Visual Analysis of a Large and Noisy Dataset

1. Introduction

Visual data analysis provides a variety of techniques for scientific inquiry. These techniques include interactive exploration, graphical representation, and pictorial representation. But the use of visual data analysis as a research method is likely to produce biased results when a large and noisy dataset is investigated. This can be evident when a data analyst is not able to fully visualise the objects and associated attributes of interest.

Also, the inability of a visual apparatus to effectively depict all features of objects in a large and noisy dataset may lead to inaccurate analysis. For example, the percentage values associated with a thousand or more different attributes will not be effectively depicted in a pie chart or bar chart. More so, the inability of the data analyst to effectively visualise all the many-valued attributes in a large dataset, may lead to inaccurate analysis. For example, the data analyst may ascribe an inconsistent object as consistent or incomplete data as complete, where he did not visualise the inconsistency or incompleteness associated with the object. This work explores the classical (conventional) and non-classical FCA approaches for dealing with IID in a large dataset. It examines how the classical and non-classical FCA approaches are applied on a dataset; notably, the dataset from the e-Mouse Atlas Gene Expression Database¹ (EMAGE).

EMAGE is a free online database which stores gene expressions information in the developing mouse embryo. It acquires the gene expression information from experimental results in journal publications, screening projects, and laboratory reports among others. Such gene expressions information is annotated with terms (standardised ontology) from Gene Ontology². The gene expressions information includes detected and not detected expressions information in tissues of the different Theiler Stages in the mouse. Detected gene expression information can be described as strong, moderate, or weak while not detected gene expression information can indicate that the experiment has not been performed, the experiment has been performed but the result is not clear, or the experiment has been performed but the result is not published by EMAGE.

EMAGE provides a real-life case study for illustrating how FCA approaches can be used to visually analyse IID. Its data contain objects which are associated with many-valued attributes. For example, a tissue (object) in an EMAGE database can be assigned gene expression information (attributes), such as detected or not detected, which is associated with many values. A gene detected in a tissue at a particular Theiler Stage may be associated with strong, moderate, or weak expression information. IID can be evident in EMAGE. For example, a tissue from a particular Theiler Stage can be detected as strong in one experiment and also as weak in another experiment. Such contradictory information is evident in the EMAGE database (McLeod and Burger 2011). Also, the information in EMAGE can be incomplete because not all the experiments about every gene in every tissue of every Theiler Stage have been performed. The possible causes of IID in EMAGE are described in (Nwagwu, 2013; McLeod and Burger, 2011).

This work does not only explore the classical and non-classical FCA approaches for dealing with IID, but also describes an automated method for the mining and visualising IID, as a means of dealing with the incomplete (partial) visual analysis of a large and noisy dataset. This approach enables the data analyst to adequately verify the soundness of the information from his investigated dataset as to avoid inaccurate conclusions. It is shown in this work, how FCA can be used to mine and visualise the IID from a noisy context whose objects are associated with mutually exclusive many-valued attributes.