在過去的關聯規則研究中，大部分的研究都著重在概念階層中最底層的交易內的挖掘工作。在這裡，我們將著重在挖掘跨概念階層的非重複性的交易間關聯規則。意為我們可以在不同交易間與任一層概念階層中挖掘到此種關聯規則。為避免產生大量不有趣的冗餘重複規則，我們沿用林聖傑（2005）計算概念階層中父層產品與子層產品的趣度來避免產生帶有冗餘重複資訊的關聯規則。 在零售店中並非所有產品都擁有相同或相似的特性。有些商品整年皆販售，有些商品則依照季節或特定節日銷售。為了尋找販售期間短但有趣的產品，我們將依照各產品的銷售期間的交易記錄來計算支持度。同時也根據其產品特性設定不同的門檻限制。為了提升挖掘的效率與有趣度，我們選擇在產生關聯規則前使用gap事先篩選有趣的產品來產生關聯規則。而非傳統上將所有的關聯規則產出後再進行關聯規則間的有趣度比較。 本研究提出一個以FP-tree為基礎的演算法，名為ITCL_FP-tree，結合產品銷售期間、多門檻限制與gap來採擷跨階層的非重複性交易間規聯規則。我們利用實際的資料驗證出ITCL_FP-tree在使用gap的情況下可以刪除50 %到70 %不等的冗餘重複或不有趣的規則。其決定frequent items的運算時間與傳統的演算法不相上下，但產生規則所需的運算時間則大幅減少。當資料量越大時，產生規則就越有效率。從實驗的結果中可以顯示使用gap的確可以幫助使用者更有效率地挖掘出有趣但不帶有冗餘重複資訊的交易間的關聯規則。 Most of previous studies on mining association rules are mining intra-transaction associations at the atomic level of concept hierarchy. In this study, we will mine the non-redundant inter-transaction cross-level association rules. An inter-transaction cross-level association rule describes the association relationships among different transactions and the rules among concepts at any level of a hierarchy. Additional step in pruning redundant rule is usually carried out after rules are found. However, this kind of mining may cause generating a large number of potential redundant rules. In retailing, an item may not be carried in the entire year in the shop. Therefore, mining the rules under such situations requires solving the rare item problem. Since all items in the database may not have the same natures or similar frequencies. In real-life applications, some items may appear very frequently and others may appear rarely. To find frequent items which appear rarely, we first identify the appearance period of each item, and then calculate the item’s support value. Multiple minimum support (MIS) is used to reflect the distinct nature of each item. In order to mine interesting rules and to improve the mining efficiency, we adopt the concept of gap to prune redundant and uninteresting items before rule generation rather than remove uninteresting rules after rule mining. Finally, we implement an FP-tree based algorithm, ITCL_FP-tree, on real data. Our experiment shows that we can prune out almost 50 to 70 percent of the redundant and uninteresting rules. The runtime of determining frequent items or generating rule is shorter than the one by using the traditional mining procedures even when the number of transactions is large. The result indicates that we can discover inter-transaction association rules with non-redundant knowledge.
