On the Approaches to Triggering Serendipity in Recommender Systems and Their Impacts to User Satisfaction

本研究主要針對兩種主要的推薦系統策略：協同過濾及內容導向，並在推薦過程中導入隨機性與降低準確性的方法，藉以觀察隨機性或準確性的降低對於刺激推薦商品中意外驚喜之發生，及對傳統用於評估推薦結果品質的各項指標如滿意度、購買意願等之影響。本研究採取實驗法以驗證假設，受測者隨機分配各特定專為實驗設計之推薦系統後，於一個虛擬電影租賃網站進行購買決策行為。待實驗結束，受測者以填寫問卷的方式回報其感興趣程度、滿意度與購買意願等指標。實驗結果證實意外驚喜的提升與其他各項指標間存在互換關係。除此之外，協同過濾型的推薦系統配合降低準確性的作法，是最適合刺激意外驚喜發生的推薦系統策略；這樣的組合能夠在不犧牲現有推薦品質的情況下提高意外驚喜出現的比例。最後，針對推薦的候選商品加上特定過濾條件如較高商品評價之門檻，將有助於減緩上述意外驚喜與其他衡量指標間之互換關係。本研究的結果對於推薦系統中意外驚喜的相關研究有重要意涵。This study focuses on two main recommender paradigms: collaborative-filtering and content-based, and introduces the “Role of chance” approach and the “Anomalies and exceptions” approach. The above two approaches are integrated in this study to form a theoretical model that examines their effects on triggering serendipity and the subsequent effects on several metrics such as user satisfaction and willingness to pay. An experiment was conducted to test the model. Participants were grouped by each recommender conditions and were asked to make a purchase at a simulated online retailer. After the experiment, participants were asked to complete a survey to report their interest, satisfactory and willingness to pay levels. Results indicate that there might be a trade-off relationship between serendipity and other metrics. In addition, collaborative-filtering recommenders which adopted the “Anomalies and exceptions” approach seem to be the most suitable combination to introduce serendipity. Finally, setting a threshold to filter products among recommendation candidates such as high rating would ease the trade-off. Our findings have major implications for the ongoing research on serendipity of recommendations