Mining people's semantic trajectory behaviours from geotagged photographs

Web 2.0 technology has changed the way users use the web. In Web 2.0, users can create their own data and upload to the web. The new web technology promotes the evolution of social media applications built on Web 2.0. Social media allow the creation and exchange of user-generated content. One main type of social media is photo-sharing websites where the main objective is the sharing of photo media content between users. Through these websites users store and manage their photos, and share and communicate with friends, families and colleagues. The development of mobile devices facilitates people's easy usage of the Internet. This has lead to dramatic growth in the number of user-generated photos in the website. This massive collection of photo data may enclose people's movement behaviours, which are useful to domain experts and areas such as traffic management and tourism. However, this large and complex dataset requires advanced techniques to extract the hidden useful knowledge from the big data. Some previous studies have been conducted, and various approaches have been proposed to extract people's movement behaviours from online geotagged photos. These studies are mainly about three topics. The first topic is to reveal the spatial behaviours of people that the approaches detect the spatial locations that people prefer to visit (Kisilevich, Mansmann, and Keim 2010; Lee, Cai, and Lee 2013). The second topic is to find out the spatial place association rules that determine the sets of places visited together in people's movements. The third topic is to discover people's dynamic spatiotemporal movement behaviours including the spatio-temporal traffic ow (Girardin et al. 2008b) and frequent spatio-temporal movement patterns (Zheng, Zha, and Chua 2012; Cai et al. 2014; Bermingham and Lee 2014). However, previous approaches lack consideration of the additional aspatial semantics information of trajectories. They are traditional geometric-feature analyses. The main drawback of previous methods is that their result patterns contain only pure geometric data, without meaningful semantics information about the mobility. Most applications analyses require complementing trajectory with additional information from the application context. The contextual information provides useful knowledge about moving behaviours with richer and more meaningful semantic information and the semantic-level patterns. This thesis aims to develop a systematic framework for extraction of people's movement patterns with meaningful and understandable semantics information. We add the aspatial semantics annotations to trajectories and analyse trajectories with spatial, temporal and aspatial features together. We aim to find the semantics-enhanced movement patterns, including semantic sequential patterns, semantic common patterns and semantic trajectory patterns. Finally, this thesis also aims to build an itinerary recommender system based on the extracted trajectory patterns. In this thesis, we propose a systematic framework for discovery of people's semantic mobility patterns from geotagged photos. The framework has four main functions for extraction of the three semantic patterns and for building the recommender system, respectively. At the first step, the framework builds spatio-temporal trajectory data from the geotagged photos. Then, we add background geographic information, place type annotation and multiple environmental contextual data to the raw trajectories to generate people's semantic trajectories. From the semantic trajectories, the framework's first main function is to find out the frequent semantic sequential patterns. This thesis proposes a sequential pattern mining method to extract semantic sequential patterns, which are sequences of stops that frequently occur in people's trajectories. This method can deal with multi-dimensional semantic trajectories. The extracted groups of patterns include not only the basic patterns, which contain geographic place category information only, but also the multi-dimensional semantic patterns, which are associated with flexible combinations of frequent environmental contextual information. The framework's second main function is to reveal the semantic common patterns. This thesis proposes a semantic trajectory clustering approach to find semantic common patterns in the semantic trajectories. The common pattern shows the common track drawn from many people having similar trajectories. A distance function is designed and proposed for the multidimensional semantic trajectories. The third main function of the framework is to extract the semantic trajectory patterns. This thesis presents a semantic trajectory pattern mining method to find frequent trajectory patterns from semantic trajectories. A semantic trajectory pattern demonstrates a frequently visited sequence of stops with typical transition time information. The transition time shows the time interval between two stops that indicates temporal behaviour of people's mobility. Finally, this framework builds a recommendation system based on the extracted semantically enhanced movement patterns. The system provides users with suggestions about travel itineraries including travel route and time interval information between two stops. The system is semantic-aware, allowing users to customise sets of place types that they want to visit in the trip and to set up travel duration. We conduct experiments to evaluate proposed methods using real photo dataset from Flickr1. The experimental results prove the effectiveness of our framework. The results show that the proposed semantics added trajectory analysis methods can extract detailed and semantically enhanced semantic patterns that not only show people's semantic-level mobility patterns, but also provide rich meaningful information and better understanding of people's movements. The results also demonstrate that our recommender system effectively generates a set of customised and targeted semantic-level itineraries that meet the user-specified constraints and with an efficiency itinerary generation property. In addition, our system produces higher place type-layer itineraries with richer meaningful information about travel contexts