Back to gmdhshell.com
After the computation has finished, visualization helps to explore computational results using the plot, the table of predictions (Table), the table of variables' importance (Importance), the performance panel and a special tool called model browser.
In this article visual elements of the GMDH Shell window show results of Time series forecasting example - Top500.
The model browser enables a user to select a target variable and explore its models. All highly ranked models can be viewed using the Model rank parameter of the Model browser.
Visualization panels dynamically show structure, coefficients, statistics and a plot of any selected Model rank. In case of series of simulations you can also browse model produced for different time points using the ID control. Any changes in the Model browser will be visualized in related panels immediately.
When the Keyboard control button is clicked, Up/Down arrow keys change Model rank and Left/Right arrows change ID. Mouse wheel changes Model rank.
The model selected in the Model browser appears in the Plot as a blue curve with red addition that is model forecast.
At the Plot tab the current model predicts observation #30. You can see that model ID #29+1 is selected in the Model browser (see above picture), '+1' means the model predicts one step ahead. Red points are post-processed (averaged) predictions, they belong to a series of simulations performed for different time points, i.e. #29+1, #30+1, #31+1 and finally #32+1 that predicts unknown value. You can switch to any of these models in the Model browser. Gray curve represents known observations of the target variable. We predict last three gray-colored observations on assumption that we don't know their actual values.
The current model (Rank 1, Time point 24) is a sum of 4 components selected out of 61 available components.
Testing performance of the current model in terms of the selected validation criterion.
Should be interpreted as the following model:
Rmax[t] = 1.078*Year[t-1] + 4.2021*RMax[t-7] - 0.4531*RPeak[t-5].
The performance panel shows how accurate our models at the known part of data in terms of different error measures are.
| Error measure | Mean | Root mean square |
| Absolute | MAE: Mean absolute error | RMSE: Root mean square error |
| Range percentage | NMAE: Normalized mean absolute error | NRMSE: Normalized root mean square error |
| Target percentage | MAPE: Mean absolute percentage error | RMSPE: Root mean square percentage error |
In order to see the accuracy estimations in the first two sections (Post-processed predictions and Current model predictions) we should provide the Performance panel with at least a small number of actual values of the target variable that we are trying to model. This requirement is met if Preprocess is set to hold-out some instances from the Solver or, if we apply a model to a new data file.
The performance panel shows Maximal positive, Maximal negative, Mean absolute and Root mean squared values of error. Error values are either absolute or normalized by range of the output variable or normalized by values of the target variable. The range of target variables is always calculated only for data-points used for learning, i.e. data points that fall under training and testing parts.
For classification problems the number of model misses is measured. Integer values of target variables will be used as different classes for model performance measuring. Class A always corresponds to the smallest integer class, for example 0 if a two-class problem consists of classes 0 and 1.
The importance tab shows the absolute number of times a certain model component was used in the set of obtained models. In case of a linear base model it represents the importance of plain variables.
Enumeration of prediction coincides with plot axes at the Plot tab.
Unique data row identifiers.
Actual values of the target variable.
Post-processed predictions of the target variable.
Predictions of a model that is currently selected in the Model browser.
A multi-target report allows a user to see in one table target names, predictions, past performance and more. You can use the configuration button to tune the appearance of a report. Also, you can save a multi-target report to html file or print it. To save it in the pdf format, we recommend using one of third party virtual pdf printers.