You may have noticed there are very frequent new releases of Tiberius. This is because we are constantly developing new features and also responding to requests from our users.
We also guarantee Tiberius to never have a known bug - if you can find any then they will be fixed and a new release made available immediately.
Please contact us if you have any suggestions for improving Tiberius.
On the Cards...
- Let us know your needs!
- Neural Net models can now be exported to Excel 2010+
- SQL code generated from the Scorecard Builder algorithm now compatible with Teradata
- The options menu now allows specification of the database 'field name brackets' for compatibility with more databases
- Closer integration with SQL Sever
- Ensemble Optimiser
- Many minor small changes made during the last 18 months!
- Now only need to log into a database (Oracle, Teradata etc.) once during a Tiberius session - no need to re enter password if changing tables.
- SVM now works for those locales where '.' is not the decimal point.
- Data 'Sort Order' correctly remembered between sessions if more than one variable to sort by.
- Modification of sql code generated for MS SQL Server for NN and logistic regression ensembles (to overcome a bug in MS SQL Server which fails to parse the sql if there are more than 500 variables in the model). Classifiers >> Neural Network / Logistic Regression
- We used this ensembling in the KKDCup ongoing competition.
- Individual neuron weights can now be manually frozen (graph type>>weights : right click on the top weight).
- Exporting of neural network models to Excel 2007 now possible (Tools>>Export Model).
- Filtering, Sorting and Sampling of data prior to loading - only load the data you want from your database (Inputs screen)
- Loading of data speed increase
- Several other speed enhancements
- Data can now be loaded from Teradata
- Favourites list added for Teradata and Oracle tables (tools>>database favourites)
- Decision tree rings can now be saved as an image (classifiers>>decision tree)
- Classifier input list can now be filtered by descision tree variables(classifiers>>decision tree)
- %class v %population now added as a standard plot in the model monitor. (classifiers>>model monitor (gini curves))
- Variables can be randomly omitted from models when building ensembles of Neural Nets or Logistic Regressions. (classifiers>>Neural Net / Logistic Regression)
- Scorecard Builder Auto Build algorithm speed up (classifiers>>Scorecard Builder)
- The 'ranking' tab of the variable ranker and the 'importance' tab of the classifier algorithms can now be used to reduce the initial candidate variables in a model (classifiers>>).
- 'Scorecard Monitor' renamed 'Model Monitor' (classifiers>>model monitor)
- Options to generate just the requested outputs in the Model Monitor (classifiers>>model monitor>>plots)
- Model Monitor can now linearly calibrate scores that represent probabilities or odds, similarly to the log odds calibration (classifiers>>model monitor>>odds)
- If your score is a probability, then the scorecard monitor plots the expected line to compare with the actual. Calibration similar to log odds will be added soon. (classifiers>>scorecard monitor>>odds>>probabilities)
- Option added to disable the message box that appears on the completion of a batch run. This means automation made easier for coders (options>>clues)
- Made more 'Vista' friendly.
- Bug fixed with opening saved weights files if locale set to Italian.
- More frequent screen refreshing during loading (so with thousands of fields it doesn't appear as if Tiberius has frozen).
- Calculating the variable importance in the NN,SVM and Logistic Regression classifier algorithms is now optional as with many fields it can save time.
- Status bar added to Classifier algorithms to display available memory and elapsed time.
- Classifier algorithms generated sql scoring code slightly modified to better deal with date fields.
- SPSS scoring script now generated for the data recoder (data>>data recoding) and logistic regression (classifiers>>logistic regression) and neural network (classifiers>>neural net) classifiers.
- In the classifiers algorithms, the distribution of the categorical variables can now be copied to the clipboard.
- Rules that give a specified lift are now displayed automatically on the decision tree. Adjust the central lift gauge to the lift required to see the rules.
- VB.NET and C# added as an extra format for the neural network scoring code (Tools>>Export Model).
- VB scoring code added for Support Vector Machine (Classifiers>>Support Vector Machine).
- PHP added as an extra format for the neural network scoring code (Tools>>Export Model).
- The traning range, or random training percent, is now stored in the weights file for the main neural network. The random selections will be different each time the model is loaded, which is useful for building ensembles utilising the batch training feature.
- There is now the ability to set the error being minimised while creating a batch file.
- The ability to build ensembles of models (a type of Bagging, or Bootstrap AGGregatING) has been added to the neural and logistic regression classifier algorithms. See Classifiers>>Neural/Logistic.
- For the classifier models, the data behind the lift charts on the 'Gini' tab is now available in a grid that can be copied and pasted.
- For the classifier models, the lift charts on the 'Gini' tab were calculating the decile lift based on the lower valued class. This has now been changed to the class of the higher value (ie 1's instead of 0's).
- For the classifier models, the generated scoring code may create extra fields based on the original field name by adding a prefix to it. This can be problamatic for SASŪ code if the original field name is long. There is now an option ('Minimum generated field name length') to restrict this prefix to 2 characters ('?_'). Thus any generated code should execute in SASŪ as long as the original field names have 30 characters or less.
- The 'error being minimised' is now remembered when a model is reloaded.
- Excel2007 (*.xlsx,*.xlsm,*.xlsb) and Access2007 (*.accdb) added to the list of default data types that can be opened. If you want to open one of these files but do not have Office2007 then you will need to install the microsoft office 2007 drivers. If you have Excel2007 and want to load more than 255 columns then we'll have to wait for Microsoft to catch up with themselves ;-).
- The R2 statistic is now displayed on the Graph Type 'X-Y' so this error metric can be seen for the models during the training process.
- Both the R2 and correlation (=R) statistics are now displayed on the Graph Type 'xy1'.
- A new button 'Background' on the data summary form (Data>>Data Summary) that enables the plot background colour to be changed.
- A new button 'Reload' in the main neural network 'Variable Importance' form. This will reload the data with just those variables that are selected, enabling easy elimination of useless data.
- New batch model building abilities under 'tools' drop down menu.
- Utilities folder installed containing Excel file with demo code that demonstrates how to control Tiberius from external applications.