EVALUATING SAMPLING METHODS FOR REUSING KNOWLEDGE FROM LARGE AND ILL-STRUCTURED QUALITATIVE DATA SETS

摘要

The desire to use ever growing qualitative data sets of user generated content in the engineering design process in a computationally effective manner makes it increasingly necessary to draw representative samples. This work investigated the ability of alternative sampling algorithms to draw samples with conformance to characteristics of the original data set. Sampling methods investigated included: random sampling, interval sampling, fixed-increment (or systematic) sampling method, and stratified sampling. Data collected through the Vehicle Owner's Questionnaire, a survey administered by the U.S. National Highway Traffic Safety Administration, is used as a case study throughout this paper. The paper demonstrates that existing statistical methods may be used to evaluate goodness of fit for samples drawn from large bodies of qualitative data. Evaluation of goodness of fit not only provides confidence that a sample is representative of the data set from which it is drawn, but also yields valuable realtime feedback during the sampling process. This investigation revealed two interesting and counterintuitive trends in sampling algorithm performance. The first is that larger sample sizes do not necessarily lead to improved goodness of fit. The second is that depending on the details of implementation, data cleansing may degrade performance of data sampling algorithms rather than improving it. This work illustrates the importance of aligning sampling procedures to data structures and validating the conformance of samples to characteristics of the larger data set to avoid drawing erroneous conclusions based on unexpectedly biased samples of data.