獨立音樂聽眾偏好結構對於推薦策略之效能影響：以 StreetVoice 為例

Abstract

不同於過往的研究關注在推薦系統演算法的效能表現，本研究更以此為基礎納入音樂聽眾的個體差異因素，試圖確認不同聽眾偏好的結構樣態，是否會有對應合適的推薦策略。另外音樂聽眾研究上，大多將獨立音樂視為主流音樂的分支，又或是聚焦在論述獨立音樂的本質，而鮮少研究著墨於獨立音樂聽眾研究。於此本研究以心理量表與使用者興趣檔量化得知獨立音樂聽眾在偏好結構的特徵表現，並驗證構念與操作化的合適性如何，同時在推薦策略方面，欲釐清不同策略在推薦效能上的表現優劣以及特性，且其中推薦效能之間是否彼此皆為正向關聯，進而揭曉聽眾偏好結構是否對於推薦策略的表現有調節影響。本研究採重複測量設計的實驗法，以StreetVoice獨立音樂平台的使用者為研究對象，共徵集126位研究參與者，測量每位參與者在四種偏好結構的表現－偏好多樣性、偏好開放性、音樂自我認同與音樂涉入程度，並同時給予三種推薦策略－基於使用者協同過濾、基於內容過濾與熱門播放排行，邀請參與者評估各策略在五種推薦效能的表現－未知歌曲比率、使用者滿意度、意外驚喜性、多樣性與新穎性。經統計分析結果發現，聽眾在四種偏好結構的表現上彼此間具有正向關聯，然而在量表與興趣檔指標的量測上未有一致結果，而不同推薦策略確實存在不同的效能表現，但推薦效能之間不一定為正向的關係。最後本研究所關注的主軸，發現聽眾偏好結構的納入，實則會對推薦策略的效能表現有調節影響。

Different from previous studies focused on the effectiveness of recommendation algorithms, this research employs individual differences of music audience as moderator, attains to confirm different preference structures of the audience would correspond to a fitting recommendation strategy. Besides, in other audience research, most of them view independent music as a branch of mainstream music or focus on addressing the concept of independent music, few of them concentrated on independent music audiences. Thus, our research introduces psychological scales and user profiles to quantify the behavior of independent music audiences in the preference structure, plus validates the suitability of construct and operationalization. Meanwhile, it is necessary to clarify whether recommendation strategies have different performances and characteristics, also prove whether the recommendation effectiveness metrics are positively correlated to each other. Furthermore, reveal the hypothesis that assumes the preference structure of the audience has a moderation effect on the effectiveness of the recommendation strategy. This research adopts the within-subject design experiment method, takes users of independent music platform StreetVoice as a research subject, totally recruiting 126 research participants, and measuring the behavior of each participant in four preference structures, such as preference diversity, openness to novelty, music self-identity and music involvement. By given three recommendation strategies, including user-based collaborative filtering, content-based filtering, and top popular charts. Inviting participants to evaluate the performance of each strategy in five recommendation effectiveness, such as unknown ratio, user satisfaction, serendipity, diversity, and novelty. Eventually, findings emerged from the statistical analysis, discover that the audiences have a positive correlation with each other in the behavior of the four preference structures, however, there is no consistent result in the measurement of scale and metrics of user profiles. Additionally, different recommendation strategies are confirmed to have different performances, yet the relationship between these performances is not necessarily positive. Ultimately, it is worth noting that preference structures of music audiences evidenced has a moderation effect on the performance of recommendation strategies.