Efficient and effective trip planning: a data-driven approach

<p>With the increasing pervasiveness of intelligent transportation, a series of innovative services are developed to facilitate people's daily life. This has enabled the advent of a location-based service: trip planning, which helps plan and manage users' itineraries. Existing literatures commonly focus on a small phase, such as making effective trip destinations recommendation during pre-trip phase, or finding the most suitable travel routes during in-trip phase. However, the combination of three important phases (i.e., pre-trip, in-trip and post-trip) is rarely considered to benefit the wellbeings of both the system and end-users.</p> <p>In this thesis, we will stand at the viewpoint of a trip planning service provider by considering the practical requirements. Our overarching aim is to develop a cohesive, efficient and effective trip planning workflow from the perspective of three consecutive phases: pre-trip, in-trip, and post-trip. In particular, the following problems are investigated: (1) pre-trip recommending top-k POIs for users; (2) in-trip dispatching drivers under peak demand; (3) in-trip finding the pick-up location with streaming users; (4) post-trip identifying top-k traffic bottlenecks on road network.</p> <p>Before the trip, users are willing to be recommended several related POIs as the trip destinations. This problem will become more challenging when users explore the map to find attractions by frequent interactive operations like zoom in/out. Since users can be located in different zoom levels in the process of data exploration, they may have different expectations on the recommended POIs with varying spatial granularities. For example, one user may want to go to a region like Melbourne CBD, while another user may expect to be recommended a particular POI like a restaurant. In this thesis, we make effective POI recommendation in a multi-level manner, where each level corresponds to a zoom level in a map or a particular spatial granularity, and we return top-k POIs from each level in a unified framework.</p> <p>In the trip, the system is in charge of dispatching suitable drivers to serve users. However, peak demand in rush hour period may bring huge burden to the system, namely the user demand is much greater than the number of available drivers. We take a special focus on an intelligent transportation mode: dynamic ridesharing, where users with similar itineraries and time schedules can share the trip together to reduce the travel cost. In this thesis, we address this problem by finding bilateral matchings between users and drivers to alleviate the overload issue efficiently and effectively. Besides, we consider to report the update time, which is the runtime to update drivers' trip schedules as per incoming rider requests. The update time is an important metric to demonstrate the system stability and has been neglected by the existing literatures. An alternative strategy to mitigate the peak demand in the trip is to find the pick-up location with most users and then dispatch more drivers to the location. In such a way, more users in high-demand areas can take a ride to make a trip. This transforms into a variant of MaxBRkNN problem under streaming users scenario. In this thesis, we find the optimal location which the most users regard as one of the kNNs efficiently.</p> <p>After the trip, the system can leverage and analyze the historical trip data to help improve the user experience. Specifically, we aim to identify traffic bottlenecks in a road network, which can easily lead to traffic congestion by traffic flow propagation and then bring unsatisfying trip experience. These traffic bottlenecks can be used as signals to remind drivers to adjust the current trip dynamically, or help road expansion in physical infrastructure. In this thesis, we deal with the traffic bottleneck identification problem efficiently and effectively by considering the influence among edges. We compute the spread influence of each edge, and then select a set of seed edges with the largest influence.</p>