Fitting functions with a configurable XGBoost regressor

This post deals with the approximation of scalar mathematical functions to one or more real variables using a XGBoost regressor without writing code but only acting on the command line of Python scripts that implement the functionality of:

  • Regressor Configuration and Training
  • Prediction and error calculation
The goal is to demonstrate that regression with XGBoost achieve low error values with extremely short learning times. Although MLP (Multi Layer Perceptron) type neural networks can be considered universal function approximators (see Fitting with highly configurable multi layer perceptrons on this website), the XGBoost regressor can arrive at lower error values than an MLP and with a significantly lower computational cost than that of an MLP.

In the real world the datasets pre-exist the learning phase, in fact these are obtained by extracting data from production databases or Excel files, from the output of measuring instruments, from data-loggers connected to electronic sensors and so on, and then used for the next stages of learning; but since the focus here is the regression itself and not the fitting of a real phenomenon, the datasets used in this post have been synthetically generated from mathematical functions: this has the advantage of being able to stress the algorithm and see for which types of datasets the algorithm has acceptable error values and for which instead the algorithm is struggling.

The code described by this post requires Python version 3 and uses the XGBoost library; it also requires the SciKit-Learn, NumPy, Pandas, MatPlotLib and JobLib libraries.
To get the code please see the paragraph Download of the complete code at the end of this post.

For the generation of synthetic training, validation and test datasets, common tools will be used: fx_gen.py for real-valued scalar generator functions of a real-valued variable and fxy_gen.py for the real-valued scalar generator functions of two real-valued variables.
Also for the visualization of the results, and precisely for the comparison of the test dataset with the prediction, the general tools will be used: fx_scatter.py for real-valued scalar generator functions of a real-valued variable e fxy_scatter.py for the real-valued scalar generator functions of two real-valued variables.

Regressor Configuration and Training

In this chapter the program fit_func_miso.py is presented and it is technically a wrapper of the class XGBRFRegressor of the XGBoost library and whose purpose is to allow the use of the regression of the underlying regressor to fit functions without having to write code but only acting on the command line.
In fact through the argument --xgbparams the user passes a series of hyper-parameters to adjust the behavior of the'underlying XGBoost regressor algorithm and others to configure its learning phase. In addition to the parameters of the underlying regressor the program supports its own arguments to allow the user to pass the training dataset and optionally the validation dataset, on which file to save the trained model, the metrics to calculate during the training, constraints for regularization (e.g. early stop) and parameters for diagnostic.

The program fit_func_miso.py, as well as the underlying XGBoost regressor, is of type M.I.S.O., i.e. Multiple Input Single Output: it is designed to fit a function of the form $y=f(x_1, ..., x_n)$ where the number of independent variables is arbitrarily large while the output dependent variable is only one.
The format of the input datasets is in csv format (with header), with n + 1 columns, of which the first n ones contain the values of the n independent variables and the last column, the n+1, containing the values of the dependent variable.

Usage of the fit_func_miso.py program

To get the program usage you can run this following command:

$ python fit_func_miso.py --help
and the output got is:

usage: fit_func_miso.py [-h] [--version] --trainds TRAIN_DATASET_FILENAME
                        --modelout MODEL_FILE [--valds VAL_DATASET_FILENAME]
                        [--metrics VAL_METRICS [VAL_METRICS ...]]
                        [--dumpout DUMPOUT_PATH]
                        [--earlystop EARLY_STOPPING_ROUNDS]
                        [--xgbparams XGB_PARAMS]

fit_func_miso.py fits a multiple-input single-output scalar function dataset
using a configurable XGBoost

optional arguments:
  -h, --help            show this help message and exit
  --version             show program's version number and exit
  --trainds TRAIN_DATASET_FILENAME
                        Train dataset file (csv format)
  --modelout MODEL_FILE
                        Output model file
  --valds VAL_DATASET_FILENAME
                        Validation dataset file (csv format)
  --metrics VAL_METRICS [VAL_METRICS ...]
                        List of built-in evaluation metrics to apply to
                        validation dataset
  --dumpout DUMPOUT_PATH
                        Dump directory (directory to store metric values)
  --earlystop EARLY_STOPPING_ROUNDS
                        Number of round for early stopping
  --xgbparams XGB_PARAMS
                        Parameters of XGBoost constructor
Where:
  • -h, --help: shows the usage of the program and ends the execution.

  • --version: shows the version of the program and ends the execution.

  • --trainds: path (relative or absolute) of a two-column csv file (with header) that contains the dataset to be used for the training; this file can be generated synthetically e.g. via the program fx_gen.py. or be a dataset actually obtained by measuring a scalar and real phenomenon that depends on a single real-valued variable.

  • --modelout: path (relative or absolute) to a file where to save the trained model in joblib format (.jl).

  • --valds: path (relative or absolute) of a two-column csv file (with header) that contains the dataset to be used for validation.

  • --metrics: list of metrics to be calculated on the training dataset and, if present, also on the validation dataset; the list of supported metrics is defined in XGBoost Parameters. under eval_metric.

  • --dumpout: path (relative or absolute) of the directory where to save the metric values as the ages go by; the program dumps_scatter.py will use the contents of this directory to display the metric graphs.

  • --earlystop: how many iterations can be performed before the algorithm begins to enter the overfitting phase.

  • --xgbparams: list of parameters to pass to XGBoost regression algorithm; see documentation of XGBRegressor.

Prediction and error calculation

In this chapter the program predict_func_miso.py is presented and whose purpose is to make predictions on a test dataset applying it to a previously trained XGBoost regressor model via the program fit_func_miso.py, always without having to write code but only through the command line.
In fact, this program supports arguments through which the user passes the previously trained model, the test dataset and the error measurements to be calculated between the predictions and the true values.
The format of the incoming test datasets is identical to the one of the program fit_func_miso.py; of course here the last column (the value of the dependent variable) is only used to compare the predicted values with the true values by calculating passed error measurements.

Usage of the predict_func_miso.py program

To get the program usage you can run this following command:

$ python predict_func_miso.py --help
and the output got is:

usage: predict_func_miso.py [-h] [--version] --model MODEL_FILE --ds
                            DF_PREDICTION --predictionout PREDICTION_DATA_FILE
                            [--measures MEASURES [MEASURES ...]]

predict_func_miso.py makes prediction of the values of a multiple-input
single-output (scalar) function with a pretrained XGBoost model

optional arguments:
  -h, --help            show this help message and exit
  --version             show program's version number and exit
  --model MODEL_FILE    model file
  --ds DF_PREDICTION    dataset file (csv format)
  --predictionout PREDICTION_DATA_FILE
                        prediction data file (csv format)
  --measures MEASURES [MEASURES ...]
                        List of built-in sklearn regression metrics to compare
                        prediction with input dataset
Where:
  • -h, --help: shows the usage of the program and ends the execution.

  • --version: shows the version of the program and ends the execution.

  • --model: path (relative or absolute) to the file in joblib (.jl) format of the model generated by fit_func_miso.py.

  • --ds: path (relative or absolute) of the csv file (with header) that contains the input dataset on which to calculate the prediction.

  • --predictionout: path (relative or absolute) of the csv file to generate that will contain the prediction, that is the approximation of the function applied to the input dataset.

  • --measures: list of measurements to be calculated by comparing the true values of the input dataset and the predicted output values; the list of supported metrics is defined in SciKit Learn Regression Metrics.

An example of using of all the programs

Suppose you want to approximate the function $$f(x)=x sin \frac{1}{x}$$ in the range $[-0.4,0.4]$ Keeping in mind that np is the alias of NumPy library, the translation of this function in lambda body Python syntax is:

x * np.sin(1 / x)
To generate the training dataset, run the following command:

$ python fx_gen.py \
  --dsout mytrain.csv \
  --funcx "x * np.sin(1 / x)" \
  --xbegin -0.4 \
  --xend 0.4 \
  --xstep 0.00031
instead to generate the test dataset, run the following command:

$ python fx_gen.py \
  --dsout mytrain.csv \
  --funcx "x * np.sin(1 / x)" \
  --xbegin -0.4 \
  --xend 0.4 \
  --xstep 0.00073
Note that the discretization step of the test dataset is larger than that of training and it is a normal fact because the training, to be accurate, it must be run on more data. Also note that it is appropriate for the discretization step of the test dataset is not a multiple of the training one in order to ensure that the test dataset contains most of the data that is not present in training dataset, and this makes prediction more interesting.
To this we intend to make a regression by fit_func_miso.py passing to the underlying regressor: n_estimators: 100, max_depth: 7; then run the following command:

$ python fit_func_miso.py \
  --trainds mytrain.csv \
  --modelout mymodel.jl \
  --xgbparams "'n_estimators': 100, 'max_depth': 7" 
and at the end of the execution the saved mymodel.jl file contains the model of the XGBoost regressor configured and trained.

Now we intend to perform the prediction and calculation of the error using the measurements mean_absolute_error and mean_squared_error; then execute the following command:

$ python predict_func_miso.py \
  --model mymodel.jl \
  --ds mytest.csv \
  --predictionout mypred.csv \
  --measures mean_absolute_error mean_squared_error
and at the end of the execution the saved mypred.csv file contains the prediction performed by applying the model on the test data; the output of the program displays the error measures passed through the argument --measures and they are very small: the first one around $1.5 \cdot 10^{-3}$ and the second one around $5.5 \cdot 10^{-6}$
Note: Given the stochastic nature of the training phase, your specific results may vary. Consider running the training phase a few times.
Finally you want to make the comparative display of the test dataset with the prediction; therefore run the following command:

$ python fx_scatter.py \
  --ds mytest.csv \
  --prediction mypred.csv \
  --title "XGBoost (estimators: 100, max depth: 7)" \
  --xlabel "x" \
  --ylabel "y=x sin(1/x)"
that shows the dispersion graphs of the test dataset and the superimposed prediction: in blue the one of the test dataset, in red the prediction. The comparison of the two graphs clearly shows that the approximation has reached very high levels, as the low error measurements already indicated; only around zero is some inaccuracy observed and is due to the oscillation of the function that is there very strong.

Note: Given the stochastic nature of the training phase, your specific results may vary. Consider running the training phase a few times.

Figure with dispersion graphs generated by the program fx_scatter.py showing the fitting in red overlay of the function $f(x)=x sin \frac{1}{x}$ and the original function below in blue.

In the folders one-variable-function/examples and two-variables-function/examples there are shell scripts that show the use of these programs in cascade respectively for datasets generated with real-valued scalar functions of a real-valued variable and for datasets generated with real-valued scalar functions of two real-valued variables.


Download of the complete code

The complete code is available at GitHub.
These materials are distributed under MIT license; feel free to use, share, fork and adapt these materials as you see fit.
Also please feel free to submit pull-requests and bug-reports to this GitHub repository or contact me on my social media channels available on the top right corner of this page.