An approach to evaluating motion pattern detection techniques in spatio-temporal data

Abstract

This paper presents a method to evaluate a geographic knowledge discovery approach for exploring the motion of point objects. The goal is to provide a means of considering the significance of motion patterns, described through their interestingness. We use Monte-Carlo simulations of constrained random walks to generate populations of synthetic lifelines, using the statistical properties of real observational data as constraints. Pattern occurrence in the synthetic data is then compared with observational data to assess the potential interestingness of the found patterns. We use motion data from wildlife biology and spatialisation in political science for the evaluation. The results of the numerical experiments show that the interestingness of found motion patterns is largely dependant on the configuration of the pattern matching process, which includes the pattern extent, the temporal granularity, and the classification schema used for the motion attributes azimuth and speed. The results of the numerical experiments allow interestingness to be attached only to some of the patterns found—other patterns were suggested to be not interesting. The evaluation method helps in estimating useful configurations of the pattern detection process. This work emphasises the need to further investigate the statistical aspects of the problem under study in (geographic) knowledge discovery.

Introduction

Location aware devices are becoming ubiquitous and will increase our capability to collect spatio-temporal motion data by many orders of magnitude. The ubiquity of such devices is reflected by the fact that in the summer of 2004 the Japanese Telecommunications Council declared that all mobile phones introduced in Japan after 2007 should have self-locating functionality. Studies of so-called moving point objects (MPO), incorporating information about changing positions of discrete objects in time and space, have been identified as a key emerging research area in GIScience (Miller, 2003). By studying MPOs through time, individual geospatial lifelines can be derived from large datasets collected, for example, from people carrying GPS-enabled phones and PDAs (e.g., Dykes & Mountain, 2003), tracked animals in field studies (e.g., Wentz, Campell, & Houston, 2003), or even tracked football players in sports scene analysis (e.g., Iwase & Saito, 2003).

It has been recognised that not only is spatial data special, but also the handling and analysing of spatio-temporal data, and above all motion data, requires the development of new concepts (Frank, 2001, Mark, 2003). Traditional analytical methods for spatial and spatio-temporal data were developed in an era when data collection was expensive and computational power was weak (Miller & Han, 2001). Miller and Han thus reason that “traditional spatial analytical techniques cannot easily discover new and unexpected patterns, trends and relationships that can be hidden deep within very large and diverse geographic datasets” (Miller & Han, 2001, p. 3).

The integration of knowledge from the field of GIScience about space–time together with the emerging field of knowledge discovery in databases opens up the possibility for geographic knowledge discovery (GKD). Applications which generate large volumes of spatio-temporal data, such as high-resolution (in time and space) satellite-based systems (Griffiths & Mather, 2000), and in our case the analysis of geospatial lifelines (Hornsby and Egenhofer, 2002, Mark, 1998) face multiple challenges in the storage and exploration of high-volume spatio-temporal datasets and thus present excellent cases for the application of GKD.

It has been recognized in the knowledge discovery in databases (KDD) literature that discovery systems can generate a glut of patterns, most of which are of no interest to the user (Silberschatz and Tuzhilin, 1996, Padmanabhan, 2004). Thus, it is recommended that data mining be carried out with regard to the statistical aspects of the problem (Fayyad, Piatetsky-Shapiro, & Smyth, 1996).

In this paper work is presented which extends the relative motion (REMO) GKD approach developed to identify motion patterns in groups of MPOs (Laube and Imfeld, 2002, Laube et al., 2005, Laube et al., 2004). The REMO approach proposed so far (Section 2.2) allows the user to search large volumes of data for instances of pre-defined motion patterns, which are constructed on the basis of existing knowledge about the motion of the objects under study. However, the user has no means by which to estimate the significance (e.g., the uniqueness) of the extracted patterns. The number and the extent of patterns found may depend significantly on both the motion data and the parameterisation of the pattern detection process. Many more patterns may be identified in a space where motion is constrained, for example, on a football pitch, than in the seemingly chaotic motion of children in a playground. A central question for GKD is therefore, how can we assess the significance of patterns extracted from such cases? The central issue of this paper is therefore to provide a means of considering the significance of REMO patterns.

Our approach focuses on the use of Monte-Carlo simulations to generate synthetic lifelines constrained by the statistical properties of real observational data. Pattern occurrence in the simulated data is then compared with observational data to assess the potential significance of the patterns. Finally, having identified potentially interesting patterns we return to the observational data to investigate whether these patterns have meaning in terms of the system under investigation.

The paper is structured as follows. Section 2 provides a literature overview on geographic knowledge discovery in general, the REMO GKD approach, pattern interestingness, and on different potential approaches to simulating geospatial lifelines. In Section 3 the central ideas of our evaluation approach are introduced. The data used in the study are introduced in Section 4. Section 5 describes the methodology for producing constrained random walks to produce synthetic lifelines, and their use in Monte-Carlo simulation for numerical experiments. The results of these experiments are presented in Section 6, and their meaning and application to the observational data is discussed in Section 7 before examining the general applicability of these results to the field of GKD.