• Mainwindow
• Statistics
• Neuronal Network - Regression
• Neuronal Network - Cluster
• Support Vector Machine - Regression
• Support Vector Machine - Cluster

Set the environment variable SADATO

Before you can start you have to define the variable SADATO in the .bashrc. This variable indicates to the directories:
such as: csv, databases, help_html, models

Open the file .bashrc with an editor (vi or emacs or a.o.) and add the following instruction:

• export SADATO=/home/user/path

Main window

The input files are often simple text files (.csv, .txt). SADATO will convert it to a unique database table. This helps to handle all the inputs in a uniform way. A text file should have 10 000 rows maximal.
You create a database by clicking , the database will be opened with a dummy table inside named 'last_mohican', witch is deletable if you have a second table imported. Only one database can be opened at the same time.
An existing database will be opened by clicking . When a database is open you can select a table and in the table the columns for modeling. To select rows a where clause is also supported. The where clause means only the condition of a where part of a SQL-statement. For example the where field "year>=2000" will select all data with the year>=2000.
Use the right mouse push button to delete a complete table from the database. You can also delete some rows from data table. Additionally data values are editable. After editing don't forget to save the database table or with a new tablename .

• Statistics
• Neuronal Network - Regression
• Neuronal Network - Cluster
• Support Vector Machine - Regression
• Support Vector Machine - Cluster

Statistics

Explanatory statistics should be used as a starting point for modeling. Here you can see results of several statistic calculations for your selected data set, such as

• Summarize (min, max, mean, std dev, variance)
• Correlation of each column with each other
• Multiple linear Regression
  It's possible to save the coefficients of the model in order to load them for applying with your data set.
• Principal Component Analysis (PCA)

Neuronal Network (NN) Regression

The neural network based on the FANN (http://leenissen.dk/fann/). The number of hidden layers can be 1,2,or 3. Different training algorithms are available. The activation function type for the hidden and output layer can be selected and some training parameters can be changed. This should be done during the training session to optimize the results. A very important rule plays the cross validation which in included in SADATO. The cross validation can be used to avoid over training.

There are two ways to work with a regression neural network:

1. Create a NN depending on parameters of specification, activation and training algorithm. Train the generated NN, MSE, RMSE and R² will be shown. Have a look at Result, Visualization and Cross Validation (if you have a target column in your dataset)
2. Load a trained NN Reg model. See Visualization and Apply the model on your data set (with or without target). You can save the data to a text file for use with a statistical toolbox for further analysis as example.

One important point using artificial neural networks or support vector machines in SADATO was the implementation of a visualization component to help the user to understand the models. The user can control the output and the parameter of the plot to check the results of the training and to restart the training if something seems to be wrong.

NN Cluster

Alternatively to regression problems are cluster problems. The number of cluster of the output is equal to the number of output neurons of the neural networks. SADATO can automatically estimate the number of clusters. The training result is visualized using a so called confusion matrix, which shows the desired cluster versus the modeled cluster outputs.

There are two ways to work cluster neural network:

1. Create a NN depending on parameters of specification, activation and training algorithm. Train the generated NN and have a look at Result and Cross Validation (if you have a target column in your data set)
2. Load a trained NN Cluster model. Apply the model on your data set (with or without target). You can save the data to a text file.

Support Vector Machine (SVM) Regression

Support vector machines are mathematically sound implementations of a training algorithm and based on the libsvm (http://www.csie.ntu.edu.tw/~cjlin/libsvm/). A SVM can be used for a cluster problem as well as for a regression problem. The implementation of artificial neural networks and SVM in one tool allows the direct comparison between them. Sometimes are SVM more suited for a problem other times are artificial neural networks more powerful.

There are two ways to work with a regression problem:

1. Create a SVM depending on structure and parameters. Train the generated SVM, MSE, RMSE and R² will bee shown. Have a look at Visualization. Optimization and Cross Validation are only possible, if data set has a target column
2. Load a trained SVM model. Apply the model on your data set (with or without target). You can save the data to a text file.

SVM Cluster

1. Create a SVM depending on structure and parameters. Train the generated SVM, the Confusion Matrix will bee shown. Optimization and Cross Validation are only possible, if data set has a target column
2. Load a trained SVM Cluster model. Apply the model on your data set (with or without target). You can save the data to a text file.

Back to top

SADATO is open source software (GPL V3) and under continuous development. The open source idea allows the modification of the code, the integration of SADATO in your software and the enhancement of SADATO by the user. We are thinking about the integration of the models produced by SADATO in SAMT (http://www.samt-lsa.org). An other idea is the integration of new methods of data prepossessing using wavelets in SADATO. Also under consideration is the implementation of the former neural networks toolbox "NNQT" in SADATO. The SADATO library, the basis for the GUI is under development and will be extended with a radial basis function neural network in the next weeks. SADATO itself is therefore in an early stage of its development and we will be pleased to get some response about error and wishes from the users. Please do not hesitate to write a email to: Ralf Wieland (rwieland@zalf.de) or to Karin Groth (kgroth@zalf.de).