New productivity features in VisuMap 3.0

I have just posted a short video that shows some new productivity features implemented in VisuMap version 3.0.

A layman's introduction to principal component analysis.

I have just uploaded a very short introduction for PCA to youtube. The video is probably the shortest introduction to PCA (1.5min). This introduction emphasizes the geometrical aspects, instead of the usual statistical nature. I especially like the statement that the covariance matrix is just a method to measure the average extend of an object (a point cloud) along any axis.

VisuMap 3.0 Released

Today we have released VisuMap 3.0. This is a major upgrade from the previous version 2.7.

The main change in this release is the implementation of the Atlas service. An atlas in VisuMap 3.0 is frame-less window that enables users to organize and compose data views of different types. The following picture shows the snapshot of a sample atlas:


Atlas service in VisuMap is thus similar as dashboard service in some other visualization centric software applications. It allows users to group visual information according to the subject instead of information type itself. And more importantly for VisuMap users, atlas service enables users to combine multiple algorithms to generate and compose heterogeneously structured information visualization. For instance, this release includes a script (PartitionAtlas.js) that automatically partitions a dataset in several clusters, and generates various maps for each cluster.

Together with this release we have extensive updated the scripting interface. Unfortunately, some old scripts may need minor adjustments to convert to this new release. We have also extended the data format extensively to accommodate the new service. However, the data format has kept compatible in the sense that old dataset files can be loaded into new version, but datasets of new format can not be read by previous versions.

Click the following image to watch a short video that shows how to create an atlas interactively:



Here is another video that shows how a script uses kmean algorithm to partition a dataset and generate different data views for the partitions.

This video shows using pre-configured SOG cluster to create mutiple data views for a dataset.

Data Clustering with Self-Organizing Graph

We have just released VisuMap version 2.7.832. In this release we have added a very interesting clustering service called the self-organizing graph (SOG). SOG is basically an extension to the self-ogranizing map (SOM) that simulates a homogeneous artificial neural network. Unlike SOM, SOG simulates a network of arbitrary structure. The network can be in fact any weighted undirected graph. The network is a kind of parameter for the algorithm: the user defines the network depending on his/her knowledge or interest about the dataset.

The following picture illustrates an application of SOG. On the left side is a map of a dataset that shows roughly two data point clouds. The right side shows the network we used to classify the dataset. During the learning process each node of the network will become associated with some data points. After the learning process has finished each data point will be classified as their corresponding nodes in the SOG network.


We see that the SOG algorithm not only correctly classified the two data point clouds, but also captured those data points located a the peripheries of the clouds (data points shown as blue triangles).

The following is a short video tutorial for the SOG clustering service in VisuMap:




Comparing with traditional clustering algorithms, like k-Mean and hierarchical clustering algorithms, SOG provides more flexible clustering service. With SOG the user can not only specify the number of clusters, but also the size and other geometric or topological relationships between the clusters. In fact, SOG enables the user to specify a model to reflect existing knowledge and interests about the data.

There are many variants of the self-organizing map that use irregular networks. Those algorithms often employ adaptive strategies to generate the irregular networks depending on the data. Those irregular network are results of those algorithms. Contrary to those algorithms SOG uses the network as input for the algorithm, the user can express his/her interest and knowledge about the data through the network.

SOG is our first attempt to provide a new kind of clustering service. While its framework resembles to provide general pattern matching capability. The learning algorithm it employed has only limited capability to capture structural information. In view of this there are still many challenges and potentials for SOG algorithm in the future.