Traffic State Prediction Evaluator

We have implemented several evaluation loss functions so that different models under the same task can be compared under the same standard.

Evaluation Metrics

For the task of traffic state prediction, this evaluator implements a series of evaluation indicators:

Evaluation Metrics

Formula

MAE(Mean Absolute Error)

\[MAE=\frac{1}{n}\sum_{i=1}^n|\hat{y_{i}}-y_i|\]

MSE(Mean Squared Error)

\[MSE=\frac{1}{n}\sum_{i=1}^n(\hat{y_{i}}-y_i)^2\]

RMSE(Rooted Mean Squared Error)

\[RMSE=\sqrt{\frac{1}{n}\sum_{i=1}^n(\hat{y_{i}}-y_i)^2}\]

MAPE(Mean Absolute Percent Error)

\[MAPE=\frac{1}{n}\sum_{i=1}^n|\frac{\hat{y_{i}}-y_i}{y_i}|*100\%\]

R2(Coefficient of Determination)

\[R^2=1-\frac{\sum_{i=1}^n(y_i-\hat{y_i})^2}{\sum_{i=1}^n(y_i-\bar{y})^2}\]

EVAR(Explained variance score)

\[EVAR =1-\frac{Var(y_i-\hat{y_i})}{Var(y_i)}\]

The ground-truth value is \(y=\{y_1,y_2,...,y_n\}\), the prediction value is \(\hat{y} = \{\hat{y_1}, \hat{y_2}, ..., \hat{y_n}\}\)\(n\)is the number of samples, the mean value is \(\bar{y}=\frac{1}{n}\sum_{i=1}^ny_i\), the variance is \(Var(y_i)=\frac{1}{n}\sum_{i=1}^n(y_{i}-\bar{y})^2\).

Evaluation Settings

The following are parameters involved in the evaluator:

Location: libcity/config/evaluator/TrafficStateEvaluator.json

  • metrics: Array of evaluation metrics, allowed_metricsin evaluator class indicates the type of metrics that the task can accept, and metrics cannot exceed this range.

  • mode: Evaluation mode, traffic state prediction is generally a prediction of multiple time steps. If set to average, it means calculating the average result of the previous n time steps, and set to singleto calculate the n-th time step evaluation results. The default is average. The current evaluator will return the results of all time steps. For example, if the total time step is 3, the averagemode returns [average loss of previous 1 time step, average loss of previous 2 time steps,average loss of previous 3 steps], The singlemode returns [loss at the first time step, loss at the second time step, loss at the third time step].