針對情感商品的情境感知推薦系統-以流行音樂為例

情感商品，如音樂、電影、小說…等，與一般單純為了使用功能的功能商品有很大的不同。因為情感商品的評價與個人感受有關，情感商品在網路上通常會存在比較多見仁見智的評論；商品的效用也更與商品本身內容及通常能帶給使用者什麼感覺來的有關；而當下的使用情境適合那些情感商品更是情感商品與功能商品很不相同的地方。一般考慮到網路評論的推薦系統，通常只有單獨從網路評論裏分析與評論與商品屬性的關係，較少同時考慮商品其它實質內容像是音樂的歌詞、音律對商品評價的影響。而這類推薦系統主要在找正負傾向規則，也較少討論找出像是「聽了讓人感到很遺憾」這種引發人類情緒的情感商品效用規則。另外傳統的推薦系統通常也會忽略在不同的情境之下，人們的商品偏好會有很大差別的現象；像在運動時喜歡聽很動感音樂的人，並不一定在睡前的情境也會喜歡類似的音樂。在這樣的背景之下，如何建立一個能同時有效的從網路評論及音樂內容找出考慮到商品能有的心情效能並能感知情境的內容推薦系統是有其研發重要性的。本研究以流行音樂這個情感商品為例，提出一個雛型架構來達成以上的目標。首先本研究先建立能了解網路評論狀況的情感標籤分類器，用於隨時了解某商品目前網路評論的情感傾向；同時本研究也建立一個同時考慮到音樂歌詞及音質特性的音樂內容分類器，用於從音樂的內容特徵來得到某音樂商品可能音樂情感傾向。經過資料的收集、分析與訓練，整體的網路評論情緒傾向分類器經過測試平均的F1有70.09%的分類成功率；而音樂內容情感傾向分類器的巨觀平均分類成功率F1有74.89%、微觀平均分類成功率F1則是到達了80.13%。利用這二個分類器，加上本研究建立的運動及安眠情境感知雛型系統與偏好資料，本研究提出了符合情感商品特性的情境感知商品推薦系統。最後本研究設計了三個實驗來驗證情緒分類器及情境感知系統的有效性。從情緒音樂實驗中我們發現利用由本研究分類器所分類出的喜悅與平靜的音樂，可以有效的降低受測者的悲傷度並增加控制度；另外從實驗結果中可以得知音樂推薦的順序雖沒有明顯影響悲傷度及控制度的變化，但會影響收聽者的滿意度。從運動及安眠情境實驗的訪談資料中，我們可以得到一些對本研究的情境感知雛形系統的正面測試回饋結果。這對未來開發類似系統會有重要的幫助。

關鍵字

意見分析；音樂推薦系統；情境感知

並列摘要

Emotional products, such as music, movies, novels… etc., are quite different from functional products. Because the evaluation of emotional products is related to personal feelings, emotional products have much more kinds of user reviews on the Internet than functional products. People will more likely choose an emotional product because of the content and People will more likely use different emotional products then functional products in the different situations. In the past researches, an opionion mining based recommendation system usually only use web reviews to recommend products, there were not many studies use both user reviews and the content of the product at the same time. This kind of recommendation system was also only driven by positive or negative tendency rules, and there are also few discussions to find out the emotional rules, such like “this music is very happy to hear.” In addition, the traditional recommendation system usually ignores the fact that people's preference for emotional products will be very different in different contexts; people who like to listen to rock music during exercise are not necessarily like rock music at bedtime. Under such a background, how to establish a context-awareness recommendation system that can effectively and effectively help people to choose emotional products by using both online reviews and product content is very important. In this study, we took pop music as a case of emotional products and we had proposed a recommendation prototype system use both web reviews opinion mining and a lyrics and sound content-based tags classifier to be the recommendation sources. After data collection, analysis and training processes, the overall web reviews opinion classifier accuracy average F1 score is 70.09%; the music content emotional tags classifier accuracy marco-average F1 score is 74.89% and micro-average F1 score is 80.13%. Using these two classifiers, a context-aware prototype system was proposed in this study, we had completed a context-aware product recommendation system that meets the characteristics of emotional goods. Finally, we designed three experiments to verify the effectiveness of the emotional classifier and context-aware system. In the emotional music experiment, we found that by using joy and calm music content classifier trained in this study can effectively help to reduce the sadness and increase the degree of control. In addition, we can also found that the recomendtion order joy-clam and clam-joy was no significant different in emotion regulation, but it will affect satisfaction. From the interview data of the exercise and sleeping context experiments, we got some positive feedback results for the context-aware prototype system of this study. This will be some help in developing similar systems in the future.