Pattern Recognition Approach to Data Interpretation

Scientific research has become an area of enormous data production. Reams of data are routinely generated as the trend towards quantification within the sciences has increased. The need for sound mathematical methods for analyzing these data is crucial. Often data are continually produced without stopping for such analyses. The result can be the production of large amounts of inferior data. Studying the mathematical patterns underlying the data can often help to determine the best next step in the analysis and to draw meaningful conclusions from the data already gathered. Moreover, such studies may reveal that better experimental designs can be devised and implemented effectively. Also, underlying properties of the data, not directly measurable, but related to the data being produced, may be studied, and predictions related to the scientific content of the data, and future data, become possible.

Specific techniques from the three computer packages—SPSS, BMDP, and ARTHUR—will be discussed in this chapter. A single, comprehensive problem will be tackled, with the goal being a thorough, statistical analysis of the data base. A reasonably complex example was chosen, to illustrate both the results as well as the problems encountered in a pattern recognition study. The statistical evaluation of the data base used for the examples was performed in the authors’ laboratory. Studies were made on the Maricopa County Arizona Health Department air pollution data base. Samples were taken at six-day intervals from January 1975 through December of 1977. Four monitoring stations were used, designated as MC (Maricopa County), MS (Mesa), NS (North Scottsdale), and SC (Scottsdale). All four are located within the Metropolitan Phoenix area with NS representing the most “pristine” area from a pollution standpoint. Only data from days where measurements at all four stations were collected are included in the study. Variables (and their abbreviations to be used) in the study were 24 high volume collection samples analyzed for manganese (MN), copper (CU), lead (PB), total particulates (PART), total organics (ORG), sulfates (SUL), nitrates (NIT), and chlorides (CL). Carbon monoxide (CO) data were also collected. The above abbreviations were used in the programs and will be utilized throughout the text.

In this chapter, the “how to do it” information will be given. Enough detail will be provided to enable the beginner to obtain results. This chapter is to be used in conjunction with details in the SPSS, BMDP, and ARTHUR manuals. It will be assumed that the user knows how to punch computer cards, or code information into the CRT, or other type of computer input device. The user must have available the information on how to access the computer to be used for the analysis. This information can be obtained at the user’s computing facility. The user must find out how to assign SPSS, BMDP, and ARTHUR at his given facility. This chapter will be concerned with the cards or commands necessary to complete successfully runs for the three programs after they have been assigned. Throughout this chapter, we will discuss only computer cards as the input medium. Other methods will utilize similar package commands. See the manual for details.

The possible applications of pattern recognition techniques to the natural sciences are limited only by the human imagination. All facets of science are generating reams of data which increases the need for more sophisticated methods of data reduction and analysis. The possibilities are infinite. Therefore, it is not the purpose of this chapter to give an exhaustive list of approaches and applications, but rather to give an idea of the possibilities that exist, and some insight into reference source materials.