An Empirical Study in Transaction Data Augmentation Based on Fuzzy Set Theory for Association Rule Mining

Frequent item-sets are sparse in real life transaction datasets, and data prerecession including data augmentation skills are necessary in order to carry out efficient mining. For example, the raw data (e.g. Power Mart transaction data set) contained the data field of price (Numerical), where a keyboard of price USD ${\$}$29.1 was mapped to price level of low to middle (e.g. “L2M” = 1), and product type of “KB”. Adopting the mapping for association rule mining, the top frequent keyboards of low to middle price level generated had a very low support = 0.12. In this work, we adopted the proposed data augmentation to map a numerical price to fuzzy price levels, where the keyboard of price USD ${\$}$29.1 was described by the different price levels at the interval [0, 1] (e.g. “LOW”= 0.0, “L2M” = 1.0, “MID” = 0.2, “M2H” = 0.0, and “HIG” =0.0). The advantages of making use of fuzzy set theory lied in two aspects: firstly the ambiguity of the vague boundaries among different price levels were handled, and secondly the experimental results showed the effectiveness of the fuzzy logic (e.g. much larger sets of frequent item-sets were generated, and more association rules could be mined in data analysis). The experimental results showed that the no. of frequent item sets and no. of association rules both increased greatly. The trapezoidal and linear membership functions were chosen, based on market research, objective preference, as well as the Power Mart transaction data set.