利用多商店的電子發票交易資料探討產品關係的表徵學習及其應用

Abstract

產品關係是零售商在制定決策時的重要依據，透過分析產品關係，零售商能夠識別競爭產品、優化產品陳列、並制定綑綁銷售和促銷策略，從而提升效益與收入。然而，傳統零售商通常僅擁有有限的交易資料，獲取額外外部數據往往需要高昂的成本。在這種缺乏豐富產品資訊的情境下，準確分析產品關係成為一項挑戰。此外，由於產品更新迅速，新品不斷湧現，進一步增加了分析的難度。隨著多通路行銷與全通路行銷策略的興起，多商店的情境逐漸增加，分析中必須考量不同商店間的差異。 受到自然語言處理領域語言模型成功應用於電子商務的啟發，本研究採用Word2Vec框架，結合商店資訊，從多商店的電子發票資料集中捕捉產品共現關係(Co-occurrence)，進而學習低維度的產品與商店表徵。我們還設計了兩種利用產品名稱進行表徵編碼的模型，包括從頭訓練(Training from Scratch)的模型與基於BERT編碼的模型，以解決新產品的問題。 我們通過三項真實世界任務來評估所學表徵的性能：(1)識別替代品與互補品；(2)將個體層級的產品關係遷移至聚合層級的缺失銷售資料預測；(3)多商店的產品配對推薦。實驗結果顯示，僅利用產品ID產生表徵的簡單模型在三項任務中均表現優於其他模型。此外，我們的模型展現了將個體層級的產品關係遷移至聚合層級的缺失銷售預測的潛力。進一步分析顯示，使用產品名稱進行編碼的模型表現不佳的原因在於其表徵呈現各向異性分布(Anisotropic Distribution)，降低了表徵的表達性，且模型未能充分捕捉產品間的關係。未來，我們計畫針對這些問題改進模型架構，並將預訓練的表徵與時間序列方法相結合，以提升銷售預測的準確性及適應真實世界情境的能力。

Product relationships are crucial for retailers when making marketing decisions, as analyzing these relationships can help identify competitors, optimize product display, and design bundling and promotional strategies, enhancing profitability and revenue. However, traditional retailers often rely on limited transaction data, and obtaining external data can be costly, making analyzing product relationships become a challenge. Furthermore, the rapid turnover of products and the frequent introduction of new products complicate the analysis even more. With the rise of multichannel and omnichannel marketing strategies, the multi-store scenario is expected to grow. Store heterogeneity may not be ignored in the analysis. Inspired by the success of language models in natural language processing applied to e-commerce, this study leverages the Word2Vec framework to integrate store information and capture product co-occurrence relationships from a multi-store e-invoice dataset, enabling the learning of low-dimensional representations of products and stores. We also developed two models that encode product names into product representations: a model trained from scratch and a BERT-based encoding model, designed to address the challenges of unseen products. We evaluate the learned representations across three real-world tasks: (1) identifying substitutes and complements, (2) transferring from individual-level relationship to aggregate-level missing sales prediction, and (3) multi-store matching recommendation. The experimental results show that the simple model, which generates product representations solely from product IDs, outperforms other models across all three tasks. Additionally, our model demonstrates potential in transferring individual-level product relationships to aggregate-level missing sales prediction. Further analysis reveals that the models using product name encoding underperform due to anisotropic distribution of representations, which reduce their expressiveness and hindered the capture of product relationships. In the future, we aim to address these issues by refining the approach and integrating our pre-trained representations with time-series methods to enhance sales prediction accuracy and better adapt to real-world scenarios.