Neural Network and Multivariate Analyses: Pattern Recognition in Academic and Social Research

Abstract

Neural networks are the modern tools that focus most heavily on the logical structure of measurement/assessment, as well as the actual results we attempt to identify by way of scientific inquiry. Employing the Self-Organizing Map (SOM) neural network, we reexamined a well-recognized and commonly employed dataset from a popular applied multivariate statistics text by Stevens (2009). Using this textbook dataset as an exemplar, we provide a preliminary guide to neural networking approaches to the analysis of behavioral outcomes. When employing conventional multivariate procedures only, the sample dataset demonstrated significant familywise error rates; however, these outcomes did not provide sufficient information for identifying the curvilinear patterns that existed within these records. When converted to natural logs and reanalyzed by the SOM, the exemplar dataset showed the actual best fit performance patterns exhibited by all members of the experimental and control groups. The SOM and related neural network algorithms appear to have unique potential in the recognition of nonlinear but unified data patterns frequently exhibited within academic and social outcomes. In particular, the SOM allows the researcher to conduct a "finer grain" analysis identifying critically important similarities and differences that can inform treatment well beyond the probability values derived from conventional statistical techniques.

The Emergence of Stimulus Relations: Human and Computer Learning

“…During the last 30 years, neural network technologies have become increasingly pervasive throughout the scientific and academic world. To list but a small fraction of the available examples, computer learning has been applied successfully to the understanding and prediction of environmental changes (e.g., Allamehzadeh & Mokhtari, 2003), the prediction of financial crises (e.g., Arciniegas, Daniel, & Embrechts, 2001;Erdal & Ekinci, 2013), and the identification and prediction of medical and physiological conditions (e.g., Abbass, 2002;Huang et al, 2013;Ninness et al, 2012;Wolberg, 1992;You & Rumbe, 2010). Precise forecasts have been demonstrated when attempting to predict severe weather patterns (e.g., Knutti, Stocker, Joos, & Plattner, 2003;Maqsood, Khan, & Abraham, 2004), predict the demand for electricity during severe weather conditions (e.g., Khan & Ondrusek, 2000;Oğcu, Demirel, & Zaim, 2012), identify probable terrorist locations (e.g., Guo, Liao, & Morgan, 2007), and classify and predict voting patterns (e.g., Ninness et al, 2012).…”

“…Within the area of instructional technology, there has recently been a renaissance of applications aimed at enhancing educational interventions. To list but a very few examples, neural networking systems have been applied to classifying and predicting student math skills during interactive training (e.g., Ninness et al, 2005), developing student learning models and predicting knowledge (e.g., Desmarais, Meshkinfam, & Gagnon, 2006), adapting computer student assessment and prediction (e.g., Desmarais & Pu, 2005), classifying and predicting student musical skills (e.g., Ninness, et al, 2013), developing instructional systems providing concurrent assessment and tutoring (e.g., Feng, Heffernan, & Koedinger, 2009), developing online personalized learning systems (e.g., Heller, Steiner, Hockemeyer, & Albert, 2006), and analyzing online cognitive tutors (e.g., Aleven, 2013;Koedinger, Corbett, & Perfetti, 2012).…”

“…One such source might be afforded by means of computational modeling" (p. 15). In fact, there has been considerable emphasis placed on exploring how artificial neural networks in the form of CMs might further our understanding of derived stimulus relations (e.g., Cullinan, Barnes, Hampson, & Lyddy, 1994;Lyddy & Barnes-Holmes, 2007;Lyddy, Barnes-Holmes, & Hampson, 2001;Tovar & Chavez, 2012;Vernucio & Debert, 2016) as well as neural network applications focusing on highly specialized problems within behavior analysis and related fields (e.g., Burgos, 2007;Donahoe & Burgos, 2000;Ninness, Henderson, Ninness, & Halle, 2015;Ninness et al, 2012Ninness et al, , 2013.…”

“…Consistent with the acquisition of new skills by human participants, before a CM can solve novel problems, it must undergo a series of supervised training trials. As described by Ninness et al (2013), "neural networks must undergo a series of training sessions with training values in conjunction with expected output patterns before they are capable of providing solutions to problems" (p. 52). During this series of training trials, referred to as epochs, an algorithm known as feedforward backpropagation (Rumelhart, Hinton, & Williams, 1986) compares the CM's calculated values with the target values and identifies the magnitude and direction of difference.…”

Probability Pyramiding Revisited: Univariate, Multivariate, and Neural Network Analyses of Complex Data

“…As discussed by Ninness et al (2013), in the behavioral and related sciences, where the number of multivariate, nonindependent, and nonlinear variables are continually rising, the sheer volume of new types of academic measurements is almost overwhelming (James, 1985). The exasperating, yet unavoidable, fact is that an increasing part of the data we collect in an effort to answer our complex academic questions have become a substantial part of those very questions (Gigerenzer, 2004).…”