Hierarchical Cooperative Discovery of Personal Places from Location Traces

Opportunistic Discovery of Personal Places Using Multi-Source Sensor Data

“…(1) approaches that rely entirely on GPS data for localization [1], [5], [6], [21], [22], [23] and (2) approaches that use other sources of data, such as Wi-Fi routers, cell towers, or Bluetooth beacons in addition to GPS data [3], [8], [9], [12], [24], [25].…”

“…Additionally, Wi-Fi has a relatively large transmission range, which makes Wi-Fi-based approaches less reliable when deciding if a user remained at the same place or not. In [9], [12], we introduced a new approach, where we first use a hierarchy of 1-hop and 2-hop Bluetooth fingerprint matching techniques, followed by Wi-Fi fingerprint matching, in order to discover a group of co-located users of a particular user. Then we borrow the location points from the co-located users to fill the gaps of the target user.…”

“…In order to obtain such POIs, a user's collected location traces can be grouped into different clusters (each representing a POI), using various spatio-temporal clustering approaches [5], [8]. However, previous work has shown that on average about 40-70% of the time the collected traces will suffer from extensive gaps [5], [9], often because of users' presence in GPS-denied areas (e.g., indoors) or other situations that prevent data collection (e.g., a smartphone placed into a low-power state). Spatio-temporal clustering on such incomplete traces will often yield a large number of small clusters, where these small clusters should actually be grouped together to form a large cluster.…”

“…This can be achieved by carefully augmenting the location traces to reduce or completely remove these data gaps, e.g., by using other sources of information to determine if two small clusters should be merged together or not. Researchers have proposed various methods for merging several small clusters into fewer larger clusters using various types of "secondary data", e.g., using techniques such as fingerprint matching [10], [11] or location data sharing among colocated users [9], [12]. However, these approaches have various limitations, e.g., fingerprint maps may be incomplete or outdated and traces of co-located users (for location sharing) may not always be available.…”

“…Compared to our previously proposed approach [13], in this paper we also use physical activity data (i.e., step counts) due to its wider availability compared to sleep and battery charging data, which have a skewed spatio-temporal distribution (as shown in [13]). For a given user, we first use the location traces of the user to determine a series of stay point (sp) clusters using our previously proposed Enhanced Stay Point Clustering approach (described in [9]). We then extract the sleep and battery charging status as well as the step count data from the same user during gaps between consecutive sp-clusters and perform a cluster merging technique referred to as Opportunistic Stay Point Clustering, which is fully described in Section 3.3 of this paper.…”