Main current DaSCI research lines:
- Interpretability and explainability: One of the most prominent demands from companies is the ability to understand the model acting as an anomaly detector. This goal implies two complementary objectives: Firstly, improving model interpretability; and secondly, providing insight into the model’s functioning, making it explainable.
- Deep learning for complex anomalies: DL is a powerful tool for anomaly detection, applied in two scenarios: static data (using autoencoders) and time series (using convolutional and recurrent neural networks). GANs have also gained popularity in anomaly detection due to their effectiveness with limited labelled data.
- False positive mitigation: This task aims at reducing the number of false positives labelled by the anomaly detector.
- Big Data scenarios and model fusion in distributed environments: With the explosion of data, popularisation of sensors, and automation in data acquisition and storage, the anomaly detection problem has become a Big Data problem, where the fusion of information/models will strongly impact the quality of the final system.
