Provide algorithmic foundation for no-config and low maintenance machine learning applications consumable in a distributed and real time software intelligence platform.
The application of machine learning and AI in a large-scale software intelligence platform, requires the methods to be used to be scalable (in the number of models to be handled), explainable, diagnosable, able to deal with uncertainty and easily maintainable. We research methods for forecasting, anomaly detection, outlier detection for time series, event streams, and log data for various application areas in software intelligence (e.g. cloud-native security), which fulfil the earlier mentioned requirements.
Many existing applications of machine learning and AI solve one problem with one purpose-built model. The situation encountered in large-scale SaaS based multi-tenant software intelligence solutions like Dynatrace is somewhat different: we have to deal with a large number - think millions - of small to medium sized (machine learning) problems, which typically have to be tackled by individual models because the tasks are very specific and data privacy regulations have to be taken into account. Hence, all the large-scale training, updating and model maintenance related tasks must be fully automated. This requires research on learning algorithms, which for a particular class of problems (e.g. time series forecasting) are
- scalable: a small set of computationally efficient algorithms can deal with a large heterogeneity of tasks. One approach we are currently investigating are parameter free (in the sense of no strong assumption on the underlying data distribution) models.
- diagnosable: success and non-success of any model training and update must be easily captured by interpretable metrics
- explainable: the decision within the software intelligence platform derived from the model output can be understood by the domain experts
- able to deal with uncertainty: we research methods which are inspired by Bayesian approaches which give approximate probabilistic predictions but are less computationally expensive
- maintainable and observable: model- retraining and updates shall be triggered by data driven assessment of the model’s suitability to the current inbound stream of data. This research stream is very much related to the detection model drifts and predicting unknown states.