Show simple item record

dc.contributor Wang, Xuyu en_US
dc.contributor.advisor Chen, Haiquan en_US
dc.contributor.author Pandya, Darshit A.
dc.date.accessioned 2020-03-11T16:11:08Z
dc.date.available 2020-03-11T16:11:08Z
dc.date.issued 2020-03-11
dc.date.submitted 2019-12-05
dc.identifier.uri http://hdl.handle.net/10211.3/215255
dc.description Project (M.S., Computer Science)--California State University, Sacramento, 2019. en_US
dc.description.abstract These days, when the concept of vehicle sharing has been exploited by people a lot, it becomes necessary to predict the flow of the vehicles to match the demand with the availability. Nowadays, in metros like San Francisco and New York, peak hour traffic problem has been a major concern for most of the daily commuters and residents. Bikes are an easy solution for the short commutes and hence should be made available in the designated stations all day round. This project visits the problem of predicting the bike flow at station-level using a novel architecture of Spatial-temporal Graph Convolutional Network. In this project, we consider two types of relations, the temporal patterns and the spatial correlation between the stations to predict the next hour rides. The graph constructed as an input to the model has nodes which are bike stations and edges with weights as the distance between the stations. We consider the graph to be a fully connected graph. The input graph along with the temporal sequences serves as an input to the model with complete convolutional structures. Using the Euclidean haversine distance algorithm, we calculate the great circle distance between the stations in miles. Unlike normal time series analysis, we use distance graph as a relationship matrix between each of the stations. The model was evaluated using the open sourced ford go (now Lyft) bike demand dataset of City of San Francisco with more than 3 million trips data to predict the next hour ride at both station-level and cluster-level. Extensive experiments validate that our STGCN-based model outperformed the baseline methods in terms of RMSE in three different settings, i.e., at the top 10 most popular stations, at the cluster level, and at the station level. en_US
dc.description.sponsorship Computer Science en_US
dc.language.iso en_US en_US
dc.subject Neural networks (Computer science)--Industrial applications en_US
dc.subject Bicycle traffic flow en_US
dc.subject Data mining en_US
dc.subject Temporal databases en_US
dc.title Station-graph: bike flow prediction using spatio-temporal graph convolutional network en_US
dc.type Project en_US

Files in this item


This item appears in the following Collection(s)

Show simple item record

Search DSpace

My Account

RSS Feeds