Add a new model, namely AER, an AutoEncoder Regressor model.
Description
Prediction-based and reconstruction-based anomaly scoring methods each have their own success and limitations. AER is a model that can predict both prediction-based and reconstruction-based anomaly scores and combine those scores to leverage the success of both scoring methods. Assuming the input length is n, the output is n+2; hence we predict the next values at the start and the end while reconstructing the input in the middle.
The ideas introduced in this model are:
Masking anomaly scores: Exponential weighted moving average function used to smooth the anomaly score produces unwanted false positives at the beginning. Masking the anomaly scores with the minimum of the anomaly scores reduces false positives.
Bi-directional Regression: Prediction-based models require at least n observations before making the first prediction at the n+1 index. If n is large, there are many false negatives at the start of the anomalies. If we perform regression in both directions (forward f and reverse r), we can fill in the gap at the beginning of the anomaly scores. Overlapping scores in the middle are averaged.
RetroSearch is an open source project built by @garambo | Open a GitHub Issue
Search and Browse the WWW like it's 1997 | Search results from DuckDuckGo
HTML:
3.2
| Encoding:
UTF-8
| Version:
0.7.4
