Implemented Dataset Class

Here we introduce the functions of several dataset classes that have been implemented.

• AbstractDataset

  Base class for all dataset classes. Note that this is an abstract class and can not be used directly.

• TrajectoryDataset

  Base class for all trajectory location prediction tasks. The raw trajectory records will be cut according to the set window_size and cut_method, which means the months-long trajectory in the raw data set will be cut into sub-trajectories that meet the single travel time/distance length. After the cutting is completed, the corresponding trajectory spatiotemporal feature encoder will be called according to the set traj_encoder parameter to extract the features of the trajectory and generate model input.

• TrafficStateDataset

  One of base class for all traffic state prediction tasks. Note that this is an abstract class and cannot be used directly. By default, the data of input_window is used to predict the data corresponding to output_window. The Batch object generated by this class contains two keys, one X and one y. Here input_window and output_window are parameters for data, see here for details.

• TrafficStateCPTDataset

  Another base class for all traffic state prediction tasks. Note that this is an abstract class and cannot be used directly. Part of the traffic prediction model realizes prediction by modeling the closeness/period/trend. By default, the data of len_closeness/len_period/len_trend is used to predict the data at the current moment(a single-step forecast). The Batch object generated by this class contains 4 keys: X, y, X_ext, y_ext. Here len_closeness/len_period/len_trend are parameters for data, see here for details.

• TrafficStatePointDataset

  A class inherited TrafficStateDataset for traffic state prediction. The dataset is used for point-based/segment-based/region-based dataset as long as the spatial dimension of this data set is 1-dimensional. The shape of tensor in the Batch object generated by this class is 3-dimensional, namely space_dim, time_dim, feature_dim.

• TrafficStateGridDataset

  A class inherited TrafficStateDataset for traffic state prediction. The dataset is used for grid-based dataset. The shape of tensor in the Batch object generated by this class is 3-dimensional or 4-dimensional depends on parameter use_row_column. If set use_row_column=True, then the 4 dimensions are grid_row_dim, grid_column_dim, time_dim, feature_dim. Otherwise, the 3 dimensions are space_dim, time_dim, feature_dim, in this case the grid is renumbered in one dimension.

• TrafficStateOdDataset

  A class inherited TrafficStateDataset for traffic state prediction. The dataset is used for od-based dataset, which means origin and destination. The shape of tensor in the Batch object generated by this class is 4-dimensional, namely origin_dim, destination_dim, time_dim, feature_dim.

• TrafficStateGridOdDataset

  A class inherited TrafficStateDataset for traffic state prediction. The dataset is used for grid-od-based dataset. The shape of tensor in the Batch object generated by this class is 4-dimensional or 6-dimensional depends on parameter use_row_column. If set use_row_column=True, then the 6 dimensions are origin_grid_row_dim, origin_grid_column_dim, destination_grid_row_dim, destination_grid_column_dim, time_dim, feature_dim. Otherwise, the 4 dimensions are origin_dim,  destination_dim, time_dim, feature_dim, in this case the grid is renumbered in one dimension.

• MapMatchingDataset

  Base class for all map matching tasks. This class generates a dictionary which contains 3 keys:rd_nwk, trajectory and route, representing road network, trajectory of GPS samples and ground truth respectively. if delta_time=True is set, trajectory will include a time column indicating the reading of the seconds. delta_time is a parameters for dataset, see here for details. see here for introduction of standard data input.

• ETADataset

  Base class for all estimated time of arrival (ETA) tasks. Function _load_dyna will load the trajectory information. Function _encode_traj will call the corresponding trajectory spatio-temporal feature encoder according to the parameter eta_encoder to extract the features of the trajectory. The extracted features will be divided in training data, evaluation data and test data to generate model input.