以表徵學習方法實現考量多樣性偏好的推薦系統

Abstract

在推薦系統中，如何同時兼顧準確性與多樣性一直是長期存在的挑戰。雖然多樣化方法能提升內容的多元性，但常忽略使用者對多樣性的個別偏好，導致推薦結果不符合使用者的興趣，降低使用者滿意度。近期個人化的多樣化方法研究試圖透過建模使用者的多樣性偏好來解決此問題，但大多數方法仍依賴於後處理的優化策略，將使用者的多樣性偏好視為外部訊號，於分離的第二階段根據此資訊重新排序推薦結果。 本研究提出一種全新的多目標學習架構，將細緻的多樣性偏好內嵌至訓練過程中。透過設計多樣性偏好建模任務，以建構使用者對商品類別屬性的興趣分布，我們的方法能學得同時捕捉「相關性」與「多樣性偏好」的豐富表徵。 在真實世界資料集上的實驗結果顯示，我們的方法不僅能同時提升準確性與個人化多樣化表現，亦能有效緩解兩者間的權衡困境。這些成果突顯了將多樣性偏好視為可學習訊號的重要性，為打造更具適應性與使用者導向的推薦系統開啟了新的方向。

Balancing accuracy and diversity has been a long-standing challenge in recommender systems. While diversification methods enrich content variety, they often ignore user-specific preferences for diversity, leading to irrelevant recommendations and degraded user satisfaction. Recent efforts in personalized diversification attempt to address this by modeling users’ diversity preferences, but most rely on post-processing heuristics or external signals detached from core user modeling. In this work, we propose a novel multi-objective learning framework that internalizes fine-grained diversity preferences directly into the training process. By designing diversity preference modeling tasks that model user interest distributions across multiple item attributes, our method learns richer user/item representations that jointly capture both relevance and diversity preferences. Extensive experiments on a real-world dataset demonstrate that our approach not only improves accuracy and diversification personalization simultaneously, but also mitigates the typical trade-off between them. Our results highlight the value of treating diversity preference as a learnable signal, opening new possibilities for more adaptive, user-centric recommender systems.