Standard Track

In the field of traffic big data, there have been long-standing phenomena such as inconsistent evaluation data sets, inconsistent evaluation indicators, and inconsistent preprocessing of data sets, resulting in poor performance comparability of different models. Therefore, in order to solve the above problems, this project has implemented a set of standard pipelines (tracks) for each task.

On the standard track, the original data set, standard data module (Data module), and standard evaluation module (Evaluator module) provided by the project are used to constrain different models to use the same data input and evaluation indicators to improve the comparability of evaluation results.

The standard data input format and evaluation input format for different tasks are explained below :

Traffic State Prediction

According to the different spatial structure of traffic data, traffic state data can generally be represented by tensors in the following formats:

  • A three-dimensional tensor shaped like (N,T,F), T is the length of time, F is the feature dimension, and N is the number of sensors.

  • A four-dimensional tensor shaped like (T,F,I,J), T is the length of time, F is the feature dimension, and I,J represents the row and column index of the grid data.

  • A four-dimensional tensor shaped like (T,F,S,T), T is the length of time, F is the feature dimension, and S,T represents the id of the origin and destination of the od data.

  • A six-dimensional tensor shaped like (T,F,SI,SJ,TI,TJ), T is the length of time, F is the feature dimension, SI,SJ,TI,TJ represents the row and column index of the origin and destination of the grid-od data.

The standard data input format is a dictionary-like Batch object instance. The key names of this object are as follows:

  • X: The multi-dimensional tensor input by the model, shape = (batch_size, T_in, space_dim, feature_dim), each dimension represents the total number of samples in the batch, the width of the input time window, the spatial dimension, and the data feature dimension. In particular, the spatial dimension can be N or I, J or S, T or SI, SJ, TI, TJ as mentioned above.

  • y: The multi-dimensional tensor that the model expects to output, shape = (batch_size, T_out, space_dim, feature_dim), each dimension represents the total number of samples in the batch, the width of the output time window, the spatial dimension, and the data feature dimension. Among them, the spatial dimension can be N or I, J or S, T or SI, SJ, TI, TJ as mentioned above.

  • X_ext: Optional external data, shape = (batch_size, T_in, ext_dim), each dimension represents the total number of samples in the batch, the width of the input time window, and the feature dimension of the external data. In particular, some models may directly incorporate X_ext into X as the input of the model.

  • y_ext: Optional external data, shape = (batch_size, T_out, ext_dim), each dimension represents the total number of samples in the batch, the width of the output time window, and the feature dimension of the external data.

The standard evaluation input format is a dictionary object, and the dictionary has the following key names:

  • y_true: The ground-truth value, the format is the same as the y in the input.

  • y_pred: The prediction value, the format is the same as the y in the input.

Trajectory Location prediction

The standard data input format is a dictionary-like Batch object instance. The key names of this object are as follows:

  • history_loc: Historical trajectory location information, shape = (batch_size, history_len), history_len is the length of the historical trajectory.

  • history_tim: Historical trajectory time information, shape = (batch_size, history_len).

  • current_loc: Current trajectory location information, shape = (batch_size, current_len), current_len is the length of the current trajectory.

  • current_tim: Current trajectory time information, shape = (batch_size, current_len).

  • uid: The id of the user for each trajectory, shape = (batch_size).

  • target: Expected next hop location, shape = (batch_size) .

The standard evaluation input format is a dictionary object, and the dictionary has the following key names:

  • uid: The id of the user for each trajectory, shape = (batch_size).

  • loc_true: Expected next hop location, shape = (batch_size).

  • loc_pred: Model prediction output, shape = (batch_size, output_dim).

Map Matching

The standard data input format is a dictionary. The key names of this object are as follows:

  • trajectory: The format of trajectory can be denoted as {"usr_id":{"traj_id":{numpy.ndarray}}}. That is to say, the key of trajectory is usr_id. Each usr_id has a dictionary, the key of which is traj_id. for each traj_id, there’s anumpy.ndarray, representing a sequence of chronologically ordered spatial points sampled from a continuously moving object, with columns=(index,longitude,latitude,time) or columns=(index,longitude,latitude). The length of trajectory is noted as num_sample.

  • rd_nwk: A road network with type networkx.classes.digraph.DiGraph.

  • route: The format of route can be denoted as {"usr_id":{"traj_id":{numpy.ndarray}}}. It is similar to trajectory. The value of route is a numpy.ndarray of geo_id with shape=(num_road,), representing ground truth. num_road is the length of real route.

The standard evaluation input format is a dictionary object, and the dictionary has the following key names:

  • result: The format of result is almost the same as that of route in standard data input. The value of result is a numpy.ndarray of geo_id with shape=(num_sample,), representing matching result. num_sample is the number of GPS samples in the trajectory.

  • route: As depicted in standard data input.

  • rd_nwk: As depicted in standard data input.

Estimated Time of Arrival

The standard data input format is a dictionary-like Batch object instance. The key names of this object are as follows:

  • current_loc/(current_longi, current_lati): Trajectory location information, shape = (batch_size, traj_len), traj_len is the length of the trajectory.

  • uid: The id of the user for each trajectory, shape = (batch_size).

  • timeid(weekid): Time information when the trajectory starts, shape = (batch_size).

  • dist: The total distance of the trajectory, shape = (batch_size).

  • Other information, such as current_dis(the distance from starting point to current point, shape = (batch_size, traj_len))， current_state(the current taxi state is available or unavailable, shape = (batch_size, traj_len)). (Optional)

The standard evaluation input format is a dictionary object, and the dictionary has the following key names:

  • y_true: The real travel time from starting point to finishing point, shape = (batch_size).

  • y_pred: The predicted travel time from starting point to finishing point, shape = (batch_size).