Unsupervised online activity discovery using temporal behaviour assumption

Abstract

We present a novel unsupervised approach, UnADevs, for discovering activity clusters corresponding to periodic and stationary activities in streaming sensor data. Such activities usually last for some time, which is exploited by our method; it includes mechanisms to regulate sensitivity to brief outliers and can discover multiple clusters overlapping in time to better deal with deviations from nominal behaviour. The method was evaluated on two activity datasets containing large number of activities (14 and 33 respectively) against online agglomerative clustering and DBSCAN. In a multi-criteria evaluation, our approach achieved significantly better performance on majority of the measures, with the advantages that: (i) it does not require to specify the number of clusters beforehand (it is open ended); (ii) it is online and can find clusters in real time; (iii) it has constant time complexity; (iv) and it is memory efficient as it does not keep the data samples in memory. Overall, it has managed to discover 616 of the total 717 activities. Because it discovers clusters of activities in real time, it is ideal to work alongside an active learning system.

Discovery and Recognition of Emerging Human Activities Using a Hierarchical Mixture of Directional Statistical Models

“…Similarly, Gjoreski et al have used an agglomerative clustering technique to enable real-time clustering of streaming data [5]. To validate clusters for potential new activities, they have proposed two temporal assumptions on human activities; that is, a human activity usually lasts for a certain period of time and there should not be frequent transitions between activities.…”

“…where n K+1 is the size of the data instances belonging to the new activity K + 1 and N = K a=1 n a denotes the initial training data size. The extended vector is just the renormalised proportions of the activity instances based on (5). Note that to make the update possible, we need to keep the training data size N as an extra parameter.…”

“…Note that the activity proportion φ for the hierarchical mixture needs to be updated as well at the end as the individual activity count is changed. One just needs to re-estimate φ according to (5) with the updated n a . The above two update procedures are similar as they both require fitting MCIvMFs to the newly discovered data.…”

Lifelong Learning in Sensor-Based Human Activity Recognition

“…Similarly, Gjoreski et al have used an agglomerative clustering technique to enable real-time clustering of streaming data [10]. To validate clusters for potential new activities, they have proposed two temporal assumptions on human activities; that is, a human activity usually lasts for a certain period of time and there should not be frequent transitions between activities.…”

Collecting Labels for Rare Anomalies via Direct Human Feedback—An Industrial Application Study

“…In Reference [88], the authors proposed a technique for online activity discovery based on clustering assumptions of labels in successive signal windows. Although their approach is memory efficient and has constant time complexity, it is not applicable in our scenario due to the fact that reoccurring activities have lead each time to a newly created cluster segment with the methods introduced in Reference [88]. This does not allow to model normal behavior as a single class in reoccurring cluster segments and to distinguish it from other, abnormal signal classes.…”