Data fusion with PARAFAC and transfer of stacked local classifiers
Date
2013
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Some data analysis methods yield pooror only adequate information on their own but with data fusion, multiple datasetscan be merged to possibly yield more information than when used alone. Data fusion can even be used to merge reduced representations of different parts of the same dataset. Data fusion yields improved results in situations where each set of datato be merged contains information unique from each other. Stacked Partial Least Squares Discriminant-Based Classification (SPLSDA) transfer attempts to use data fusion to aid in applying previous analysis to new data. Data transfer or model transfer allows for use of datasets taken under different conditions or on different instruments, ifthis variation can be accounted for. SPLSDA transfer is based upon a previously developed classification transfer approach which uses a reduced dimensional representation for each different section of the data, in order to classify new samples taken under different conditions. A similar method is Interval Partial Least Squares (IPLS) with the exception that these intervals collectively cover all of the data. Only some of the data in each section is used as much of the data contains veryredundant information or is uninformative which can hinder the classification model. The purpose of SPLSDA transfer is for transferring new infrared samples into an existent database, which were collected on a different instrument. Classification models can be fused from the infrared spectra of new samples and converted to allow classification using an existing database. Another method involving data fusion is Parallel Factor Analysis (PARAFAC), which can be used to analyze multiple datasets simultaneously to find the causes of underlying variation in the sample. PARAFAC is used here to determine the concentration of three specific compounds found in a growth medium along with other unknown compounds.