2014年协同过滤技术综述：超越用户-物品矩阵的创新与发展

需积分: 10 37 浏览量更新于2024-07-18 收藏 1.2MB PDF 举报

2014年的这篇经典综述名为《CollaborativeFilteringbeyondtheUser-ItemMatrix:ASurveyoftheStateoftheArtandFutureChallenges》（超越用户-项目矩阵：推荐系统的最新进展与未来挑战），作者是来自代尔夫特理工大学的研究者YUESHI、MARTHALARSON和ALANHANJALIC。该论文深入探讨了过去二十年间在推荐系统领域取得的显著成就，尤其是在协同过滤技术（Collaborative Filtering）方面的研究。协同过滤作为推荐系统的核心框架，其基础在于用户-项目（U-I）矩阵，这个矩阵记录了每个用户对集合中项目的个人偏好，它对于生成个性化推荐至关重要。然而，随着技术的发展，新的推荐场景开始涌现，这些场景提供了超越传统U-I矩阵的有价值信息，主要分为两类：一类是关于用户和项目的丰富元数据，包括用户的个人信息、兴趣偏好、历史行为等；另一类是关于用户和项目互动的信息，如用户之间的社交网络关系、项目之间的关联性等。文章对这些新兴的推荐场景进行了详尽的总结和分析，重点介绍了如何利用这些额外信息来改进和扩展现有的协同过滤算法。作者讨论了诸如基于内容的推荐（Content-based Recommendations）、矩阵分解方法（如SVD和PMF）、混合推荐（Hybrid Recommendations）以及考虑到用户行为动态变化的实时推荐策略等新型技术。此外，还涉及到了一些未来挑战，如数据稀疏性问题、隐私保护、算法的可解释性和公平性等问题。这篇综述不仅概述了当时协同过滤领域的研究现状，而且对未来的研究方向和技术发展提出了富有洞察的见解，对于理解并应用现代推荐系统具有重要的参考价值。通过阅读这篇论文，读者能够对基于协同过滤的推荐算法有更深入的理解，并为开发适应新场景的创新算法奠定基础。

3:8 Y. S h i e t a l .

Fig. 2. A conceptual view of the algorithms incorporating the rich side information of users and items to

extend memory-based CF.

user’s comment on a hotel that explains why she gave a 5-star rating to the hotel may

indicate her preference to hotels in a particular situation.

4. STATE-OF-THE-ART ALGORITHMS

For each of the categories of information sources that have just been discussed, a

category of algorithms can be identiﬁed that have been developed to exploit it. The two

categories of algorithms are discussed in this section.

4.1. Collaborative Filtering Using Side Information on Users and Items

4.1.1. Extending Memory-Based Collaborative Filtering.

Recall that traditional memory-

based CF approaches measure the similarity between users or items based on the

U-I matrix. As mentioned earlier, side information can be exploited for calculating or

reﬁning the similarities and thereby improving recommendations. Conceptually, we

can represent the algorithms that extend memory-based CF for incorporating the rich

side information in a diagram, as shown in Figure 2. For instance, in a movie recom-

mendation domain, a user can be represented by the tags that she assigned to movies

in the past, whereas a movie can be represented by its genres.

To the best of our knowledge, the ﬁrst work in this direction was carried out by

Melville et al. [2002], who proposed ﬁrst predicting the missing values in the U-I matrix

by using the side information of items (e.g., categories like title and genre, referred

to in this work as “content features”) and then deploying user-based CF to generate

recommendations. Similar approaches have been studied for different use cases [Burke

2002; Tso-Sutter and Schmidt-Thieme 2006]. Another category of approaches focused

on tags as side information [Firan et al. 2007; Shepitsen et al. 2008]. Similar to ratings,

tag-derived information also encodes connections between users and items. Tso-Sutter

et al. [2008] proposed a similarity fusion strategy that calculates the user-user (or item-

item) similarity based on both tags and ratings within a memory-based CF framework.

Subsequently, Tagommenders [Sen et al. 2009] were proposed as a group of tag-based

recommendation algorithms that utilize the inferred preferences for tags to predict the

users’ preferences for items. In order to address the noisiness of tags, schemes for tag

enhancement have been introduced. For example, different weights have been used

for different tags, which are then combined with traditional memory-based CF [Liang

et al. 2010]. In addition to tags, geotags have also been exploited. Speciﬁcally, they have

been used to extend memory-based CF for personalized location prediction [Clements

et al. 2010b], shop recommendation [Takeuchi and Sugimoto 2006], and restaurant

recommendation [Horozov et al. 2006].

ACM Computing Surveys, Vol. 47, No. 1, Article 3, Publication date: April 2014.

Collaborative Filtering beyond the User-Item Matrix 3:9

Fig. 3. A conceptual view of the algorithms incorporating the rich side information of users and items to

extend model-based CF.

As mentioned in Section 3.1.1, social networks are one of the most important sources

of side information. In analogy to nearest neighbors in user-based CF, a social net-

work naturally provides the neighborhood (i.e., similar users) for each user—for ex-

ample, the t rustees of each user in a trust network [Massa and Bhattacharjee 2004].

TidalTrust [Golbeck 2005] and MoleTrust [Massa and Avesani 2007] are the two most

well known models that predict a user’s rating on an item by aggregating the ratings

from the user’s trustees (or connected users in other types of networks). Subsequently,

TrustWalker [Jamali and Ester 2009a], a random walk model that integrates trust

information with traditional item-based CF, has also been proposed. Using both the

social network and the user-contributed tags, item recommendations can be generated

by taking into account user-user similarities based on the two different information

sources, as proposed by Guy et al. [2010].

4.1.2. Extending Model-Based Collaborative Filtering.

In addition to memory-based CF, ex-

tensive work has also been devoted to extending model-based CF approaches to in-

corporate rich side information. In general, the algorithms in this direction provide

recommendations to individual users by training a predictive function that is based

on all of the variables, including the U-I matrix and the side information of users and

items, as illustrated in Figure 3.

The earliest work in this direction focused on the classiﬁcation perspective. For

example, each user can be treated as a separate classiﬁcation problem [Basu et al.

1998; Billsus and Pazzani 1998]. Under t his approach, each i tem is treated as a training

instance and is represented by a set of features, such as its ratings from all users and

associated side information (e.g., attributes). Then, a classiﬁer is trained for each user

and used to predict relevant items for recommendation. Similarly, one can treat each

item or each rating as a classiﬁcation problem [Basilico and Hofmann 2004; Pazzani

1999] by taking all available ratings and/or side information of users and items as

features of training instances. The key drawback of these classiﬁcation models lies in

their limited ability to generalize, and speciﬁcally, in the danger of overﬁtting, because

these classiﬁers are usually trained on a limited number of high-dimensional feature

vectors.

In addition to classiﬁcation models, models that are based on topics have also been

proposed. Popescul et al. [2001] have extended the aspect/topic model of Hofmann

et al. [1999] to incorporate the item side information. Wang et al. [2006] proposed a

generalization of latent semantic analysis (LSA) [Hofmann 2004] for this purpose. CF

with side information was shown to be an application of this generalized LSA model.

ACM Computing Surveys, Vol. 47, No. 1, Article 3, Publication date: April 2014.

剩余44页未读，继续阅读

u010012178

粉丝: 0
资源: 2

2014年协同过滤技术综述：超越用户-物品矩阵的创新与发展

协同过滤推荐算法综述

智能推荐系统中协同过滤算法综述.docx

基于协同过滤音乐推荐文献综述

关于协同过滤算法的优缺点

基于用户协同过滤和基于物品协同过滤的区别

协同过滤算法Java

协同过滤算法参考文献

Python 协同过滤算法

Amazon推荐系统,关于协同过滤算法

近年来关于基于用户的协同过滤的推荐算法的中文参考文献

最新资源