透過社群媒體評論及選擇性使用者偏好建立 Embedding-based 音樂推薦系統

Embedding-based Music Recommendation System by Leveraging the Social Media Review and the Selective User Preference

謝昀哲(YUN-ZHE XIE)

指導教授 ： 謝俊霖

國立臺灣大學/管理學院/資訊管理學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202102159

摘要

隨著線上音樂串流以及訂閱制成了人們聽音樂的主流方式之後，人們有了比以往更多的選擇，這也同時使音樂推薦系統比以往更為重要。音樂推薦系統不僅能幫助使用者在眾多選擇之中快速找到自己可能有興趣的音樂，也可以找出使用者的潛在興趣。
在我們的研究之中，我們建立了一個Embedding-based的J-pop音樂推薦系統，為了能更精確地捕捉使用者的偏好。我們以歌曲的其他資訊，例如歌手或作曲當作使用者可能會喜歡一首歌曲的原因，並且以專業領域知識考慮了這些資訊之間的關聯。除此之外，我們另外蒐集了社群媒體上的評論來代表每一首歌的客觀感受。我們發現人們經常在社群媒體上面描述一首歌曲對於他們的感受，而這些感受可以帶來與其他使用者偏好不同的資訊。在推薦的時候，我們分別考慮了使用者的長期偏好以及短期偏好。長期偏好代表了使用者喜歡的音樂風格，而短期偏好則是指出了使用者最近生活周遭發生的事會影響到近期使用者所聽的歌曲類型或主題。此外，我們也認為使用者會因為不同的原因和偏好而喜歡不同的歌曲。也就是說使用者偏好是會隨著被推薦的歌曲而有所不同的，我們將這個概念稱為選擇性偏好。
我們透過深度學習模型建立一個Embedding-based的推薦系統。該模型包含了知識圖譜及注意力機制。接著我們透過實驗切除法(ablation experiment)來評估我們實作的每個概念對模型成效的影響。實驗成功之後，我們進一步去分析模型的結果及發現來證實這些機制運作的合理性。

關鍵字

音樂推薦系統 ； 使用者偏好 ； 注意力機制 ； 社群媒體分析 ； 知識圖譜 ； 機器學習

並列摘要

With online music streaming and subscription becoming the mainstream way of listening to music, people have more songs to choose from. Therefore, the importance of the music recommendation system is more important than before. It can not only reduce the time for users to find songs that they are interested in but explore the potential interests of users.
In our research, we build a personalization embedding-based recommendation system for J-pop music, and we will focus on how to capture the user preference more precisely. We use the side information of the song to represent the reason that people may be interested in the song. We consider the relation between these user preferences by domain knowledge. Moreover, we collect reviews on social media as one of the user preferences. We found that people tend to write reviews on social media platforms to describe the feelings of the song, and it brings information that is different from the traditional user preference. When recommending, we consider the long-term and short-term user preferences. The long-term user preference indicates the user's music habit, and the event or the purpose of users listening to music will impact the short-term user preference. Besides, we assume that people listen to different types of music for different reasons and preferences. It means that the user preference will vary with the song, and we called it selective user preference.
We design an embedding-based recommendation system by deep learning model. The model includes a knowledge graph and the attention mechanism. Then, we hold an ablation experiment to evaluate the performance of the concept we implement. After successful on the ablation experiment, we discuss the result and provide some findings to prove the concept is reasonable.

並列關鍵字

Music Recommendation System ； User Preference ； Attention Mechanism ； Social Media Analysis ； Knowledge Graph ； Machine Learning

參考文獻

Ashby. (2000). A Stochastic Version of General Recognition Theory. Journal of mathematical psychology, 44 2, 310-329.

Bahdanau, D., Cho, K., Bengio, Y. (2015). Neural Machine Translation by Jointly Learning to Align and Translate. CoRR, abs/1409.0473.

Bordes, A., Usunier, N., García-Durán, A., Weston, J., Yakhnenko, O. (2013). Translating Embeddings for Modeling Multi-relational Data. NIPS,

Celma, Ò. (2010). Music Recommendation and Discovery - The Long Tail, Long Fail, and Long Play in the Digital Music Space.

Celma, Ò., Cano, P. (2008). From hits to niches?: or how popular artists can bias music recommendation and discovery. NETFLIX '08,