Abstract
摘要:隨著音樂串流服務的蓬勃發展,使用者能夠在隨意情境下享受聆聽音樂的樂趣。然而音樂資料庫的大量滋長,也往往讓使用者無所適從,不知如何找到合適的音樂或該聽何種音樂。音樂檢索及推薦系統的目的即在於幫助使用者能夠快速找到合適的音樂,對不知道該聽什麼樣歌曲的使用者,也能分析其喜好,推薦合適的音樂,以提昇聆聽體驗與樂趣。
現有的推薦系統根據使用者過去的聆聽資料,以推薦特性相似的音樂為主。然而,人們的音樂偏好會跟著情境而異,在運動時、參加派對時、就寢前等不同的情境下,會有不一樣的需求。因此,我們將探討使用者的需求、聆聽情境、以及音樂歌單間的相關性,藉以提昇音樂推薦系統的效能。此外,我們也將透過使用者與推薦系統間的互動資料,增進對使用需求的了解,藉以提供貼心的音樂饗宴。同時,我們也將透過巨量音樂資料分析及機器學習,幫助使用者嘗試外在世界的新穎音樂,使其聆聽體驗更加豐彩。
本實驗室長期投入音樂相關的研究,多年來已經累積豐富的經驗與紮實的基礎。近年來,我們更與台灣首屈一指的音樂串流服務公司KKBOX建立良好的合作關係,進行巨量聆聽及歌曲資料分析,獲得寶貴經驗,有利此研究計畫的推行。我們將利用深度學習以及類神經網路技術發展互動與情境式的音樂推薦系統。
本計畫將分三年執行,內容包括(1)使用者聆聽資料分析,包括使用者需求探勘以及聆聽情境偵測;(2)音樂聆聽情境分析,包含歌單主題探勘與辨識、歌詞關鍵字分析以及自動歌單產生;(3)整合上述研究成果,建立互動與情境式音樂推薦系統。
Abstract: The advancement of online music streaming technology and serves leads to a new music listening era in which users can access music anytime and anywhere through any smart device. As the world is now populated with abundant music collections in the electronic form, music selection and search is no longer an easy task. The purpose of a music retrieval and recommendation system is to remove such hurdle so that even if a user does not know what music to listen to, the music recommended by the system is likely to be liked by the user.
The importance of elevating user experience to an unprecedented level in the new era cannot be overstated. However, most existing music recommendation systems recommend music based on the past preference record of a user; only the hit rate of each song is considered. The truth is that the music people listen to varies with the context in which the music is listened to. For example, songs played during a sports event can be very different from those played during a party or before sleep. In other words, the music need of a user is related to the context of music listening. Because most people listen to a set of music pieces in each period of time, the notion of playlist plays an important role in user preference analysis as well. We are interested in investigating the relation between user need, listening context, and music playlist as well as various contextual topics. The findings will help us to develop a context-aware music recommendation system. In this work, we will also consider the fact that most smart devices have a display panel that serves as a simple and easy platform for user interaction. Therefore, we will investigate interactive recommendation in which the need or preference of a user is learned from the interaction between the user and the system. The world is full of wonders; it should be no different for music listening. Therefore, our system will be built in such a way that it will help the users to discover novel music.
Our lab has engaged in music data mining and music emotion recognition for many years. The recent collaboration with KKBOX using real-world big data further enriches our experiences in this field. We are strongly motivated to apply the advancement in deep learning and neural network to music recommendation. This three-year project has three major parts: (a) analysis of user listening history, including user need mining and listening context detection; (b) investigation of contextual topics, including playlist topic mining and recognition, lyrics topic analysis, and automatic playlist generation; (c) system development, mainly building an interactive, context-aware music recommendation system.
Keyword(s)
音樂檢索
音樂推薦
深度學習
類神經網路
情境式推薦
互動式推薦。
Music recommendation
deep learning
neural network
information retrieval.