基於電子發票交易數據之產品關係建模

Abstract

產品之間的關係形塑整體的市場結構，而顧客的消費行為正是這些關係建立的基礎。對零售商而言，掌握市場結構及產品間的互補性與替代性，有助於商品陳列規劃、促銷策略制定，以及競爭位置的判斷。過去許多研究多仰賴線上電商平台所提供的豐富顧客行為資料，例如瀏覽紀錄，以推論產品關係。然而，對於實體通路業者而言，這類資料通常難以取得，因此如何在僅有的交易資料中挖掘出有效的產品關係，成為一項重要的挑戰。 本研究參考自然語言處理領域的技術，將一筆交易視為段落，由其中的商品名稱短句組成，藉由微調預訓練語言模型（Pre-trained Language Model, PLM）BERT，從商品的共現關係和語意學習產品表徵。實驗設計包含三項任務評估模型能力：交易補全任務、互補品與替代品辨識、產品關係分群，前兩項任務中結果顯示模型能掌握商品間的共現關係，且透過先前文獻中提出的互補性與可交換性指標衡量產品間的關係，也顯示模型對互補品、替代品具有一定程度的辨識能力。最後，本研究提出一套基於互補性與可交換性所定義的產品相似度，基於相似度對商品進行分群與可視化，觀察模型對商品學習到的分佈情形，此方法比直接使用潛在空間表示分群效果更佳，亦能更直觀地反映產品間的替代與競爭關係，提供商家識別市場競品、了解市場結構的有效方式。

The relationships among products shape the overall market structure, with customer purchasing behavior serving as the foundation for establishing such relationships. For retailers, understanding the market structure and identifying complementary and substitutable products is essential for product display planning, promotional strategies, and assessing competitive positioning. While several previous studies infer product relationships using rich customer behavior data such as browsing histories available on online platforms, such data is often inaccessible to brick-and-mortar retailers. As a result, uncovering meaningful product relationships using only transaction data poses a significant challenge. This study draws on techniques from natural language processing by treating each transaction as a paragraph composed of short product name phrases. We fine-tune a pre-trained language model (PLM), specifically BERT, to learn product representations based on co-occurrence patterns and semantic information. The model is evaluated through three tasks: transaction completion, complements and substitutes identification, and product clustering. Results from the first two tasks show that the model effectively captures co-occurrence relationships and shows a certain ability to identify complements and substitutes using the complementarity and exchangeability metrics proposed in prior literature. Finally, we propose a similarity metric derived from these two metrics to perform product clustering and visualization. This similarity-based approach yields better clustering performance than directly applying latent space representations and offers a more intuitive and interpretable view of product relationships. It enables retailers to better understand competitive dynamics and identify key market competitors.