Unsupervised Machine Learning for Developing Personalised Behaviour Models Using Activity Data

Abstract

The goal of this study is to address two major issues that undermine the large scale deployment of smart home sensing solutions in people’s homes. These include the costs associated with having to install and maintain a large number of sensors, and the pragmatics of annotating numerous sensor data streams for activity classification. Our aim was therefore to propose a method to describe individual users’ behavioural patterns starting from unannotated data analysis of a minimal number of sensors and a ”blind” approach for activity recognition. The methodology included processing and analysing sensor data from 17 older adults living in community-based housing to extract activity information at different times of the day. The findings illustrate that 55 days of sensor data from a sensor configuration comprising three sensors, and extracting appropriate features including a “busyness” measure, are adequate to build robust models which can be used for clustering individuals based on their behaviour patterns with a high degree of accuracy (>85%). The obtained clusters can be used to describe individual behaviour over different times of the day. This approach suggests a scalable solution to support optimising the personalisation of care by utilising low-cost sensing and analysis. This approach could be used to track a person’s needs over time and fine-tune their care plan on an ongoing basis in a cost-effective manner.

Big Data for Energy Management and Energy-Efficient Buildings

“…The research of Lapalu et al focused on the unsupervised mining of activities for smart home prediction [38]. Unsupervised ML algorithms for developing personalized behavior models using activity data have been presented [39]. dynamic programming, Monte Carlo methods and temporal difference methods can be exploited for solving a reinforcement learning problem.…”

Tracking changes in user activity from unlabelled smart home sensor data using unsupervised learning methods

“…In order to perform HAR, periods of sensor data events that may represent activities must be extracted first. Traditional approaches for this include the use of sliding time and sensor windows as used by Yala et al, Cook and Krishnan [7,8]. These sliding windows are generally used for training supervised learning systems when activity labels are present, as the sliding windows can be chosen based on the activity labels present in the data, and windows containing noise can be removed manually.…”

“…However, it must be noted that even though both t-SNE and UMAP are both useful choices for visualisation, clustering based on their output is generally not recommended, as density information is often lost during the process [16]. A useful technique is also presented by Fiorini et al [8], where radar graphs were constructed from motion sensor data which can be used to facilitate a quick visual review of the sensor data. This technique can be used in conjunction with other visualisation techniques such as UMAP, to gain further insight into the sensor data.…”

“…In order to visualise the behavioural changes by day-of-theweek, UMAP was performed to generate data points in a two-dimensional space from the eight-dimensional feature sets of the temporal clusters. It was hypothesised that using a day-of-the-week level of granularity might help to better track changes in the longer term, because as shown in Fiorini et al [8], there can be marked differences between weekday and weekend routines.…”

“…However, most of the existing research studies that have shown promising results used context-rich information obtained from BSNs and video-based solutions, and not WSNs. Researchers such as Fiorini et al [8] used unsupervised learning with WSNs; however, in this case the authors were looking for an overall user ''busyness'' metric rather than individual user activity patterns.…”