Although the availability of sensor data is becoming prevalent across many domains, it still remains a challenge to make sense of the sensor data in an efficient and effective manner in order to provide users with relevant services. The concept of virtual sensors provides a step towards this goal, however they are often used to denote homogeneous types of data, generally retrieved from a predetermined group of sensors. The DIVS (Dynamic Intelligent Virtual Sensors) concept was introduced in previous work to extend and generalize the notion of a virtual sensor to a dynamic setting with heterogenous sensors. This paper introduces a refined version of the DIVS concept by integrating an interactive machine learning mechanism, which enables the system to take input from both the user and the physical world. The paper empirically validates some of the properties of the DIVS concept. In particular, we are concerned with the distribution of different budget allocations for labelled data, as well as proactive labelling user strategies. We report on results suggesting that a relatively good accuracy can be achieved despite a limited budget in an environment with dynamic sensor availability, while proactive labeling ensures further improvements in performance.

The advances in Internet of things lead to an increased number of devices generating and streaming data. These devices can be useful data sources for activity recognition by using machine learning. However, the set of available sensors may vary over time, e.g. due to mobility of the sensors and technical failures. Since the machine learning model uses the data streams from the sensors as input, it must be able to handle a varying number of input variables, i.e. that the feature space might change over time. Moreover, the labelled data necessary for the training is often costly to acquire. In active learning, the model is given a budget for requesting labels from an oracle, and aims to maximize accuracy by careful selection of what data instances to label. It is generally assumed that the role of the oracle only is to respond to queries and that it will always do so. In many real-world scenarios however, the oracle is a human user and the assumptions are simplifications that might not give a proper depiction of the setting. In this work we investigate different interactive machine learning strategies, out of which active learning is one, which explore the effects of an oracle that can be more proactive and factors that might influence a user to provide or withhold labels. We implement five interactive machine learning strategies as well as hybrid versions of them and evaluate them on two datasets. The results show that a more proactive user can improve the performance, especially when the user is influenced by the accuracy of earlier predictions. The experiments also highlight challenges related to evaluating performance when the set of classes is changing over time.

In interactive machine learning, human users and learning algorithms work together in order to solve challenging learning problems, e.g. with limited or no annotated data or trust issues. As annotating data can be costly, it is important to minimize the amount of annotated data needed for training while still getting a high classification accuracy. This is done by attempting to select the most informative data instances for training, where the amount of instances is limited by a labelling budget. In an online learning setting, the decision of whether or not to select an instance for labelling has to be done on-the-fly, as the data arrives in a sequential order and is only valid for a limited time period. We present a taxonomy of interactive online machine learning strategies. An interactive learning strategy determines which instances to label in an unlabelled dataset. In the taxonomy we differentiate between interactive learning strategies when the computer controls the learning process (active learning) and those when human users control the learning process (machine teaching). We then make a distinction between what triggers the learning: active learning could be triggered by uncertainty, time, or randomly, whereas machine teaching could be triggered by errors, state changes, time, or factors related to the user. We also illustrate the taxonomy by implementing versions of the different strategies and performing experiments on a benchmark dataset as well as on a synthetically generated dataset. The results show that the choice of interactive learning strategy affects performance, especially in the beginning of the online learning process, when there is a limited amount of labelled data.

In interactive machine learning it is important to select the most informative data instances to label in order to minimize the effort of the human user. There are basically two categories of interactive machine learning. In the first category, active learning, it is the computational learner that selects which data to be labelled by the human user, whereas in the second one, machine teaching, the selection is done by the human teacher. It is often assumed that the human user is a perfect oracle, i.e., a label will always be provided in accordance with the interactive learning strategy and that this label will always be correct. In real-world scenarios however, these assumptions typically do not hold. In this work, we investigate how the reliability of the user providing labels affects the performance of online machine learning. Specifically, we study reluctance, i.e., to what extent the user does not provide labels in accordance with the strategy, and fallibility, i.e., to what extent the provided labels are incorrect. We show results of experiments on a benchmark dataset as well as a synthetically created dataset. By varying the degree of reluctance and fallibility of the user, the robustness of the different interactive learning strategies and machine learning algorithms is explored. The experiments show that there is a varying robustness of the strategies and algorithms. Moreover, certain machine learning algorithms are more robust towards reluctance compared to fallibility, while the opposite is true for others