TowardtheNextGenerationofRecommenderSystems:ASurveyoftheState-of-the-ArtandPossibleExtensions资源-CSDN文库

电子商务

4星 · 超过85%的资源需积分: 16 124 浏览量 2011-02-26 14:43:06 上传评论收藏 556KB PDF 举报

资源推荐

资源详情

资源评论

Toward the Next Generation of Recommender

Systems: A Survey of the State-of-the-Art and

Possible Extensions

Gediminas Adomavicius, Member, IEEE, and Alexander Tuzhilin, Member, IEEE

Abstract—This paper presents an overview of the field of recommender systems and describes the current generation of

recommendation methods that are usually classified into the following three main categories: content-based, collaborative, and hybrid

recommendation approaches. This paper also describes various limitations of current recommendation methods and discusses

possible extensions that can improve recommendation capabilities and make recommender systems applicable to an even broader

range of applications. These extensions include, among others, an improvement of understanding of users and items, incorporation of

the contextual information into the recommendation process, support for multcriteria ratings, and a provision of more flexible and less

intrusive types of recommendations.

Index Terms—Recommender systems, collaborative filtering, rating estimation methods, extensions to recommender systems.

1INTRODUCTION

ECOMMENDER systems have become an important

research area since the appearance of the first papers

on collaborative filtering in the mid-1990s [45], [86], [97].

There has been much work done both in the industry and

academia on developing new approaches to recommender

systems over the last decade. The interest in this area still

remains high because it constitutes a problem-rich

research area and because of the abundance of practical

applications that help users to deal with information

overload and provide personalized recommendations,

content, and services to them. Examples of such applica-

tions include recommending books, CDs, and other

products at Amazon.com [61], movies by MovieLens

[67], and news at VERSIFI Technologies (formerly

AdaptiveInfo.com) [14]. Moreover, some of the vendors

have incorporated recommendation capabilities into their

commerce servers [78].

However, despite all of these advances, the current

generation of recommender systems still requires further

improvements to make recommendation methods more

effective and applicable to an even broader range of real-life

applications, including recommending vacations, certain

types of financial services to investors, and products to

purchase in a store made by a “smart” shopping cart [106].

These improvements include better methods for represent-

ing user behavior and the information about the items to be

recommended, more advanced recommendation modeling

methods, incorporation of various contextual information

into the recommendation process, utilization of multcriteria

ratings, development of less intrusive and more flexible

recommendation methods that also rely on the measures

that more effectively determine performance of recommen-

der systems.

In this paper, we describe various ways to extend the

capabilities of recommender systems. However, before

doing this, we first present a comprehensive survey of the

state-of-the-art in recommender systems in Section 2. Then,

we identify various limitations of the current generation of

recommendation methods and discuss some initial ap-

proaches to extending their capabilities in Section 3.

2THE SURVEY OF RECOMMENDER SYSTEMS

Although the roots of recommender systems can be traced

back to the extensive work in cognitive science [87],

approximation theory [81], information retrieval [89],

forecasting theories [6], and also have links to management

science [71] and to consumer choice modeling in marketing

[60], recommender systems emerged as an independent

research area in the mid-1990s when researchers started

focusing on recommendation problems that explicitly rely

on the ratings structure. In its most common formulation,

the recommendation problem is reduced to the problem of

estimating ratings for the items that have not been seen by a

user. Intuitively, this estimation is usually based on the

ratings given by this user to other items and on some other

information that will be formally described below. Once we

can estimate ratings for the yet unrated items, we can

recommend to the user the item(s) with the highest

estimated rating(s).

More formally, the recommendation problem can be

formulated as follows: Let C be the set of all users and let S

be the set of all possible items that can be recommended,

such as books, movies, or restaurants. The space S of

734 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 17, NO. 6, JUNE 2005

. G. Adomavicius is with the Carlson School of Management, University of

Minnesota, 321 19th Avenue South, Minneapolis, MN 55455.

E-mail: gedas@umn.edu.

. A. Tuzhilin is with the Stern School of Business, New York University,

44 West 4th Street, New York, NY 10012. E-mail: atuzhili@stern.nyu.edu.

Manuscript received 8 Mar. 2004; revised 14 Oct. 2004; accepted 10 Dec.

2004; published online 20 Apr. 2005.

For information on obtaining reprints of this article, please send e-mail to:

tkde@computer.org, and reference IEEECS Log Number TKDE-0071-0304.

1041-4347/05/$20.00 ß 2005 IEEE Published by the IEEE Computer Society

possible items can be very large, ranging in hundreds of

thousands or even millions of items in some applications,

such as recommending books or CDs. Similarly, the user

space can also be very large—millions in some cases. Let u

be a utility function that measures the usefulness of item s

to user c, i.e., u : C  S ! R, where R is a totally ordered set

(e.g., nonnegative integers or real numbers within a certain

range). Then, for each user c 2 C, we want to choose such

item s

2 S that maximizes the user’s utility. More formally:

8c 2 C; s

¼ arg max

s2S

uðc; sÞ: ð1Þ

In recommender systems, the utility of an item is usually

represented by a rating, which indicates how a particular user

liked a particular item, e.g., John Doe gave the movie “Harry

Potter” the rating of 7 (out of 10). However, as indicated

earlier, in general, utility can be an arbitrary function,

including a profit function. Depending on the application,

utility u can either be specified by the user, as is often done for

the user-defined ratings, or is computed by the application, as

can be the case for a profit-based utility function.

Each element of the user space C can be defined with a

profile that includes various user characteristics, such as age,

gender, income, marital status, etc. In the simplest case, the

profile can contain only a single (unique) element, such as

User ID. Similarly, each element of the item space S is

defined with a set of characteristics. For example, in a

movie recommendation application, where S is a collection

of movies, each movie can be represented not only by its ID,

but also by its title, genre, director, year of release, leading

actors, etc.

The central problem of recommender systems lies in that

utility u is usually not defined on the whole C  S space,

but only on some subset of it. This means u needs to be

extrapolated to the whole space C  S. In recommender

systems, utility is typically represented by ratings and is

initially defined only on the items previously rated by the

users. For example, in a movie recommendation application

(such as the one at MovieLens.org), users initially rate some

subset of movies that they have already seen. An example

of a user-item rating matrix for a movie recommendation

application is presented in Table 1, where ratings are

specified on the scale of 1 to 5. The “” symbol for some of

the ratings in Table 1 means that the users have not rated

the corresponding movies. Therefore, the recommendation

engine should be able to estimate (predict) the ratings of the

nonrated movie/user combinations and issue appropriate

recommendations based on these predictions.

Extrapolations from known to unknown ratings are

usually done by 1) specifying heuristics that define the

utility function and empirically validating its performance

and 2) estimating the utility function that optimizes certain

performance criterion, such as the mean square error.

Once the unknown ratings are estimated, actual

recommendations of an item to a user are made by

selecting the highest rating among all the estimated

ratings for that user, according to (1). Alternatively, we

can recommend the N best items to a user or a set of

users to an item.

The new ratings of the not-yet-rated items can be

estimated in many different ways using methods from

machine learning, approximation theory, and various

heuristics. Recommender systems are usually classified

according to their approach to rating estimation and, in the

next section, we will present such a classification that was

proposed in the literature and will provide a survey of

different types of recommender systems. The commonly

accepted formulation of the recommendation problem was

first stated in [45], [86], [97] and this problem has been

studied extensively since then. Moreover, recommender

systems are usually classified into the following categories,

based on how recommendations are made [8]:

. Content-based recommendations: The user will be

recommended items similar to the ones the user

preferred in the past;

. Collaborative recommendations:Theuserwillbe

recommended items that people with similar tastes

and preferences liked in the past;

. Hybrid approaches: These methods combine colla-

borative and content-based methods.

In addition to recommender systems that predict the

absolute values of ratings that individual users would give

to the yet unseen items (as discussed above), there has been

work done on preference-based filtering, i.e., predicting the

relative preferences of users [22], [35], [51], [52]. For

example, in a movie recommendation application, prefer-

ence-based filtering techniques would focus on predicting

the correct relative order of the movies, rather than their

individual ratings. However, this paper focuses primarily

on rating-based recommenders since it constitutes the most

popular approach to recommender systems.

2.1 Content-Based Methods

In content-based recommendation methods, the utility

uðc; sÞ of item s for user c is estimated based on the utilities

uðc; s

Þ assigned by user c to items s

2 S that are “similar”

to item s. For example, in a movie recommendation

application, in order to recommend movies to user c, the

content-based recommender system tries to understand the

commonalities among the movies user c has rated highly in

the past (specific actors, directors, genres, subject matter,

ADOMAVICIUS AND TUZHILIN: TOWARD THE NEXT GENERATION OF RECOMMENDER SYSTEMS: A SURVEY OF THE STATE-OF-THE-ART... 735

TABLE 1

A Fragment of a Rating Matrix for a Movie Recommender System

etc.). Then, only the movies that have a high degree of

similarity to whatever the user’s preferences are would be

recommended.

The content-based approach to recommendation has its

roots in information retrieval [7], [89] and information

filtering [10] research. Because of the significant and early

advancements made by the information retrieval and

filtering communities and because of the importance of

several text-based applications, many current content-based

systems focus on recommending items containing textual

information, such as documents, Web sites (URLs), and

Usenet news messages. The improvement over the tradi-

tional information retrieval approaches comes from the use

of user profiles that contain information about users’ tastes,

preferences, and needs. The profiling information can be

elicited from users explicitly, e.g., through questionnaires,

or implicitly—learned from their transactional behavior

over time.

More formally, let ContentðsÞ be an item profile, i.e., a set

of attributes characterizing item s. It is usually computed by

extracting a set of features from item s (its content) and is

used to determine the appropriateness of the item for

recommendation purposes. Since, as mentioned earlier,

content-based systems are designed mostly to recommend

text-based items, the content in these systems is usually

described with keywords. For example, a content-based

component of the Fab system [8], which recommends Web

pages to users, represents Web page content with the

100 most important words. Similarly, the Syskill & Webert

system [77] represents documents with the 128 most

informative words. The “importance” (or “informative-

ness”) of word k

in document d

is determined with some

weighting measure w

that can be defined in several

different ways.

One of the best-known measures for specifying keyword

weights in Information Retrieval is the term frequency/inverse

document frequency (TF-IDF) measure [89] that is defined as

follows: Assume that N is the total number of documents

that can be recommended to users and that keyword k

appears in n

of them. Moreover, assume that f

i;j

is the

number of times keyword k

appears in document d

. Then,

i;j

, the term frequency (or normalized frequency) of

keyword k

in document d

, is defined as

i;j

max

z;j

; ð2Þ

where the maximum is computed over the frequencies f

z;j

of all keywords k

that appear in the document d

However, keywords that appear in many documents are

not useful in distinguishing between a relevant document

and a nonrelevant one. Therefore, the measure of inverse

document frequency ðIDF

Þ is often used in combination

with simple term frequency ðTF

i;j

Þ. The inverse document

frequency for keyword k

is usually defined as

IDF

¼ log

: ð3Þ

Then, the TF-IDF weight for keyword k

in document d

defined as

i;j

¼ TF

i;j

 IDF

ð4Þ

and the content of document d

is defined as

Contentðd

Þ¼ðw

; ...w

Þ:

As stated earlier, content-based systems recommend

items similar to those that a user liked in the past [56], [69],

[77]. In particular, various candidate items are compared

with items previously rated by the user and the best-

matching item(s) are recommended. More formally, let

ContentBasedP rofileðcÞ be the profile of user c containing

tastes and preferences of this user. These profiles are

obtained by analyzing the content of the items previously

seen and rated by the user and are usually constructed

using keyword analysis techniques from information

retrieval. For example, ContentBasedP rofileðcÞ can be

defined as a vector of weights ðw

; ...;w

Þ, where each

weight w

denotes the importance of keyword k

to user c

and can be computed from individually rated content

vectors using a variety of techniques. For example, some

averaging approach, such as Rocchio algorithm [85], can be

used to compute ContentBasedP rofileðcÞ as an “average”

vector from an individual content vectors [8], [56]. On the

other hand, [77] uses a Bayesian classifier in order to

estimate the probability that a document is liked. The

Winnow algorithm [62] has also been shown to work well

for this purpose, especially in the situations where there are

many possible features [76].

In content-based systems, the utility function uðc; sÞ is

usually defined as:

uðc; sÞ¼scoreðContentBasedP rofileðcÞ; Content ðs ÞÞ: ð5Þ

Using the above-mentioned information retrieval-based

paradigm of recommending Web pages, Web site URLs,

or Usenet news messages, both ContentBasedP rofileðcÞ of

user c and Content ðs Þ of document s can be represented as

TF-IDF vectors

and

of keyword weights. Moreover,

utility function uðc; sÞ is usually represented in the

information retrieval literature by some scoring heuristic

defined in terms of vectors

and

, such as the cosine

similarity measure [7], [89]:

uðc; sÞ¼cosð

;

Þ¼



jj

i¼1

i;c

i;s

ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ

i¼1

i;c

q ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ

i¼1

i;s

;

ð6Þ

where K is the total number of keywords in the system.

For example, if user c reads many online articles on the

topic of bioinformatics, then content-based recommenda-

tion techniques will be able to recommend other bioinfor-

matics articles to user c. This is the case because these

articles will have more bioinformatics-related terms (e.g.,

“genome,” “sequencing,” “proteomics”) than articles on

other topics and, therefore, ContentBasedP rofileðcÞ,as

defined by vector

, will represent such terms k

with

high weights w

. Consequently, a recommender system

using the cosine or a related similarity measure will assign

higher utility uðc; sÞ to those articles s that have high-

weighted bioinformatics terms in

and lower utility to the

ones where bioinformatics terms are weighted less.

736 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 17, NO. 6, JUNE 2005

Besides the traditional heuristics that are based mostly

on information retrieval methods, other techniques for

content-based recommendation have also been used, such

as Bayesian classifiers [70], [77] and various machine

learning techniques, including clustering, decision trees,

and artificial neural networks [77]. These techniques differ

from information retrieval-based approaches in that they

calculate utility predictions based not on a heuristic

formula, such as a cosine similarity measure, but rather

are based on a model learned from the underlying data

using statistical learning and machine learning techni-

ques. For example, based on a set of Web pages that were

rated as “relevant” or “irrelevant” by the user, [77] uses

the naive Bayesian classifier [31] to classify unrated Web

pages. More specifically, the naive Bayesian classifier is

used to estimate the following probability that page p

belongs to a certain class C

(e.g., relevant or irrelevant)

given the set of keywords k

1;j

; ...;k

n;j

on that page:

P ðC

1;j

&...&k

n;j

Þ: ð7Þ

Moreover, [77] uses the assumption that keywords are

independent and, therefore, the above probability is

proportional to

P ðC

P ðk

x;j

Þ: ð8Þ

While the keyword independence assumption does not

necessarily apply in many applications, experimental results

demonstrate that naı

ve Bayesian classifiers still produce

high classification accuracy [77]. Furthermore, both

P ðk

x;j

Þ and P ðC

Þ can be estimated from the underlying

training data. Therefore, for each page p

, the probability

P ðC

1;j

&...&k

n;j

Þ is computed for each class C

and page p

is assigned to class C

having the highest probability [77].

While not explicitly dealing with providing recommen-

dations, the text retrieval community has contributed several

techniques that are being used in content-based recommen-

der systems. One example of such a technique would be the

research on adaptive filtering [101], [112], which focuses on

becoming more accurate at identifying relevant documents

incrementally by observing the documents one-by-one in a

continuous document stream. Another example would be

the work on threshold setting [84], [111], which focuses on

determining the extent to which documents should match a

given query in order to be relevant to the user. Other text

retrieval methods are described in [50] and can also be

found in the proceedings of the Text Retrieval Conference

(TREC) (http://trec.nist.gov).

As was observed in [8], [97], content-based recommender

systems have several limitations that are described in the

rest of this section.

2.1.1 Limited Content Analysis

Content-based techniques are limited by the features that

are explicitly associated with the objects that these systems

recommend. Therefore, in order to have a sufficient set of

features, the content must either be in a form that can be

parsed automatically by a computer (e.g., text) or the

features should be assigned to items manually. While

information retrieval techniques work well in extracting

features from text documents, some other domains have an

inherent problem with automatic feature extraction. For

example, automatic feature extraction methods are much

harder to apply to multimedia data, e.g., graphical images,

audio streams, and video streams. Moreover, it is often not

practical to assign attributes by hand due to limitations of

resources [97].

Another problem with limited content analysis is that, if

two different items are represented by the same set of

features, they are indistinguishable. Therefore, since text-

based documents are usually represented by their most

important keywords, content-based systems cannot distin-

guish between a well-written article and a badly written

one, if they happen to use the same terms [97].

2.1.2 Overspecializati on

When the system can only recommend items that score

highly against a user’s profile, the user is limited to being

recommended items that are similar to those already rated.

For example, a person with no experience with Greek

cuisine would never receive a recommendation for even the

greatest Greek restaurant in town. This problem, which has

also been studied in other domains, is often addressed by

introducing some randomness. For example, the use of

genetic algorithms has been proposed as a possible solution

in the context of information filtering [98]. In addition, the

problem with overspecialization is not only that the

content-based systems cannot recommend items that are

different from anything the user has seen before. In certain

cases, items should not be recommended if they are too

similar to something the user has already seen, such as a

different news article describing the same event. Therefore,

some content-based recommender systems, such as Daily-

Learner [13], filter out items not only if they are too different

from the user’s preferences, but also if they are too similar

to something the user has seen before. Furthermore, Zhang

et al. [112] provide a set of five redundancy measures to

evaluate whether a document that is deemed to be relevant

contains some novel information as well. In summary, the

diversity of recommendations is often a desirable feature in

recommender systems. Ideally, the user should be pre-

sented with a range of options and not with a homogeneous

set of alternatives. For example, it is not necessarily a good

idea to recommend all movies by Woody Allen to a user

who liked one of them.

2.1.3 New User Problem

The user has to rate a sufficient number of items before a

content-based recommender system can really understand

the user’s preferences and present the user with reliable

recommendations. Therefore, a new user, having very few

ratings, would not be able to get accurate recommendations.

2.2 Collaborative Methods

Unlike content-based recommendation methods, collabora-

tive recommender systems (or collaborative filtering systems)

try to predict the utility of items for a particular user based

on the items previously rated by other users. More formally,

the utility uðc; sÞ of item s for user c is estimated based on

the utilities uðc

;sÞ assigned to item s by those users c

2 C

who are “similar” to user c. For example, in a movie

ADOMAVICIUS AND TUZHILIN: TOWARD THE NEXT GENERATION OF RECOMMENDER SYSTEMS: A SURVEY OF THE STATE-OF-THE-ART... 737

剩余15页未读，继续阅读

评论收藏

内容反馈

zcy304907240

2012-09-24

还可以，但是有点难度，有些地方看不懂
draflyPHD

2012-11-27

推荐算法的经典论文
gumutengfei

2012-09-13

是真的论文资源，受到推荐来搜索下载的
c8s123

2014-10-09

现在看有点老了，可以看看2012的一个综述
bludes

2015-09-11

google都没有下载的，都是要收费的，太感谢了。

yangruihaha

粉丝: 0
资源: 3

Toward the Next Generation of Recommender Systems: A Survey of t...

最新资源

Toward the Next Generation of Recommender Systems: A Survey of t...

推荐系统的循序进阶读物

推荐系统的循序进阶读物（从入门到精通）

Internet of things and big data analytics toward next-generation intelligence

The Art of Memory Forensic

The Art of Memory Forensics

Supervisory Control of Concurrent Systems-A Petri Net Structural Approach

UML and the Unified Process.pdf

Space-Time Codes and MIMO Systems 书和代码 纯英文版

The entity-relationship model toward a unified view of data

Trends and Advances in Information Systems and Technologies Vol1-Springer2018

Yunus, Muhammad - Creating A World Without Poverty; Social Business and the Future of Capitalism

the entity-relationship model ---toward a unified view of data

The Psychology of Money

Big Data and Computational Intelligence in Networking-CRC(2018).pdf

Project Management of Complex and Embedded Systems

Architects of Intelligence: The truth about AI from the people building it

A Review of Computer Vision Techniques for the Analysis of Urban Traffic

微信小程序源码-合集6.rar

微信小程序源码-合集5.rar

微信小程序源码-合集4.rar

微信小程序源码-合集3.rar

微信小程序源码-合集2.rar

浏览器插件 Auto Refresh Plus 7.4.4 ctx

品优购项目 素材及代码

学生宿舍管理系统源码文件

全能电子地图下载器1.9.5完善版.zip

大学生网页设计大作业-5个网页设计制作作品自己任选

数据可视化大屏资料合集（网上收集的几十种样式的html静态页）

最新资源

Space-Time Codes and MIMO Systems 书和代码纯英文版

品优购项目素材及代码