Public dataset for multi-modal mobility

👋 Read the quickstart first. It has examples that can give you a quick overview.

Background information

1. Evaluation procedure
2. Evaluation metrics

Data characteristics

This dataset contains data from 3 artificial timelines. The timelines cover 15 separate modes, including recently popular modes such as e-scooter and e-bike. We had two main goals for the data collection.

Dwell time

Instead of focusing only on trips, we wanted to evaluate a timeline that included significant dwell time. We could see from our calibration runs that android appears to have built-in duty cycling. Including significant dwell time would allow us to capture the impact of this context sensitive behavior. Therefore, we structured our timeline trips as round trips to libraries with an intermediate dwell time ~ 3x the mean travel time to the location.

Broad range of modes

Since we are creating artificial trips, we can structure them to maximize mode variety. In order to efficiently cover this space, we tried to ensure that no mode was repeated. We only had to include commuter rail twice since there were few other transit options to reach the starting point chosen.

Multi-modal transfers

Detecting multi-modal transfers is tricky because there isn't a clear signal similar to a trip end. We ensure that there are many transition examples by emphasizing multi-modal transfers while constructing our artificial trips.

A brief summary of the timelines is as below. The details are in the dataset (summary and filled).

id	Description	Outgoing trip modes	Incoming trip modes	Travel time	Dwell time	Overall time
unimodal_trip_car_bike_mtv_la	Suburban round trip	car (suburban street)	bicycle	40 mins	1.5 hr	2 hrs
car_scooter_brex_san_jose	Downtown library	car (freeway)	escooter, bus rapid transit	3 hrs	3 hrs	6 hrs
train_bus_ebike_mtv_ucb	Multi-modal trip across the bay	suburb walk, commuter train, subway, city bus, university walk	ebike, express bus, downtown walk, light rail, commuter train with tunnels, suburb walk	6 hrs	6 hrs	12 hrs

We currently have the following data from these timelines.

id	accuracy control	HAHFDC	HAMFDC	MAHFDC	MAMFDC	total runs	travel hrs	travel hrs for android + iOS	total hrs	total hrs for android + iOS
unimodal_trip_car_bike_mtv_la	6	6	3	3	0	18	~ 12	~ 24	36	72
car_scooter_brex_san_jose	7	6	3	4	1	21	63	126	126	252
train_bus_ebike_mtv_ucb	12	6	6	6	6	36	216	432	432	864
Total								582		1188

• I think that the travel time is ~ 10 mins there by car and ~ 30 mins back by bike, so a dwell time of 1 hour there. But we also wait for 30 mins after coming back to ensure that the trip end is detected. Should verify against the actual data.

Reroutes

Since the data was collected over multiple months, there were small reroutes to some of the trajectories. Notably:

• I used a ramp instead of stairs to wheel the non-bikeshare e-bike to the bikeshare dock (see Acknowledgements below)
• The express bus was rerouted from the Temporary Transbay Terminal to the Salesforce Transit Center
• I was not allowed to ride the LBNL shuttle after the first round and the replacement UC Berkeley shuttle route was off by a block

This will be automatically handled in the spec.

Acknowledgements

I would like to thank Harrison Liew and gennui raffill for their assistance in salvaging this data collection effort after the Lyft e-bike debacle. The third timeline includes a one-way ride from the UC Berkeley campus to the Transbay bus stop on an e-bike. I used Lyft bikeshare e-bikes for the first round of data collection, but then a couple of e-bikes caught on fire, and Lyft pulled all e-bikes from service. I was now stuck since using a regular bike would preclude comparisons with the first round.

Fortunately, Harrison and gennui stepped in and helped out by lending an e-bike and riding the bike back to campus to complete the round trip. Neither rain nor heat nor lost phones or unexpected band practice stayed us from the slow collection of all four quadrants of the frequency/accuracy combinations. And Lyft had still not put e-bikes into service at this time, so this dataset would not have been possible without their help.

Please cite the following work when using this dataset:

• Shankari, K. and Fuerst, Jonathan and Fadel Argerich, Mauricio and Avramidis, Eleftherios and Zhang, Jesse. (2020). MobilityNet: Towards a Public Dataset for Multimodal Mobility Research. Climate Change AI 2020 @ ICLR20 (Spotlight Talk).