運用層級貝氏定理建立顧客購賣期間預測模型， 以 3C 賣場顧客為例

Abstract

近年來，3C 產業在數位科技快速發展地帶動下更加的蓬勃發展，加上消費升 級趨勢以及大眾經濟實力的提升，3C 產品已逐漸成為生活必需品之一。各大 3C 賣場更是無所不用其極地滿足消費者需求，透過各式各樣的行銷方法、品牌設 計、購買體驗創新，更甚於在店面運營端也費盡苦心，使其在擴大營收同時，也 能盡己所能地降低成本。 在購買基數快速成長下，各大 3C 賣場更面對劇烈的購買需求變動，連帶使 店面運營人員配置帶來更多不確定性，需要相對應之模型協助公司掌握顧客之購 買行為，得以更準確地配置店面人力，進而達到管控營運成本的目的。更進一步 而言，透過顧客購買行為數據，也可掌控與決定每期的庫量，如此一來就能夠 降低過多產品的貨成本，也能避免發生缺貨狀況。 本研究使用某機構提供的 2006-2007 年 3C 賣場購買資料庫，以購買顧客為研 究對象，期望建立 3C 賣場顧客之購買行為預測模型。此模型以層級貝氏統計為 基礎，透過先驗以及後驗分配的設定，利用馬可夫鏈蒙地卡羅方法(Markov Chain Monte Carlo)模擬參數的聯合後驗分配。以吉氏抽樣(Gibbs Sampling)連續抽樣，進 而收斂至目標分配。再根據其後驗機率分佈，結合危險率函數(Hazard Ratio)建立 其購賣期間的預測模型，以期幫助 3C 賣場掌握其顧客之購買行為，藉以更準確 的進行每期 3C 賣場人潮之預測，進一步做服務人員配置規劃。

In recent years, the 3C industry has become more vigorous with the rapid growth of digital technology. In addition to the trend of the consumption-upgrade and the improvement of the economy, 3C products have gradually become one of the necessities. The major 3C stores are omnipotent to meet the needs of consumers. Through various marketing campaigns, brand design, and shopping experience innovation, 3C store brands are also working on the supply chain management to reduce their costs. At the same time, 3C stores also desire to reduce operating costs as possible. Under the rapid growth of the purchases, major 3C stores are facing more dramatic demand, which brings more uncertainty to the store operator configuration. This is the reason that the corresponding model is needed to help the company grasp the customer's purchase behavior more accurately, and therefore achieve the purpose of controlling the operating costs. Furthermore, through the purchase data of customers, it is also possible to determine the inventory of each period, which can reduce the inventory cost of excessive products and avoid the shortage of stocks. This study uses the 3C store purchase database (2006-2007) provided by an institution, and expects to establish a 3C store customer purchase behavior prediction model. Based on the hierarchical Bayesian statistics, this model uses the Markov Chain Monte Carlo method to simulate the joint post-allocation of the parameters through the prior and post-mortem allocations. Using the Gibbs Sampling to do the continuous sampling, and then converge to the target allocation. Based on the posterior probability distribution, combined with the Hazard Ratio to establish a forecasting model during the purchase period. With this model, this study is expected to help 3C stores to grasp the purchasing behavior of its customers, so as to forecast the crowds in each period more accurately, furthermore, to plan the service staffing more accurately.