http://vdisk.weibo.com/s/ljXQ7
twiiter big learning with graphs
There are different approaches to music recommendation,including:
(1)Expert-based approaches that are based on human annotation of music data such as the Music Genome Project,which is used by the Internet music service Pandora,com
(2)Social-based approaches that characterize items based on textual attributes,cultural information,social tags and other kinds of web-based annotations
(3)Content-based approaches that analyze the audio content for characterizing tracks and establishing item-to-item similarity
(4)Collaborative Filtering(CF) methods that analyze listening patterns(or other sorts of feedback) by many users,in order to establish similarities across users and items
Each of the above approaches comes with distinct disadvantages.An expert-based approach is etremely time consuming and hard to scale.Social tagging is unlikely to provide a complete characterization of music.Content based approaches,even at their best,miss important relations and distinctions humans make between songs;the reader is referred to the interesting piece by Slaney on the topic. The ability to directly model user desires makes CF among the more accurate and serendipitous approaches to recommendation. In addition CF naturally learns popularity trends by observing actual user behavior,which the orher approaches overlook. However,its availability is severely hindered by the need to have a substantial amount of relevant user behavior on record.The approach described in this paper aims at alleviating this central issue of CF.
Latent factor models constitute one of the leading CF techniques.They characterize both items and users as vectors in a space automatically inferred from observed data patterns. The latent space representation strives to capture the semantics of items and users,which drive their observed interactions.In music,factors might measure obvious dimensions such as genres,tempo or pitch or less well defined dimensions such as the emotions provoked by the song or the social target audience.One of the most successful realizations of latent factor models is based on matrix factorization(MF).These methods have become very popular in recent years by combining good scalability with predictive accuracy.In addition,MF provides a substantial expressive power that allows modeling specific data characteristics such as temporal effects,item taxonomy and attributes,social relations,and 3-way interactions.
本文摘自Yehuda大神的最新paper《Build your own music recommender by modeling Internet radio streams》
