| 摘要: | 市場區隔是企業根據客戶所需要的產品或服務，將市場區割成不同群組， 使得每個群組內之消費者均有類似需求或購買行為。對個別區隔市場設計專屬的行銷組合，更能有效運用企業的行銷資源，企業更能做好行銷工作。一個合適的市場行銷策略，可以應用到各個領域的客戶，以增加銷售量。這個過程可由四個活動，即市場區隔的顧客辨識，了解該市場區隔的顧客輪廓及特徵，評價該市場區隔的顧客和目標市場的策略及資源配置。市場區隔通常採用集群技術，使得每個群組內之消費者均有類似購買行為、需求或喜好。這種資訊可幫助企業做選擇性營銷的決策過程，如目錄設計，交叉銷售，或安排他們的貨架空間等可能會導致銷售量的增加。增加銷售不僅只是削減價格，若根據客戶需要和喜好，提供一定的數量之熱門項目的混搭銷售，也是一種受到客戶高度讚賞的促銷方式。類似品項及其不同數量的組合包裝是行銷重要策略之一，例如可從一組合為2瓶A牌酒及2包B牌菸或另一組合為2瓶C牌酒及2包D牌菸去選購，符合消費者多樣化需求。但是以往的促銷組合包裝通常來自設計者喜好或觀念，如何從日常大量交易中自動有效得到受顧客歡迎的類似品項及其不同數量的組合包裝是本研究主要課題。 但分析連鎖零售商店擷取的交易資料集是一大挑戰，因為經常包含稀疏的數據，係因交易資料由大量的產品組成，每筆交易包含極少量產品項目，類似稀疏資料集，導致每個項目只出現於小部分的交易資料。以購物籃技術分析的結果，通常得到極端低項目支持度。 資料關聯技術之購物籃分析大都以品項來區別哪些產品客戶會一起購買，有 時因忽略物件階層式架構關係產生相似度的問題，而嘗試將品項提升至類別來增加支持度而產生關聯規則，亦不考慮購買數量，但如此亦失去了可能更明細的關聯品項的決策性資訊。而資料分群技術主要功能則可利用成員的相似度將群集間之差異及群內相似性找出來。 但大部分的研究專注在分群技術及相似度衡量上，忽略了物件間階層式架構 的概念及品項購買數量問題，例如實務上從最終消費者的觀點來看產品或物件相似度,其實是有不同意義的,而且在零售業眾多的產品類別中，實務上也幾乎不可能為平衡式之階層架構。 固本研究提出一個架構來衡量非平衡式之階層架構物件之相似度，從消費者觀點出發，貼近人類直覺，並考慮品項購買數量問題，據以分群，協助分群意義具體化，對決策者提供有用的資訊。 Market segmentation is an important marketing process for enterprises to identify and group customers according to the products or services they need so that suitable market stimuli can be applied to each individual segment of customers to increase sales volumes. The process is composed of four activities, namely, segment identification, segment characterization, segment evaluation and target segment evaluation. Segment identification is usually performed with clustering, which groups customers with similar transactions to the same segment. Such information can lead to increased sales by helping retailers do selective marketing and can help in many business decision-making processes, such as catalog design, cross-marketing and arrange their shelf space. A mixture of popular items with certain quantities in a package according to needs and preferences from customer is a highly appreciated promotion approach to increase sales instead of just cutting price. What popular items and how many quantities supposed to be packed together should be emerged from customer buying behaviors instead of from designers’ perspective. Performing segment identification from transaction data is difficult because a typical retailer usually carries tens of thousands of goods whereas a transaction typically contains less than a hundred items. Besides the goods that are purchased, the quantities that are consumed also play an important role in distinguishing customers. To reduce the issues of low cardinality and high intra distancing of transaction clustering, the majority of shopping basket analysis attempted to import a product hierarchy to a higher concept level, or aggregate transactions to a customer level to alleviate the sparcity from transactions. However, among the enormous volume of retail products, it is almost impossible to have a balanced hierarchical structure. In empirical practices, similarity in the hierarchy from the perspective of consumers (bottom-up) is also quite different from designers (top-down). Aggregate transactions to a higher concept level may also lose some detail information. Previous studies have seldom been found to have applied a combined quantity and similarity concept from a bottom-up perspective on an unbalanced tree to clustering transactions. This study presents an algorithm for mining quantitative similar clusters via an improved clustering algorithm which tracks top k clusters with its own quality of intra-similarity. This algorithm is based on a QSKM (Quantity Sensitive kth matched similar pair) similarity measures which we derived for transactions with purchased quantities using an unbalanced hierarchical structure from a consumer’s perspective. From our experiments, we found that QSKM measures outweighed traditional similarity measures in finding the clusters of similar products and quantities purchased together from a real-life transaction database, and also discovered up to 5 clusters with enough coverage from sparse data which cannot be discovered by using traditional similarity measure or frequent patterns. Besides, the cluster intra-similarities are better than GCS (General Cosine Similarity). |
