1. Using Google’s Aggregated and Anonymized Trip Data to Estimate Dynamic Origin-Destination Matrices for San Francisco
TRB Applications Conference 2017
Bhargava Sana, Joe Castiglione,
Dan Tischler, Drew Cooper
Good Afternoon! I am here to share our recent experience with a new passive data source from Google to estimate OD matrices for San Francisco Bay Area.
SAN FRANCISCO COUNTY TRANSPORTATION AUTHORITY
May 16, 2017

2. OD Data Collection Conventional methods
License plate surveys
Roadside interviews
Emerging passive data collection methods
Bluetooth detectors
Cell phone Call Detail Records (CDR) data
GPS, Wi-Fi detectors
Historically, trip OD info available only through expensive and time-consuming surveys
Recently, use of passive data collection technologies is gaining momentum

3. Google’s Better Cities Program Aggregated and Anonymized Trip (AAT) Data
Minimize congestion, improve safety and reduce infrastructure spending
Aggregated and Anonymized Trip (AAT) information from location reports
Extract data from moving users
Clean data and snap to road network
Aggregate OD trip counts
Apply differential privacy filters and minimum trips threshold
Partnered with Google
Better Cities program made available aggregated and anonymized OD flow information from location reports of mobile device users
Reports based on data from users who chose to store their location information from Google-enabled devices on Google servers
Easy to turn it on in account settings and record trips fairly accurately; No active navigation constraint
Describe AAT creation process
AAT DATA

4. Google AAT Dataset
Hourly AAT data for six months (Apr-Jun and Sep-Nov 2015)
Flow data provided as relative trips as opposed to absolute counts
Convert relative flows to trips using HH travel survey?
Hourly flow data provided for 6 months in 2015
85 districts defined for 9-county San Francisco Bay Area: combination of Tract and County boundaries
Google provided relative flows rather than absolute counts stating that location reports account for only a sample of travelers
But we need trips for planning purposes; maybe relative flows could be converted to trips using regional HH travel survey?
A quick comparison with CHTS flows shows that the relative magnitude of AAT flows is reasonable
CHTS is sparse with data for less than half the total OD district pairs (7,225) but AAT has all ODs

5. Relative Flow Conversion Model
𝑷𝑻 𝒐𝒅𝒕 = β 𝒕 𝑹𝑭 𝒐𝒅𝒕 +ϵ ; 𝟎≤𝐭≤𝟐𝟑
𝑃𝑇 𝑜𝑑𝑡 : Person trips between o-d for hour-of-day t
𝛽 𝑡 : Coefficient of AAT relative flow for t
𝑅𝐹 𝑜𝑑𝑡 : Avg AAT flow between o-d for hour-of-day t
𝜖 : Error term
No geographic constants applied
District- and County-level regression models estimated
20% sample used for validation
The relationship between relative flows and trips appeared to vary by geography as well as time-of-day
However only TOD coefficients used so that the conversion model is independent of OD markets
Both district-level and county-level regression through origin (RTO) models were estimated
A 20% holdout sample was used for validation in each case

6. Origin-Destination-Hour
Predicted vs Observed
Origin-Destination-Hour
Scatter plots of predicted vs observed hourly OD flows show that the models perform reasonably well except county model validation sample. Please note, each point is an origin-destination-hour.
RMSE and Mean Absolute Error (MAE) show that county model is unsurprisingly better due to higher spatial aggregation

7. Summary AAT relative flow magnitudes correlated with actual trips
AAT geographic coverage significantly higher
Simple linear regression model may be used for conversion
Could support measuring longitudinal variation
Further studies
Better/smooth survey data
Compare with cell CDR data