Semi-Supervised and Explainable Machine Learning with an Application to the Low-Voltage Grid

Due to the enormous increase in available data and the computational power needed to process it, interest in the use of machine learning (ML) keeps increasing as it is able to extract valuable insights from this data. In an effort to make it easier to use ML in practice, this thesis proposes several novel ML techniques driven by two industry needs: (1) to make machine learning systems understand what the user wants and (2) to make the user understand what machine learning systems say or predict. Moreover, we also demonstrate the value of ML on a concrete contemporary use case: the reinforcement of the low-voltage grid.

In standard unsupervised clustering, it is difficult to tell the algorithm what the user expects. There are often many possible clusterings for the same set of examples. Therefore, the user has to experiment with different algorithms, similarity metrics and hyperparameter configurations to obtain the desired clustering. Active semi-supervised clustering offers an alternative: the algorithm gathers constraints by asking the user questions and uses those constraints to produce a clustering that matches the user's expectation. However, handling noise remains a challenge: if some constraints are incorrect, existing algorithms often fail to produce a satisfactory result. To address this limitation, our first contribution is a framework that makes most constraint-based clustering algorithm robust to noise. By reasoning about noise probabilistically, the framework is able to detect and correct the noisy constraints such that the clustering algorithm can produce a high-quality clustering even if some of the constraints provided by the user are incorrect.

Next to clustering, we also looked into anomaly detection (AD). Many AD algorithms only identify anomalies, they do not explain why they label a particular instance as anomalous. This makes it difficult to take action without further analysis: maybe the identified anomaly exhibits rare but completely normal behavior, e.g., maintenance works. If an anomaly detector can explain its predictions, it is immediately clear why the algorithm thinks a particular instance is anomalous. These explanations enable users to verify the algorithm's reasoning and make it easier to act upon the results. As our second contribution, we provide an extensive overview of dependency-based AD, an anomaly detection strategy that is explainable. The main idea is to model the dependencies between attributes and mark instances that violate these dependencies as anomalous. These dependencies and violations can serve directly as explanations: e.g., ``Animals with a tail usually have a backbone; however, a scorpion has a tail but no backbone''. We compare existing dependency-based detectors and highlight the strengths of every method. Moreover, we identify two weaknesses of existing models and propose two novel ideas to solve them. Although further research is still required, we believe the presented insights are very relevant for future developments in the field of dependency-based AD.

Besides, the AD literature does not agree on how to set the hyperparameters of different anomaly detectors when experimentally comparing algorithm. Most papers compare either performance using ``default'' settings, or maximal performance under optimal settings. These methodologies are either too pessimistic, assuming no tuning, or unrealistically optimistic, assuming optimal tuning; and often result in methodologically unsound and irreproducible comparisons between algorithms. As our third contribution, we therefore propose to use a small validation set to tune an anomaly detector's hyperparameters to achieve a realistic, fair, sound, and reproducible comparison.

Finally, in our last contribution, we illustrate the value of ML by applying it to help plan a cost-efficient reinforcement of the low voltage grid (LVG). To plan such a reinforcement, one needs to calculate the typical currents and voltages throughout the grid. This calculation requires the electricity consumption time series of each connected consumer as an input. However, for many consumers, these time series are unknown and need to be estimated from the available information. Therefore, we developed two methods to predict the electricity consumption time series of a particular household given consumer, weather and calendar information. The first method solves the task but is complex and slow with limited interpretability. Therefore, we also developed a second method that uses predictive clustering. This second model is significantly simpler with similar predictive performance as our first approach. Moreover, it is fast and interpretable, allowing a domain expert to inspect which time series are predicted in which circumstances. We hope our contribution will help to reduce the cost of the reinforcement of the LVG while also giving distribution grid operators more insight into their data.