News Letter: November 2005

BayesiaLab 4.0: what's new

Data Sources
Defining several missing values	It is now possible to define several missing values for a database.

New database association wizard	Based on the new database import wizard, the database association wizard allows to link a database to an existing network with much flexibility. Modalities can be linked one by one or aggregated into a single modality. A discrete node can be mapped to an interval node. New nodes can also be added. Now, a warning mechanism allows to manage quickly and simply the different problems that can occur during an association. Each information is notified in a list and a double-click on it displays the convenient modality association editor in order to solve the problem.

Automatic aggregation improvements	The aggregation mechanism allows to put together modalities which have a similar behavior with respect to a target variable and a target modality. The computation process of the correlations has been modified and an automatic algorithm has been developed.

New automatic aggregation algorithm	A new automatic aggregation algorithm based on a decision tree generates the best aggregates given the target variable and its modality and the maximal number of wanted modalities.

New multiple aggregation	Based on the new automatic aggregation algorithm, a new interface has been added in order to perform automatic aggregation on several variables. These variables are chosen, among the selection, given their initial number of its modalities.

Aggregates saved	In order to keep and reuse the aggregates, they are saved in the bayesian network file.

Transferring discretization points	Once a variable manually discretized, it is possible to transfer its discretization points to other continuous variables. If the domains of the target variables are smaller than the current one, they will be enlarged in order to contain the domain of the initial variable.

Weights used in discretization and aggregation	If a Weight variable is defined, it is used by the decision tree and the equal frequencies discretization algorithms and also by the automatic aggregation algorithm of the discrete variables.

Select All buttons	In the discretization or the aggregation process, two buttons allow to select all the discrete variables or all the continuous ones in order to perform multiple aggregations or multiple discretizations.

Intermediate database save	During the import or association process, a button allows to save the intermediate database as it is already processed. This feature is very useful when the import process is long and has to be redone.

Automatic column resize	A double-click on the right side of the header of a column resizes all the selected columns to their header preferred size. This feature is available in import and association wizards.
Graphs
Graph editor	A new graphical analysis tool has been added: the graph editor. It allows to create different kinds of graphs with the variables of the network. It uses the values stored in the database. It also allows to show graphs with exact continuous values if the database has stored them during the import process. An option in the settings allows to change the behavior. The graphs are customizable by modifying their specific parameters. The graphs are interactive: you can display the data corresponding to the selected points you can zoom by selecting the area you need to magnify the coordinates of the cursor location in the graph are displayed the number of displayed points is indicated ...
Interface
Interval nodes displayed differently	The interval nodes (those that have numerical continuous values) are displayed now with a dashed border instead of the previous plain border to be immediately identified.

Node selection	It is now possible to select the nodes that are directly or indirectly connected to the current node or to select the nodes that belong to the current Markov blanket of the node (parents, children and spouses). These actions are available in the node's contextual menu.

Categories' colors list	The list of colors used by the categories (defined by the dictionary) has been modified in order to obtain a real difference between each color. The modifications of the list made by the users are saved.

Colored category tag displayed in the comment	To improve the readability of the graph, if color categories are defined, they are also displayed in the comments.

Editing node with comment displayed	Once the comments are displayed, the node can be moved, selected or edited by dragging, clicking or double-clicking directly on the comment.

Missing values and weights of a database	If a database associated to a network contains missing values and/or has a Weight variable, the bottom-right icon displays it. The green icon indicates that the database has a Weight variable and the orange icon indicates that the database contains missing values.
Node editor
Occurrence matrix	When a network has an associated database, a new view mode is available in the node editor. This view mode displays a matrix showing how many times the modalities of the node are in the database. The number of occurrences take into account the weights and the virtual occurrences, if any.

Aggregates editor	All the modalities' aliases, created when a database is opened or associated, can be edited with the button Aggregates in the node editor. It is available for discrete and interval nodes. You can add or remove names associated to the selected modality.
Monitors
Mean and standard deviation	If the continuous values have been loaded, the mean and the standard deviation displayed in the monitor of the interval nodes are computed by using these values.

Comments instead of names in monitors	You can switch the display of the node name in the monitors to the node comment through the contextual menu. The comment of the node is displayed instead of the name.

Comments in the monitor's tooltip	The comment of the node is displayed inside the tooltip of the monitors.
Learning
Learning optimization	The learning time has been divided by 2 on average. This optimization is very efficient for databases containing a lot of rows.

Imputation menu	The imputation algorithms can be changed even after import or association.

Automatic comment	At the end of the learning algorithm, a comment is automatically added to the network. This comment contains the database name, the date, the used parameters, etc.

Imputation improved	A new static imputation algorithm is now available and the dynamic imputation algorithm has been improved.

Semi-supervised learning	A new learning algorithm is now available. This semi-supervised learning algorithm allows to learn a network incrementally starting from a target node up to the the number of arcs allowed in depth. This algorithm is an extension of the algorithm that constructs the Markov Blanket. It is particularly useful for the fast learning of restricted networks even when there are hundreds or thousands of nodes. It also allows a more complete analysis than the one obtained by the Markov Blanket learning algorithm and can also allow to get a more precise prediction of the target if the variables of the Markov Blanket have missing values during the network use.

Static policy learning	The static policy learning has been totally recast. Based on Dynamic Programming, it allows now to compute the exact optimal action policies for static Bayesian networks.
Inference
Joint probability and sample number	The current joint probability of the network (ie. taking into account the hard positive and negative evidences, as well as the likelihoods) is displayed at the bottom-right of the window. If a database is associated to the network, an estimation of the corresponding number of samples is also displayed.

Analysis improvement	The efficiency of the analysis methods have been improved by restricting the analysis only to the dependent nodes.

Asking the preferred modality	During the target analysis or the batch processing, it is possible to have several modalities of the target node with the same probability. Now, instead of choosing the first modality, the algorithm asks the user to make the choice himself.
Settings
Language	The language of BayesiaLab can be changed and will be effective after restarting BayesiaLab. Japanese language support has also been added.

EM/Imputation iteration number	The iteration number used for the EM and imputation algorithms can be modified.

Structural complexity influence	The limits of the structural complexity influence have been modified.

Normalization of the weights	The normalization of the weights can be enabled or disabled and the normalization factor can be set.

Database save format	The format of the output files can be modified. It is possible to change the column separator and to specify if a dot must be used as end of line character.

Continuous exact values import	The continuous exact values can be imported when a database is loaded. This option allows to use the new graph editor to generate graphs with continuous dimensions and allows to compute the mean and the standard deviation of each interval node in the monitors.

News Letter: November 2005

BayesiaLab 4.0: what's new

Data Sources

Graphs

Interface

Node editor

Monitors

Learning

Inference

Settings