Processing request, please stand by

METEOR | Dynamic Bayesian Outlier Detection

METEOR - Dynamic Bayesian Outlier Detection

Delightfully Crafted by jorgeserras

Learn More Watch Video

Source Code

Multivariate Time Series

With the numerous amount of applications and demand for data analysis, a full outlier detection mechanism focused on longitudinal data is built with the aim of providing knowledge regarding data's underlying nature.

Learn More

Score Analysis

After training a DBN, each instance is scored using a sliding window approach. Score distributions are then scrutinized providing a clear boundary for anomaly classification.

Learn More


SAX discretization and dimensionality reduction is available for continuos data prior to the modeling phase.

DBN modeling

A flexible tree-augmented DBN modeling procedure with easily adapted parameters.


After performing anomaly detection, output data sets can be downloaded free of charge!

What is METEOR?

MultivariatE Time sEries OutlieR (METEOR) is a complete anomaly detection system ranging from pre-processing to post-score analysis using dynamic Bayesian networks (DBN)


Data is processed ensuring the system's assumptions including data visualization

DBN Modeling

Optimal tree-augmented DBN modeling through a straightforward user interface

Model Download

DBN structures can be downloaded and changed at any time through the whole procedure

Transition Outliers

According to the modeled DBN, data set windows are scored and analyzed

MTS Outliers

Complete Multivariate Time Series are scored and evaluated

Score Analysis

Score analysis strategies are studied and easily adapted through the user interface

Dynamic Bayesian Outlier Detection

Horizontal/Panel Data Formatting

Outlier Detection

Input Data

Pre-processing: SAX

PAA controls the dimensionality reduction. The value indicates the new length for each discretized time series. If no dimensionality reduction is desired, pick the maximum value (current length).

For most datasets, the optimal value (minimal information loss) is in the range from 5 to 8.

DBN Modeling

If unchecked, each time slice is modeled independently. Non-stationary modeling is heavy and can cause overfitting. The user is adviced to avoid it if possible.

Markov lag indicates the DBN order. A value in the range from 1 to 3 is adviced and adequate for most datasets.

The number of preceding parents should always be 1 or 2 unless the user knows that the dataset would benefit from larger values.

Outlier Detection