最佳匹配法與序列分群：電商用戶行為與運輸物流的分析

Optimal Matching & Sequence Clustering: Analyses of E-Commerce User Behaviors & Delivery Logistics

楊鈞宜(Yang Jiun-Yi)

指導教授 ： 莊皓鈞 ； 周彥君

國立政治大學/商學院/資訊管理學系/碩士(2021年)

https://doi.org/10.6814/NCCU202100754

摘要

研究動機與目的：精準行銷成效取決於消費者標籤的精準程度，而消費者分群效果受到序列資料特徵、所使用的距離計算方法等因素影響；本研究以實務案例驗證，導入不同的序列距離計算方式，是否有助於萃取消費者行為資訊差異並優化其分群效果。

研究方法：透過狀態序列轉換後，導入最佳匹配法計算序列相異度矩陣，最後分群觀察序列狀態分布比例圖表，並計算兩群間特徵指標統計顯著性，以驗證分群所得之群體特徵具有顯著差異。

研究應用場景：只要有時間戳記的日誌格式資料，透過狀態定義形成序列後，皆可以導入最佳匹配法運算出特徵，並用於分群、分類等不同算法中。

研究貢獻：本研究以「電商用戶行為序列分群」及「物流進程序列分群」兩案例，實證最佳匹配法所運算產生之特徵，相較於事件次數累計，更有助於提高分群多樣性，且能使分群群體特徵指標間產生顯著差異。

關鍵字

分群 ； 用戶行為 ； 序列 ； 序列分群 ； 最佳匹配法 ； 物流進程 ； 編輯距離 ； 相異度矩陣

並列摘要

Motivation and purpose: Ecommerce advertising retargeting performance highly depends on the quality of audience cluster tagging, and the cluster quality is affected by different sequence features transform methods. Our research aimed to validate that whether using optimal matching to generate sequence dissimilarity increased the diversity and significance of clustering results. Furthermore, we also discussed sequence clustering use cases in delivery logistics satisfaction to compare different scenario of sequence clustering.

Research method: First, we convert log data to state sequences, then compute sequence dissimilarity matrix using optimal matching algorithm. Last, run clustering algorithm and observe the state distribution plot among different clusters, with the results of Kruskal-Wallis significance test to validate that significant difference exists between key metrics of those two clusters.

Implement scenario: As long as there’s log format data with timestamps, we can transform it into state sequences through state definition, then generate dissimilarity matrix as a feature used in clustering, classification and other algorithms for increasing performance.

Research value: Our research validated that generating sequence dissimilarity by optimal matching algorithm not only increased the diversity of clustering results with more user pattern observed, but also segmented significantly different types of clusters using only state sequences data. Besides, we perform two analyses to show the entire process from data transformation, modeling and visualization.

並列關鍵字

optimal matching algorithm ； sequence data ； clustering ； user behavior ； delivery logistics ； edit distance ； dissimilarity matrix

參考文獻

[1] 普華永道會計師事務所。(民 108 年)。2019 全球消費者洞察報告。民 109 年 6 月 28 日，取自:https://www.pwc.tw/zh/publications/global-insights/2019-consumer- insights.html

[2] Google/Ipsos. (2017). Shopping Tracker. Retrieved June 28, 2020 from: https://www.thinkwithgoogle.com/data/us-shopping-behavior-statistics/

[3] Chen, Y., Fan, C., Li, Z., & Ren, L. (2020). Research on the Relationship between Precision Marketing and Company Development Ability. 2nd International Conference on Big-data Service and Intelligent Computation (BDSIC 2020). Association for Computing Machinery, New York, NY, USA, 42–48. DOI: https://doi.org/10.1145/3440054.3440062

[4] 劉洪偉、梁周揚、左妹華、陸丹、范夢婷、何銳超(民 108 年)。利用消費者瀏 覽行為識別品牌競爭關係研究。廣東工業大學學報，36(05)，1-6，13。DOI:

https://doi.org/10.12052/gdutxb.190063