Research Area: Data Mining
An Advanced Traveler Information System (ATIS) collects, processes and presents road traffic information to the users, assisting them in their trips and supporting them in making the necessary pretripanden route decisions. With this purpose in mind, traffic modeling becomes one of the most important tasks of ATIS, because it enables the description, simulation and forecasting of the traffic variables of interest. Among all the traffic variables that can be modeled (flow, road occupancy, speed, travel time, etc.), travel time acquires a special relevance in ATIS, because the concept can be easily understood by travelers. As such, travel time estimation and prediction are two of the most common traffic modeling problems ATIS has to deal with.
The importance of these two modeling problems has raised the interest of the research community in the past few years, and has thus resulted in a vast number of publications and model proposals. However, a detailed analysis and review of the state-of-the-art shows that not all the proposed models are adequate for all the study sites, traffic situations, available data, etc. In this context, the combination or fusion of models seems to be one of the most promising research lines, because it allows the use of specific and more suitable models for each case.
A specific type of combined or hybrid travel time models are those that initially preprocess the data using clustering algorithms, with the aim of identifying and separating the different traffic patterns that may be present.Then, a separate and suitable travel time model is built for each cluster,allowing the construction of more specific models for each case. Particularly, a special case of these combination models relies on the paradigm of time series clustering, in which each instance to be clustered is a whole time series, for example, the sequence of travel time measurements collected throughout a day.
To begin with, clustering a time series dataset requires making some non-trivial decisions, such as selecting a suitable distance measure. There are in-numerable distance measures specifically designed for time series data, and it has been demonstrated previously in the literature that there is no unique distance measure that is suitable for all the databases. Indeed, it seems that the specific characteristics of each database have a strong impact on the performance of the different existing measures. In this context, an interesting question that will be addressed in this dissertation is whether, given a set of characteristics of the database, it is possible to automatically select a suitable distance measure(s) from a set of candidates.
Name of the Researcher: Usue Mori Carrascal
Name of the Supervisor(s): Jose A. Lozano
Year of Completion: 2015
University: University of the Basque Count
Thesis Link: Home Page Url
