| Summary: | 碩士 === 國立成功大學 === 資訊管理研究所 === 101 === Businesses usually use data mining or variety of techniques to analyze the likely next step of customer’s behavior or trend in order to get the higher satisfaction and also increase self-profits. Therefore, the capability of data analysis is quite important in today’s situation. However, the data type may come in fuzzy which cannot be solved in traditional mathematics. Until Zadeh(1965) proposed fuzzy theory the questions finally saw the daylight. Fuzzy theory is now used in so many fields and fuzzy time series is one of them.
Time series usually oscillate between likely trends in nature. Although high order fuzzy logical relationship may capture trends in time series, it still cannot precisely predict the situation which cannot find the same FLR in training data. Even if we can find the same FLR, it doesn’t mean that the linguistic class will be the same. Therefore, in the steps of fuzzification and rule establishment we not only use the larger membership degree be a linguistic class, but also take the smaller one into consideration and separate it into major and minor linguistic class. Finally we use KNN method to search the similar FLR.
Although high order FLR may capture the likely trends, sometimes the best order which is quite large will lead to inconvenient in practice. In order to use the shortened order but also keep the high accuracy, we proposed the concept of minimum recent orders (MRO) in second stage. We use the uniqueness of LHS and consistency of RHS to find the MRO in each record in the training data to early predict the result.
|
|---|
