")
I completed a couse in machine learning and had some free time to apply it.
Refer to todo.md