Style4Rec：利用風格和購物車信息增強基於 Transformer 的電子商務推薦系統

Abstract

了解用戶的產品偏好對於產品推薦系統的有效性至關重要。精準營銷使用用戶的歷史數據來識別這些偏好,並提供與其一致的建議。然而,在用戶偏好的動態格局中,最近的瀏覽和購買記錄可能更好地反映當前的購買偏好。雖然基於Transformer 的推薦系統在順序推薦任務中取得了重大進展,但它們在使用產品圖像風格信息和購物車數據方面往往存在不足。產品圖片無疑會影響用戶的產品偏好,而款式信息在電商平台上的最終購買決策中發揮著關鍵作用。為了應對這一挑戰,我們提出了 Style4Rec,一種基於 Transformer 的電子商務推薦系統,它利用風格和購物車信息來增強現有基於 Transformer 的順序產品推薦系統的性能。我們的方法涉及創建一個基於 Transformer 的順序產品推薦系統,該系統集成了神經風格提取模塊。該模塊從產品圖像中提取風格信息並將其轉換為多層轉換器推薦器的樣式嵌入。我們根據 session 整理出數據,生成單獨的歷史行為和產品向量。通過比較這些向量,我們得出最終的產品推薦。此外,我們將購物車數據作為補充數據集,以進一步提高性能。實施 Style4Rec 帶來了顯著的改進:HR@5從 0.681 增加到 0.735,NDCG@5 從 0.594 增加到 0.674,MRR@5 從 0.559 增加到0.654。我們使用合作公司所提供的電子商務數據集對我們的模型進行了廣泛的評估,並發現 Style4Rec 在各種評估指標上都優於最先進的基於 Transformer 的順序推薦基準。因此,Style4Rec 具體推進了個人化電子商務產品推薦系統領域的實質性進步。

Understanding users’ product preferences is crucial for the effectiveness of product recommendation systems. Precision marketing leverages users’ historical data to identify these preferences and provides recommendations aligned with them. However, in the dynamic landscape of user preferences, recent browsing and purchase records may better reflect current purchasing preferences. While transformer-based recommendation systems have made significant progress in sequential recommendation tasks, they often fall short in utilizing product image style information and shopping cart data. Product images undoubtedly influence users’ product preferences, with style information playing a pivotal role in their final purchasing decisions on e-commerce platforms. In response to this challenge, we propose Style4Rec, a transformer-based e-commerce recommendation system that leverages style and shopping cart information to enhance the performance of existing transformer-based sequential product recommendation systems. Our approach involves creating a transformer-based sequential product recommendation system that integrates a neural-style extraction module. This module extracts style information from product images and converts it into style embeddings for the multi-layer transformer recommender. We organize the data based on sessions, generating separate historical behavior and product vectors. By comparing these vectors, we derive the final product recommendations. Additionally, we incorporate shopping cart data as a supplementary dataset to further enhance performance. Implementing Style4Rec has resulted in significant improvements: HR@5 increased from 0.681 to 0.735, NDCG@5 increased from 0.594 to 0.674, and MRR@5 increased from 0.559 to 0.654. We conducted an extensive evaluation of our model by using an e-commerce dataset provided by our partnering company and found that Style4Rec outperformed state-of-the-art transformer-based sequential recommendation benchmarks across various evaluation metrics. Therefore, Style4Rec represents a substantial advancement in personalized e-commerce product recommendation systems.