在數位音樂蓬勃發展的現在，協助使用者找到感興趣內容的推薦系統已日益重要。以往推薦系統大多採用協同過濾(Collaborative Filtering, CF)及物件內容相關特徵資料作為分析方法及標的。然而協同過濾方法容易受到冷啟動或者熱門度之影響，且物件內容相關特徵資料也難以深入描述物件吸引使用者之特性。在本研究中，通過將協同過濾之矩陣分解(Matrix factorization, MF)方法結合卷積神經網絡(Convolutional Neural Network, CNN)分析音樂之頻譜圖，開發了一種新型的推薦系統Spec-Rec。使用KKBOX WSDM Cup 2018: Music Recommendation and Churn Prediction 資料集進行實作驗證。最後實驗結果顯示Spec-Rec系統與競賽中之頂尖模型有相當之預測水準。且使用訓練完之模型用來預測其他不在訓練範圍內之音樂也能保持穩定之預測水準。證明本系統藉由直接分析使用者對於音樂的偏好特徵，在評價資料稀缺的狀況下也能將音樂推薦給適合的使用者。;Nowadays, digital music booms rapidly. The recommendation system which helps users to find the music they like becomes more and more important. In the past, most recommendation systems analyze by using Collaborative Filtering (CF) or related metadata. However, CF are usually affected by Data Cold Start Problem or Popularity Bias of the item. Also, the metadata cannot describe the feature of the item which attracts users’ attention. In this article, by raising a new type of recommendation system called Spec-Rec, a combination of Matrix Factorization (MF) with Collaborative Filtering and Convolutional Neural Network (CNN), to analyze the spectrogram of music. The experiment uses KKBOX WSDM Cup 2018: Music Recommendation and Churn Prediction dataset for validation. The experimental results show that the Spec-Rec is competitive with the model of the top class in the KKBOX WSDM challenge. Also, the trained model can be used to predict the music outside the training data and can provide a stable performance. According to the results, Spec-Rec system can recommend music to the adequate users by analyzing spectrogram of music and user latent vector even when there is a lack of rating data of music and user preference.
