Continuous pLSI and Smoothing Techniques for Hybrid Music Recommendation
Continuous pLSI and Smoothing Techniques for Hybrid Music Recommendation
复制标题
DOI：
--
复制
发表时间：
2009
期刊：
影响因子：
--
通讯作者：
Kazuyoshi Yoshii;Masataka Goto
中科院分区：
文献类型：
--
作者：
Kazuyoshi Yoshii;Masataka Goto
本文提出了一个扩展的概率潜在语义索引（pLSI）的混合音乐推荐，处理由用户提供的评级数据和基于内容的数据提取的音频信号。原始的pLSI可以通过将用户和项目视为遵循多项分布的离散随机变量来应用于协同过滤。在混合推荐中，有必要处理通常表示为连续向量值的音乐内容。为此，我们提出了一个连续的pLSI，采用高斯混合模型。然而，这种扩展会导致严重的局部最优问题，因为它大大增加了参数的数量。这被认为是产生“枢纽”的主要因素，“枢纽”是不适当地推荐给几乎所有用户的项目。为了解决这个问题，我们测试了三种平滑技术：多项式平滑，高斯参数化和基于艺术家的编辑聚类。实验结果表明，虽然第一种方法没有改善，但其他方法显着提高了推荐准确率，降低了hubness。这表明在实际应用中适当地限制模型的复杂度是很重要的。
This paperpresentsan extendedprobabilisticlatent semantic indexing (pLSI) for hybrid music recommendation that deals with rating data provided by users and with contentbased data extracted from audio signals. The original pLSI can be applied to collaborative filtering by treating users and items as discrete random variables that follow multinomial distributions. In hybrid recommendation, it is necessary to deal with musical contents that are usually represented as continuous vectorial values. To do this, we propose a continuous pLSI that incorporates Gaussian mixture models. This extension, however, causes a severe local optima problem because it increases the number of parameters drastically. This is considered to be a major factor generating “hubs,” which are items that are inappropriately recommended to almost all users. To solve this problem, we tested three smoothing techniques: multinomial smoothing,Gaussian parametertying,andartist-baseditem clustering. The experimentalresults revealed that although the first method improved nothing, the others significantly improved the recommendation accuracy and reduced the hubness. This indicates that it is importantto appropriately limit the model complexity to use the pLSI in practical.