以偏態加權處理音樂推薦中的冷啟動問題

Abstract

隨著線上音樂服務的巨大成長，人們現在有管道可以接觸大量風格迴異的音樂。大部分的歌曲都是大眾不知道的，他們代表著長尾分布長長的尾巴。因此，推薦系統是人們挖掘新音樂很重要的工具，但是已提出的方法主要都專注在提升整體系統的準確率，這常常使得系統推薦熱門歌因為熱門歌的歷史聆聽紀錄比較多。對於樂迷來說這樣的推薦不實用，因為他們很可能已經聽過這些歌，或者因為品味不同而對熱門歌不感興趣。另一件有趣的事是冷門歌比較不容易吸引到聆聽者，但當吸引到聆聽者時他們會認為這些冷門歌品質比主流歌好。所以，推薦這些長尾中隱藏的鑽石可以促進使用者對系統的信心。 在推薦系統中對於新的使用者或商品沒有歷史紀錄而難以推薦的問題稱為冷啟動問題。通常要處理冷啟動需要有額外的內容資料。這篇論文中我們使用卷積遞歸神經網路提取音樂特徵，設計了一個基於內容的推薦模型，而後提出一種針對像音樂歷史資料這樣偏斜的資料集重新加權的方法應用在這個模型上。主要的想法是在訓練模型時提高冷門歌的的重要度，也就是給予他們較大的懲罰。我們實驗了所提出的基於內容推薦模型的架構以及重新加權方法的的設計，另外，我們也應用這個偏斜加權方法在其他幾個模型上，展示了它的效果不只在基於內容推薦模型，協同過濾模型混和模型也適用。實驗結果顯示，推薦系統採用我們所提出的重新加權方法後，對於較不有名的歌的推薦表現有了顯著的提升。

With the enormous development of online music services, people nowadays have access to a massive amount of music covering a wide range of genres. Most of the music in the market are unknown to the public, representing the heavy tail in the long tail distribution. Therefore, recommender systems are an important tool for people to find new music that draws interest. However, the majority of methods proposed have focused on improving overall recommendation accuracy, which often leads to recommending hit songs since hit songs have more interaction data with users. This is not practical for music nerds because they are very likely to have heard of the hit songs, or even not interested in hit songs for they have different tastes. Another interesting thing is that as these niche songs are less likely to attract listeners, the ones they attract perceive the songs as higher quality than the mainstream songs. Thus, providing recommendations of the hidden gems in the long tail helps foster customer confidence. It is called the cold start problem in recommendation system when only a little or no historical data available for new items/users. Typically additional content is needed to solve the cold start problem. In this work, we design a content-based model using a convolutional recurrent neural network (CRNN) to extract features from the audio file and then apply to it a reweighting method targeting skewed data that we proposed. The main idea is to raise the importance of lesser-known songs when training the model, in other words, to give larger penalties on them. We experiment with the architecture for our content-based model as well as the design of the reweighting function. Besides, we apply the reweighting method on other several models, demonstrating the effectiveness not only on content-based models but also collaborative filtering-based and hybrid ones. Experiments show that models with our proposed skew reweighting method significantly outperform those without reweighting on lesser-known songs.