How can we easily generate latent variables in BayesiaLab?
Quote 0 0
A Latent variable (or Factor) is a variable that is not contained in the original dataset. It is also frequently referred to as a Hidden variable. Such nodes are always shown in white in BayesiaLab.You can manually create Latent variables can by adding a node to the current network, and then defining its number of states. Furthermore, you will then need to manually define the connections with the other variables and use either Probabilities Learning (Learning | Parameter Estimation), which uses the Expectation-Maximization (EM) algorithm, or manually define the distributions via the Node Editor.The other way to create Latent variables is to use the BayesiaLab's automatic Data Clustering algorithms (Learning | Clustering) on a selected set of nodes:[* odppaqu]Data Clustering: the objective of this EM algorithm is to create a Latent variable that summarize the joint probability distribution defined by the selected nodes. This algorithm can be used to search the optimal number of states.[/*:m odppaqu][* odppaqu]K-Means: this EM algorithm consists in partitioning the (numerical) observations corresponding to the selected nodes into K clusters in which each observation belongs to the cluster with the nearest center/mean. The number of centers K is defined by the user. We have illustrated this algorithm for [url=https://forums.bayesialab.com/viewtopic.php?f=10&t=34 odppaqu]data discretization[/url odppaqu].[/*:m odppaqu][* odppaqu]Binary Clustering: this algorithm is a deterministic tool for generating binary (or boolean) latent variables by applying a deterministic function on a set of selected nodes. [/*:m odppaqu]The Multiple Clustering tools can also be used for creating Latent variables per node Class. To be used for the automatic induction of Latent variables, such Classes of nodes must follow the format, [Factor_i]. Classes can be set manually, e.g. utilizing expert knowledge, via the Class Editor in the Contextual Menu. Alternatively, Classes can be automatically defined using BayesiaLab's Variable Clustering tool.
Quote 0 0