Abstract:
Anomaly Detection (AD) in time series applications is vital given the wide usage of
time series data in the real-world setting. Usually, real-world anomalies are found in a
variety of forms making them more difficult to detect. Various anomaly detection
models were considered in research, where each of these models tends to apply a
specific assumption for detecting anomalies revealing more sensitivity towards a
specific kind of anomalies. Hence, it is required to build a model selection for different
AD models that is capable of choosing, at each time step, the suitable anomaly detection
model. Such a model selection would be able to take advantage of several AD models in
an attempt to maximize the accuracy of time series anomaly detection.
Knowing that ground truth labels in real-world applications are expensive and rare to
obtain in large quantities, it is important to design a model selection for AD that does
not rely explicitly on ground truth labels. In this regard, a model selection for six
unsupervised AD models is proposed based on Time-Series Forest and Reinforcement
Learning to choose dynamically an AD technique without relying heavily on ground
truth labels. Experiments on two synthetic and real time series datasets have been
applied. Results on the real dataset reveal that the proposed model selection outperforms
the entire AD models in terms of the F1score metric. In the synthetic dataset, the
proposed model, recording an F1score of 0.989, outperforms all the AD models except
for KNN. Different reward function scenarios, including adaptive rewards with multiple
exploration rates, have been also considered. The proposed AD model selection
approach with its original reward function yields the best overall scores. Each of the six
AD models implemented on three additional synthetic datasets having global, local, and
clustered anomalies respectively, show distinct performance depending on the type of
anomalies. This proves the significance of the proposed AD model selection framework
which maintains a high F1score on the three datasets.
