| Summary: | 碩士 === 國立台中師範學院 === 教育測驗統計研究所 === 89 === The purpose of this study is to explore how to implement pattern recognition techniques to simulated assessment data. There are two types data which are used in this study. One is the raw binary data and the other is the data estimated by IRT.
Since the dimensionality of assessment data usually is up to 20 or 30. There are two steps to build a high dimensional classification procedure. The first is feature extraction and second one is classifier selection. In feature extraction, DAFE, DAFE based on LOOC, and Rule Space are compared in this study. In classifier selection, linear classifier, quadratic classifier, and quadratic classifier based on LOOC are used.
The conclusions are
1. The best accuracies of simulated data sets are over 90%. This result is acceptable for teachers.
2. In feature extraction part, DAFE based on LOOC has the best performance when data set is singular.
3. In classifier selection part, quadratic classifier based on LOOC is the best choice.
4. LOOC can improve the performance on both singular (number of items is greater than the class sample size.) and nonsingular situations.
5. Binary data and IRT data have their own advantages. Anther experiment is needed to decide which one is better.
|
|---|
