Balancing accuracy and fairness in recommender systems: from offline to online algorithms

Ph.D.Recommender systems are critical for powering the fast-growing web and mobile segments of the economy, to connect users to the right items (videos, jobs, news articles etc.). Accuracy and fairness are two essential goals in recommender systems. Most systems focus on optimizing recommendation accuracy, yet may bring about novel challenges in ensuring fairness. The objective of this thesis is to study how to develop effective recommendation algorithms to produce accurate and fair recommendations.First, we propose a recommendation algorithm for improved accuracy and generalization ability. The algorithm combines the field information in the input item features and the high-order feature interactions. Rather than using the single point maximum a posteriori estimation, we estimate the probability distribution of the field-aware embedding parameters, where various confidence levels for feature embeddings are incorporated to enhance the robustness and recommendation quality.Next, we propose a re-ranking algorithm on top of existing recommendation algorithms for recommendation refinement. Top candidates are re-ranked to improve the accuracy of the recommendation. We develop a graph neural network based framework to explicitly model how items mutually influence each other in the lists, and thereby affect user satisfaction. We also encode personalized user preferences and intents into the item representations of the model.In practical applications, users’ feedback arrives sequentially and frequently. Unlike offline approaches that have to be re-trained completely for new records, we design a contextual bandit algorithm for the online recommendation, which takes user and item content features into consideration and automatically balances exploration and exploitation. We enhance the representability by adopting the generalized linear models, which include both linear and logistic regressions and have exhibited stronger performance in many real-world applications. We prove a sublinear upper bound on the regret of the algorithm.With increasing concerns about bias and discrimination in machine learning, recommender systems have to go beyond purely accuracy-oriented approaches and take fairness issues into account, especially in the recommendation area related to education, housing, public accommodation etc. Fair treatment for different item groups is required. However, the optimization of fairness often comes at the expense of recommendation accuracy due to the conflicting goals between recommendation accuracy and fairness. To resolve the problem, we propose a fairness-aware re-ranking algorithm to balance the recommendation quality and fairness.Nevertheless, in real recommendation applications, managing the trade-off between accuracy and fairness is not a single-shot process. Rather it is an online task, where the system could focus on improving accuracy for users with particular favor, and the lack of fairness at the time can later be compensated when recommending items to users with diversified interests. As such, we can achieve long-term system’s fairness while preserving satisfying recommendation quality. We present a reinforcement learning based model to realize it. Extensive experiments on both synthetic and real-world datasets consistently show the advantage of our proposed model over the state-of-the-art approaches.推薦系統對於推動互聯網和移動經濟的快速發展，幫助用戶發現感興趣的商品至關重要。準確性和公平性是推薦系統的兩個關鍵目標。本文的目的是研究如何設計更加有效的推薦算法以進行更加準確公平的推薦。首先，我們提出了一種用於提高準確性和泛化能力的推薦算法。該算法將商品輸入特徵中的域信息和高階特徵交互相結合。與傳統的單點最大後驗估計不同，我們提出對特徵嵌入參數的概率分佈進行估計，引入特徵嵌入的置信度以增強算法的魯棒性和推薦質量。其次，我們提出了一種可以應用在任一推薦算法上的重排序算法，以對現有的推薦結果進行優化。我們設計了一個基於圖神經網絡的算法對列表中的商品如何相互影響進行顯式建模。此外，我們還將用戶的個性化意圖編碼到了表示向量中。在實際應用場景中，用戶的新反饋會持續不斷的產生。傳統的離線學習方法需要對新樣本完全重新訓練。與此不同，我們設計了用於在線推薦場景的上下文老虎機算法。該算法將用戶和商品的內容信息考慮在內，並自動在探索和發現之間取得平衡。我們採用了廣義線性模型來增強算法的實用性 。我們證明了該算法有次線性的後悔收益保證。隨著人們對機器學習所可能帶來的社會問題如偏見歧視等的關注日益增加，推薦系統除了需要考慮準確性，也應更加註重推薦的公平性問題。推薦算法需要保證對不同受保護屬性的商品組別公平對待。然而，推薦準確性和公平性的優化目標相互衝突，提升推薦公平性通常需要以犧牲準確性為代價。為了解決該問題，我們提出了一種考慮公平性的重排序算法，以平衡推薦質量和公平性。實際上，對推薦準確性和公平性的平衡是一個長期過程。推薦系統可以在對有確定偏好的用戶推薦時專注於提高準確性，稍後在向具有廣泛興趣的用戶推薦時彌補之前缺乏的公平性。因此，我們可以在保證推薦質量的同時，實現系統長期的累積公平。我們設計了一個平衡推薦準確性和公平性的強化學習模型。合成和真實數據集上的實驗證明了新方法對於現有方法的優勢。Liu, Weiwen.Thesis Ph.D. Chinese University of Hong Kong 2020.Includes bibliographical references (leaves 155-183).Abstracts also in Chinese.Title from PDF title page (viewed on 16, September 2021)