探討偏好啟發對微時刻推薦的影響：互動式微時刻推薦系統

The influence of preference elicitation to micro-moment recommendations: An interactive MMRS

王詩堯(Wang Shih-Yao)

指導教授 ： 林怡伶

國立政治大學/商學院/資訊管理學系/碩士(2020年)

https://doi.org/10.6814/NCCU202001650

摘要

先前的研究指出推薦系統不僅應根據用戶的行為數據或受歡迎的項目進行推薦，還應符合用戶的偏好。部分推薦系統設計在使用者首次加入時調查其長期偏好。然而當在微時刻情境下，必須在限時內做出決策的壓力會導致注意力的限縮，最終會因此做出跟平時不同的選擇。這使我們相信作為決策輔助的推薦系統也應考慮短期意圖，並透過與使用者的互動來捕捉。這項研究進行了為期三週的使用者研究，以根據熱門程度、長期偏好和短期意圖來比較推薦的效果。本實驗設計了三個階段，包括進入前調查、使用聊天機器人、實驗後調查和訪談。總共招募了120名大學生，並將他們平均分配到四組之中。實驗的主要任務為透過與聊天機器人進行互動，在微時刻的各種情境下選擇一間餐廳。實驗結果顯示，MIX組（同時考慮長期偏好和短期意圖）會話的成功率比LTP組（僅捕獲長期偏好）高21.8%，並且利用更少的動作完成一輪推薦流程。另外，MIX組的所選項目在推薦列表上的平均排名最低，且推薦的點擊率最高。結果證明，這是四組中能支持使用者以較少的努力做出有效決策的最佳設計，而且該設計也是最適合支持微時刻的情境。透過證明MIX組優於LTP組，證明了在微時刻捕捉短期意圖的重要性。

關鍵字

微時刻 ； 推薦系統 ； 意圖 ； 偏好啟發 ； 互動式設計 ； 聊天機器人

並列摘要

Previous studies pointed out that recommendation systems should not only recommend by user's behavioral data or popular items but should conform to user preferences. Some recommendation systems investigate users’ long-term preferences when they first join. However, in micro-moments, giving limited available time to make decisions leads to a narrowing of attentional focus, eventually comes up with different choices. It convinces us that short-term intentions should also be taken into consideration and obtained through interactions with users. This research conducts a three-week user study to compare the effects of recommendations based on popularity, long-term preferences, and short-term intentions. Three phases including onboarding survey, chatbot use, post-experiment survey and interview were designed in this experiment. A total of 120 university students were recruited and assigned to one out of four groups. The main tasks focused on interacting with the chatbot then making choices of restaurants under various situations of micro-moments. The result shows that the sessions of the MIX group (considering both long-term preferences and short-term intentions) have a more 21.8% success ratio than the LTP group ones (capturing only the long-term preferences) and spent fewer actions in the recommendation processes. In addition, the mean of the MIX group' s selected position is the lowest, and the click-through of the MIX group is the highest. The results proved that it is the best design among four groups supporting users to make effective decisions with fewer efforts, moreover, this design is most suitable for the situation under micro-moments. Comparing the design of the LTP group, it also shows the importance of capturing short-term intentions at micro-moments.

並列關鍵字

Micro-moments ； Recommendation system ； Intention ； Preference elicitation ； Interactive design ； Chatbot

參考文獻

Amatriain, X., Pujol, J. M., & Oliver, N. (2009). I like it... I like it not: Evaluating user ratings noise in recommender systems. In International Conference on User Modeling, Adaptation, and Personalization (pp. 247–258). Springer.

Ariely, D. (2016). Time pressure: Behavioral science considerations for mobile marketing.

Baltrunas, L., Kaminskas, M., Ludwig, B., Moling, O., Ricci, F., Aydin, A., & Schwaiger, R. (2011). Incarmusic: Context-aware music recommendations in a car. In International Conference on Electronic Commerce and Web Technologies, 89–100.

Chen, J. W. (2016). A study on the selection of fast food restaurant by Utar Kampar students using analytic hierarchy process (AHP). Doctoral Dissertation, UTAR.

Chen, L., & Pu, P. (2007a). Hybrid critiquing-based recommender systems, 22–31.