[[alternative]]有效率的探勘客戶消費行為方法的研究

博士[[abstract]]資料豐富的資料庫在數位化之後已經普遍產生，如何從資料庫中挖掘重要的資訊是資料探勘的主要任務。在商業活動的應用中，我們可以從交易資料庫中分析常常一併購買的商品以及顧客在購買某些商品之後，可能也會購買其它商品的關聯行為，也就是探勘頻繁項目集，在論文中，我們提出很有效率的探勘頻繁項目集演算法，不論在執行時間或記憶體的使用量上都優於之前的研究。 然而，新的交易資料會不斷產生，而舊有的交易資料必須被移除，若重新探勘原始資料，則會浪費時間重新找尋已知的資訊。隨時間不斷產生新資料與移除舊資料的環境稱之為資料流。因此在資料流的環境下，找出所有頻繁項目集開始被學者提出研究。另外，交易資料中的頻繁項目集可能非常多，眾多資訊會造成困擾，以致無法做決策。因此學者提出封閉項目集。若項目集的支持度比其所有超集合的支持度大，則此項目集稱為封閉項目集。由頻繁封閉項目集可衍生全部的頻繁項目集。我們也提出有效率的演算法，當資料不斷新增或被移除時，原始資料庫不需要被重新讀取，只需將新增或被移除的資料與舊有的封閉項目集做運算，就可產生更新後的封閉項目集。 耗材性商品通常在所有商品中十分經常被購買，雖然單獨的獲利可能並沒有家電、電子商品這麼高，但是累積後的獲利卻不是小數值，所以若是能針對耗材性商品掌握正確的商機促銷，對於獲取重大利潤將有很大的幫助，而頻繁項目集無法提供促銷時機的資訊。因此，我們提出一個新穎的資料探勘方式，針對於某種耗材性商品，找出不同特徵的客戶對此商品的消費行為，根據客戶的背景屬性值以及此次購買某商品的數量，我們可以利用此消費行為，正確的預測出此顧客何時會再需要此商品，以掌握行銷此商品給此客戶的時機。[[abstract]]Mining frequent itemsets is to discover the groups of items appearing always together excess of a user specified threshold. Many approaches have been proposed for mining frequent itemsets by applying the FP-tree structure to improve the efficiency of the FP-Growth algorithm which needs to recursively construct sub-trees. Although these approaches do not need to recursively construct many sub-trees, they also suffer the problem of a large search space, such that the performances for the previous approaches degrade when the database is massive or the threshold for mining frequent itemsets is low. In order to reduce the search space and speed up the mining process, we propose an efficient algorithm for mining frequent itemsets based on frequent pattern tree. Our algorithm generates a sub-tree for each frequent item and then generates candidates in batch from this sub-tree. For each candidate generation, our algorithm only generates a small set of candidates, which can significantly reduce the search space. However, there may be many frequent itemsets existing in a transaction database, such that it is difficult to make a decision for a decision maker. Recently, mining frequent closed itemsets becomes a major research issue, since a set of the frequent closed itemsets is a condensed and complete representation of the frequent itemsets and all the frequent itemsets can be derived from the frequent closed itemsets. Because the transactions in a transaction database will grow rapidly in a short time, and some of the transactions may be antiquated. Consequently, the frequent closed itemsets may be changed due to the addition of the new transactions or the deletion of the old transactions from the transaction database. It is a challenge that how to update the previous closed itemsets when the transactions are added into or removed from the transaction database. We propose an efficient algorithm for incrementally mining closed itemsets without scanning the original database. Our algorithm updates closed itemsets by performing some operations on the previous closed itemsets and the added/deleted transactions without searching the previous closed itemsets. Compared with other commodities, consumable products are purchased high-frequently. Although single gains for consumable products may be lower than that of appliances or electronic products, the accumulative gains for consumable products are great. Therefore, grasping suitable timing to do sales promotion for consumable products is an important task. Sequential pattern mining only considers the sequential purchasing behaviors for most of the customers, but they cannot predict when the customer will need the products in the future. For the consumable products, the purchase time for the next transaction is usually related to the purchase quantities for this transaction. We propose a novel data mining algorithm to find the consumption behaviors for most of customers. From this information, we can predict the next purchased time for an item based on the purchased quantity of this item at this time.[[tableofcontents]]Content I List of Table III List of Figure V Chapter 1 Introduction 6 Chapter 2 Related Work 12 2.1 Mining Frequent Itemsets 12 2.2 Frequent Closed Itemsets in the Data Streams 13 2.3 Mining Consumption Behaviors 15 Chapter 3 Mining Frequent Itemsets 19 3.1 Preliminary 19 3.2 Search Space Reduced Algorithm 24 3.2.1. Item-prefix Tree Construction 25 3.2.2. Frequent Itemset Generation 28 3.3 Experimental Results 31 Chapter 4 Mining Closed Frequent Itemsets 44 4.1. Preliminary 44 4.2 MFCI Algorithm 46 4.2.1 Algorithm MFCI-add 50 4.2.1.1 Closed Itemset Generator 52 4.2.1.2 The Maintenance of Closed Itemsets after Adding a Transaction 57 4.2.2 Algorithm MFCI-del 68 4.3 Experimental Results 75 4.3.1 Experimental Results on Synthetic Datasets 76 4.3.2 Experimental Results on Real Datasets 85 Chapter 5 Mining Consumption Behaviors 91 5.1 Preliminary 91 5.2 Mining Consumption Association Rules 92 5.2.1 Mining Consumption Patterns 97 5.2.2 Mining Consumption Association Rules 102 5.3 Experimental Results 116 5.3.1 Synthetic Data Generations 116 5.3.2 Experimental Results 118 Chapter 6 Conclusions and Future Work 123 REFERENCES 127 List of Table Table 3 1 A Transaction Database TDB 21 Table 3 2 Frequent items and their support counts 21 Table 3 3 A sorted transaction database with minimum support 25% 21 Table 3 4 The memory usages and execution times for the five algorithms on the datasets with average transaction size T = 10 38 Table 3 5 The memory usages and execution times for the five algorithms on the datasets with average transaction size T = 20 38 Table 4 1 A Transaction Database 45 Table 4 2 Content Table for the transaction database in Table 4-1 46 Table 4 3 The Item Table for the transaction database in Table 4-1 47 Table 4 4 The Temp Table after adding T = {ACT} 55 Table 4 5 Transaction database D 63 Table 4 6 Content Table for the transaction database in Table 4-5 63 Table 4 7 Item Table for the transaction database in Table 4-5 63 Table 4 8 The Temp Table after adding T5 = {ACDTW} 64 Table 4 9 The Item Table after adding T5 = {ACDTW} 64 Table 4 10 The Content Table after adding T5 = {ACDTW} 65 Table 4 11 The Temp Table after adding T6 = {CDT} 66 Table 4 12 The Item Table after adding T6 = {CDT} 66 Table 4 13 The Content Table after adding T6 = {CDT} 67 Table 4 14 The Item Table after deleting T1 and T2 72 Table 4 15 The Content Table after deleting T1 and T2 73 Table 4 16 Parameters for IBM data generator 76 Table 5 1 A Transaction Table 94 Table 5 2 Transaction table for item A 95 Table 5 3 Sorted transaction table for item A 96 Table 5 4 Consumption table 97 Table 5 5 The status for 3 consumption intervals 101 Table 5 6 Customer Table 103 Table 5 7 Customer-Consumption Table 104 Table 5 8 Frequent 2-attribute value set 111 Table 5 9 The confidence for the frequent attribute value set 113 Table 5 10 Consumption association rules 114 Table 5 11 The parameters of the synthetic datasets 117 Table 5 12 The parameters for the first experiment 119 Table 5 13 The parameters settings for the second experiment 121 Table 5 14 The comparison of experimental results 122 List of Figure Figure 3 1 The FP-tree for Table 3-1 with minimum support 25% 21 Figure 3 2 The item-prefix tree for item D 28 Figure 3 3 Execution times for the five algorithms on dataset T10I2D100K 32 Figure 3 4 Execution times for the five algorithms on dataset T10I4D100K 33 Figure 3 5 Execution times for the five algorithms on dataset T10I6D100K 33 Figure 3 6 Execution times for the five algorithms on dataset T20I4D100K 36 Figure 3 7 Execution times for five algorithms on dataset T20I6D100K 36 Figure 3 8 Execution times for the five algorithms on dataset T20I8D100K 37 Figure 3 9 Execution times for the five algorithms 40 Figure 3 10 Execution times for algorithms on dataset T10I4D100K(item 10K) 40 Figure 3 11 Execution times for five algorithms on real dataset chess 42 Figure 3 12 Execution times for five algorithms on the real dataset Mushroom 42 Figure 4 1 Execution times for the three algorithms 78 Figure 4 2 Execution times for the three algorithms 79 Figure 4 3 Execution times for the three algorithms when 1K transactions are added and another 1K transactions are deleted 80 Figure 4 4 Execution times for the two algorithms 81 Figure 4 5 Execution times for the two algorithms 82 Figure 4 6 Number of closed itemsets deleted after 83 Figure 4 7 Execution times for the two algorithms after 84 Figure 4 8 Execution times for CFI-Stream-Modify and MFCI 85 Figure 4 9 Execution times of MFCI for adding 86 Figure 4 10 The total number of closed itemsets after adding 87 Figure 4 11 Execution times of MFCI for deleting 406 transactions each time 88 Figure 4 12 The total number of the closed itemsets 88 Figure 4 13 Execution times of MFCI for adding and 89 Figure 4 14 The total number of closed itemsets for each addition and deletion 90 Figure 5 1 Consumption intervals in Table 5-4 100 Figure 5 2 The MinRSup for consumption interval I 101 Figure 5 3 Numeric attribute 2-D projection 108 Figure 5 4 Algorithm CAR 116 Figure 5 5 The experimental results for Series A-D 120[[note]]學號: 895620093, 學年度: 10