Enabling the full potential of Machine Learning

When large amounts of data is collected and analysed through Machine Learning algorithms, useful information and knowledge can be generated. One example of such knowledge consists of detecting rare events and anomalies, for example deviations from any kind of normal behaviour or faulty equipment. Another example is the possibility to identify data clusters over different time scales and precision in multivariate time-series data. Such clusters can for example represent the operational behaviour of vehicles and machines, or human activities in a smart home. Extracting such high-level knowledge is essential to fully understand the operation of complex environments. But the techonology comes with several challenges, which constitute the core of the research conducted by Mohamed-Rafik Bouguelia, Associate Professor and Docent in Machine Learning at Halmstad University, Center for Applied Intelligent Systems Research (CAISR).

“Machine Learning is useful for a wide range of application domains. However, it still lacks the generalisability associated with human learning. The field would have a much broader impact if it could enable machines to easily interact with humans, adapt to changing contexts, and transfer knowledge between tasks”, says Mohamed-Rafik Bouguelia.

Algorithms that adapt to change

The research of Mohamed-Rafik Bouguelia is conducted in collaboration with several Swedish industrial partners, for example Volvo Group, Alfa Laval and Öresundskraft. One of the research projects focuses on developing ’adaptive’ Machine Learning algorithms that can closely track and react to changes in the data distribution.

“Today, there are innumerous complex machines equipped with various sensors that collect data continuously over time. To detect anomalies or symptoms that indicate imminent failures, we design algorithms that can learn what constitutes a normal or abnormal behaviour from this data. However, such data streams are subject to change over time due to unforeseen external conditions. That is why we are working on algorithms that can adapt to data changes”, says Mohamed-Rafik Bouguelia.

Human feedback is essential

Machine Learning algorithms can be used in many different areas, for example vehicles, trains, heat pumps and district heating. When monitoring district heating substations, for example, the data that is collected sometimes show atypical events that happen, but they are not necessarily anomalies. Anomaly detection is often subjective because what the user considers to be a ’relevant’ anomaly depends on the purpose of the application. Furthermore, real-world data may contain different plausible groupings, and a fully unsupervised clustering can not establish a grouping that suits the user’s needs because this requires external domain knowledge. Creating labelled data to train fully supervised machine learning algorithms is often costly and time-consuming, and such expert knowledge may not be available beforehand.

“To address these problems, we are developing interactive Machine Learning and Data Mining algorithms that include the human within the (machine) learning loop. For example, to distinguish between relevant and irrelevant anomalies, one can use algorithms that proactively communicate with a human expert to leverage her or his feedback and learn to suggest more relevant anomalies”, says Mohamed-Rafik Bouguelia.

The objective is not only to produce a more accurate anomaly detector, but also to minimise the expert effort required to investigate these anomalies. Other research directions along these lines include leveraging the human feedback to define most typical behaviours among a fleet of diverse systems, or to learn features and data representations that reveal faults and anomalies that are most relevant to the end-user.