190 likes | 205 Views
Spatiotemporal Pattern Mining For Travel Behavior Prediction. UIC IGERT Seminar 02/14/2007 Chad Williams. Agenda. What is the problem What is the goal Hypothesis and objectives Novelty of work Approach. Problem. Accurately predicting and modeling individual daily traveler behavior
E N D
Spatiotemporal Pattern Mining For Travel Behavior Prediction UIC IGERT Seminar 02/14/2007 Chad Williams
Agenda • What is the problem • What is the goal • Hypothesis and objectives • Novelty of work • Approach
Problem • Accurately predicting and modeling individual daily traveler behavior • Number of trips • Source and destination • When trips are made and how • What factors affect these decisions • Motivation • Regional planning • Intelligent Traveler's Assistant (ITA)
Current challenges • Quality of dataset - paper based travel diaries • Tendency to underestimate actual travel time • Time events occurred is imprecise/unreliable • Locations/activities forgotten during entry • Level of predictions • Transferability of predictions unreliable/difficult
Goal • Gain insight into travel decisions and behavior • Leverage richer datasets available through new technologies (GPS enabled PDA surveys) • More accurate time information • GPS coordinates of destinations and route choice • Location characteristics/Mode alternatives • Planned vs. actual travel behavior • Mine meaning from the combined stream • Enter spatiotemporal pattern mining!!
What is spatiotemporal mining? • Data mining across the dimensions of space and time simultaneously • Relation to transportation • Spatial characteristics of transportation network, and accessibility also known to affect choices • Travel behavior is known to change over time (ie. throughout the day) • Traveler profile known to influence mode choice, activity choice • Thus it is really multi-dimensional mining, but spatiotemporal mining is the catch word right now.
Hypothesis • Patterns across travelers that take all of these dimensions into account will improve the transferability of mined patterns; resulting in improved traveler behavior prediction
Research Objectives • Algorithms for mining patterns that span non-aligned information streams (more about this later) • Spatial – location, route • Temporal – activity timing, congestion fluctuation • Decision behavior – planned vs. actual • Formal models • Destination and mode choice prediction • Formal model for prediction of trip chaining • Formal measure of “interestingness” in spatiotemporal travel patterns
Prior work • Activity & travel behavior modeling • Move travel behavior prediction to individual level • Survey data used to model activity patterns • Past focus largely on activity order and frequency to understand traveler behavior (Jones et al. 1990 & Axhausen and T. Grling 1992) • Recent work has shifted focus to the scheduling process and factors which affect decisions(Mohhamadian and Doherty 2005, 2006)
Prior work (cont.) • Destination, traveler movement prediction • Short term prediction for cellular networks • General traveler movement (Liu and Maguire 1995) • Mobile user travel profiles (Bhattacharya and Das 1999) • Aggregate activity based travel (Wang and Cheng 2001) • GPS log history for individual travel prediction (Ashbrook and Starner 2002, 2003) • Information stream alignment (A Marascu and F Masseglia 2005, Kleinberg 2007)
Novelty of Work • Key differences from prior work • Intraday temporal analysis of patterns • Transference of spatial patterns • Integration of these aspects for insight into traveler decision making and behavior • Collaborative approach to behavior prediction
Research Objectives • Algorithms for mining patterns that span non-aligned information streams (more about this later) • Spatial – location, route • Temporal – activity timing, congestion fluctuation • Decision behavior – planned vs. actual • Formal models • Destination and mode choice prediction • Formal model for prediction of trip chaining • Formal measure of “interestingness” in spatiotemporal travel patterns
Why is the integration significant? • Traditional data mining techniques focus on a single plane • Classification algorithms (ex. Amazon/NetFlix) • Many “spatial dimensions” • User characteristics (income, family size, etc) • Buying/Rent history • Movie preferences/ratings • But values fixed rather than range over time • Temporal mining • Examines single item or set of items as it changes over time • Lose context of role “spatial” characteristics play in influencing these changes
Spatial transferability • Consider 2 working adults • Similar family profiles • Each work similar hours • But live and work in different areas of town or even different towns. • Which behaviors are transferable which aren’t? • Key is mining spatial similarity in conjunction with behavior similarity • Accessibility, density of wanted attractions, travel time between locations, trip chaining tolerances
Research Objectives • Algorithms for mining patterns that span non-aligned information streams (more about this later) • Spatial – location, route • Temporal – activity timing, congestion fluctuation • Decision behavior – planned vs. actual • Formal models • Destination and mode choice prediction • Formal model for prediction of trip chaining • Formal measure of “interestingness” in spatiotemporal travel patterns
Formal Models • Destination and mode choice prediction • ITA benefits • Further understanding for planners • Transit routes that are likely to get used • Factors most likely to influence mode decision • Formal model for prediction of trip chaining • Optimize individual’s schedule • Suggest routes or additional stops • Key is using insight into decision making to only suggest changes traveler actually might consider
Research Objectives • Algorithms for mining patterns that span non-aligned information streams (more about this later) • Spatial – location, route • Temporal – activity timing, congestion fluctuation • Decision behavior – planned vs. actual • Formal models • Destination and mode choice prediction • Formal model for prediction of trip chaining • Formal measure of “interestingness” in spatiotemporal travel patterns
“Interestingness” • Because of size of information space need to reduce the amount of data that is actually looked at • Metric for identifying what rules or relations are likely to be of interest and which are not • Ex. text mining -> stop words • Introduce metric for reducing these types of rules in the transportation domain
Caveat • Not a silver bullet • Unlikely to solve all of these problems • General idea is that by integrating multiple views of information that influences traveler’s behavior, we will be able to get a richer model for prediction.