Inferring dockless shared bike distributions in new cities by transfering the knowledge from the cities popular with the dockless shared bikes.
Keras implementation for the paper 'Inferring Dockless Shared Bike Distribution in New cities', which was publised on the WSDM 2018.
Foreknowing the dockless shared bike distributions in a new city is of great significance for the design of the bike delivering strategy, government regulations, renewing the traffic rules, and so on. In this paper, based on the multi-source data and known bike distirbutions in the delivered cities, we employed the convolutional neural network to model the interaction between geo-related information and the bike distributions, and then applied this learnt knowledge to the target city to infer the potential dockless shared bike distributions.
We collect muliple geo-related data from differet sources, including:
- Mobike data (by crawler);
- POI data from (Baidu Map);
- Satellite light (Light Polution);
- Road network (OpenStreetMap);
- Transportation hubs (collect manually);
- Business centers (Fliggy);
If you're interested in these data, you can crawl or get the data from our provided urls.
-
Run script:
>> python main.py --train_cities bj --test_cities nb --model_choice 0 --y_scale --epochs 200
-
Parameter description:
- model_choice: model choice for the problem;
- train_cities: source cities we would transfer knowledge from
- target_cities: target citeis that will be applied the model
- y_scale: whether to scale the target label;
- epochs: maximum training epochs;
- Bike lane planning work in KDD 2017, paper link:
@inproceedings{Bao2017Planning,
title={Planning Bike Lanes based on Sharing-Bikes' Trajectories},
author={Bao, Jie and He, Tianfu and Ruan, Sijie and Li, Yanhua and Zheng, Yu},
booktitle={ACM SIGKDD International Conference on Knowledge Discovery and Data Mining},
pages={1377-1386},
year={2017},
}
- For academic and non-commercial use only.
- For commercial use, please contact Mobike Company