OCEAN: 2 1/2D Interactive Visual Data Mining of Text Documents

OCEAN is a tool for a posteriori visual data mining that uses the output of a text miner to help users better explore a document space. Clustered documents are transformed into a hierarchical 3D representation analog to Reconfigurable Disk Trees. An intermediary document representation allows for interface customization and offers a generic approach to 3D structuring of text miner outputs. An interactive exploration is offered that avoids tree manipulation and provides local 2 1/2D view of the disks. A user evaluation is reported that outlines the benefit of controlled document hierarchy traversal and 2 1/2D document layout. 1 A Posteriori Visual Data Mining According to [1], visualization can occur before, during or after text mining algorithms. In this work, visualization takes place after the run of the text miner (TEMIS [12]). It is used to help marketing people and decision-makers at EDF Group 1 to discover the main concepts and trends in a professional area. The text miner produces a hierarchical clustering of the documents that groups similar documents into classes and subclasses. The text miner also ouptuts a dissimilarity matrix that is used by a correspondence analysis to produce a representation of these documents in a 3dimensional space. In order to help users better explore the conceptual space produced by the classifier and extract knowledge for their professional needs, we have designed an interactive visualizer with constrained navigation and assistance for knowledge discovery. The approach is illustrated on two data sets: Dw 4606 web pages grouped into 897 classes (depth=3), and Dc 1621 corporate documents grouped into 262 classes (depth=5). Despite the high number of documents, the interface offers a clear view of the data set and enables users easily discover the document space. A qualitative evalu-1 EDF Group is a European energy leader