机器学习范式在推荐系统中的应用

需积分: 10 144 浏览量更新于2024-07-21 收藏 4.69MB PDF 举报

"Machine Learning Paradigms - Applications in Recommender Systems" 是一本由 Springer 在2015年出版的专业书籍，作者是 Aristomenis S. Lampropoulos 和 George A. Tsihrintzis。本书关注的是在推荐系统中如何运用机器学习算法，特别是面对稀疏或缺乏负面样本的情况。在推荐系统领域，传统的基于用户喜好内容的机器学习方法往往在只有少量或没有负面反馈的情况下表现不佳。这本书引入了一种新颖的方法，即基于单类分类（One-Class Classification）的机器学习技术。单类分类是一种专门处理仅有正面样本而缺乏负面样本的数据问题的机器学习方法。将这种方法与推荐系统相结合，为解决“大数据”和“稀疏数据”问题提供了新的研究、创新和发展方向。本书的目标读者包括从事大规模复杂数据研究的研究人员、实践者以及研究生。无论是模式识别、机器学习还是推荐系统的专家，还是希望了解推荐系统及其应用的计算机科学领域的普通读者，都能从书中获益。此外，书中的广泛引用文献列表为读者深入研究相关领域提供了全面的参考资料。推荐系统是人工智能和数据科学的一个重要分支，它们通过分析用户的偏好来提供个性化的建议。这本书的内容可能涵盖了如何构建有效的推荐模型，如何处理数据稀疏性问题，以及如何利用单类分类算法来提高推荐精度等主题。此外，书中可能还讨论了如何应对大规模数据的挑战，以及如何将这些理论应用于实际的商业场景，如电商、媒体推荐等。 "Machine Learning Paradigms - Applications in Recommender Systems" 提供了一个深入了解推荐系统如何利用先进机器学习策略的平台，对于想要在这个领域深化知识和技能的人来说，是一本不可多得的资源。

2 1 Introduction

speciﬁc querying user needs. Thus, unlike search engines or information retrieval

systems, a RS provides information in a way that will be most appropriate and valu-

able to its users and prevents them from being overwhelmed by huge amounts of

information that, in the absence of RS, they should browse or examine. This is to

be contrasted with the target of a search engine or an information retrieval system

which is to “match” items to the user query. This means that a search engine or an

information retrieval system tries to form and return a ranked list of all those items

that match the query. Techniques of active learning such as relevance-feedback may

give these systems the ability to reﬁne their results according to the user preferences

and, thus, provide a simple form of recommendation. More complex search engines

such as GOOGLE utilize other kinds of criteria such as “authoritativeness”, which

aim at returning as many useful results as possible, but not in an individualized way.

A learning-based RS typically works as follows: (1) the recommender system

collects all given recommendations at one place and (2) applies a learning algorithm,

thereafter. Predictions are then made either with a model learnt from the dataset

(model-based predictions) using, for example, a clustering algorithm [3, 18]oron

the ﬂy (memory-based predictions) using, for example, a nearest neighbor algorithm

[3, 15]. A typical prediction can be a list of the top-N recommendations or a requested

prediction for a single item [7].

Memory-based methods store training instances during training which are can

be retrieved when making predictions. In contrast, model-based methods generalize

into a model from the training instances during training and the model needs to

be updated regularly. Then, the model is used to make predictions. Memory-based

methods learn fast but make slow predictions, while model-based methods make fast

predictions but learn slowly.

The roots of RS can be traced back to Malone et al. [11], who proposed three forms

of ﬁltering: cognitive ﬁltering (now called content-based ﬁltering), social ﬁltering

(now called collaborative ﬁltering (CF)) and economic ﬁltering. They also suggested

that the best approach was probably to combine these approaches into the category

of, so-called, hybrid RS.

1.2 Formulation of the Recommendation Problem

In general, the recommendation problem is deﬁned as the problem of estimating

ratings for the items that have not been seen by a user. This estimation is based on:

• ratings given by the user to other items,

• ratings given to an item by other users,

• and other user and item information (e.g. item characteristics, user demographics).

The recommendation problem can be formulated [1] as follows:

Let U be the set of all users U ={u

, u

, ..., u

} and let I be the set of all

possible items I ={i

, i

, ..., i

} that can be recommended, such as music ﬁles,

images, movies, etc. The space I of possible items can be very large.

1.2 Formulation of the Recommendation Problem 3

Let f be a utility function that measures the usefulness of item i to user u,

f : U × I → R, (1.1)

where R is a totally ordered set (e.g. the set of nonnegative integers or real numbers

within a certain range). Then, for each user u ∈ U, we want to choose an item i



∈ I

that maximizes the user utility function, i.e.

∀u ∈ U, i



= arg max

i∈I

f (u, i ). (1.2)

In RS, the utility of an item is usually represented by a rating, which indicates

how a particular user liked a particular item, e.g., user u

gave the object i

the rating

of R(1, 1) = 3, where R(u, i) ∈{1, 2, 3, 4, 5}.

Each user u

, where k = 1, 2, ..., m, has a list of items I

about which the user

has expressed his/her preferences. It is important to note that I

⊆ I , while it is

also possible for I

to be the null set. This latter means that users are not required

to express their preferences for all existing items.

Each element of the user space U 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.

Recommendation algorithms enhance various techniques by operating

• either on rows of the matrix R, which correspond to ratings of a single user about

different items,

• or on columns of the matrix R, which correspond to different users’ ratings for a

single item.

However, in general, the utility function can be an arbitrary function, including a

proﬁt function. Depending on the application, a utility f can either be speciﬁed by

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

as can be the case for a proﬁt-based utility function. Each element of the user space

U 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 I is deﬁned via a set of characteristics.

The central problem of RS lies in that a utility function f is usually not deﬁned on

the entire U × I space, but only on some subset of it. This means that f needs to

be generalized to the entire space U × I . In RS, a utility is typically represented by

ratings and is initially deﬁned only on the items previously rated by the users.

Generalizations from known to unknown ratings are usually done by:

• specifying heuristics that deﬁne the utility function and empirically validating its

performance, or

• estimating the utility function that optimizes a certain performance criterion, such

as Mean Absolute Error (MAE).

4 1 Introduction

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 Eq. 1.2. Alternatively, we can recommend the N best items to

a user. Additionally, we can recommend a set of users to an item.

1.2.1 The Input to a Recommender System

The input to a RS depends on the type of the ﬁltering algorithm employed. The input

belongs to one of the following categories:

1. Ratings (also called votes), which express the opinion of users on items. Ratings

are normally provided by the user and follow a speciﬁed numerical scale (example:

1-bad to 5-excellent). A common rating scheme is the binary rating scheme, which

allows only ratings of either 0 or 1. Ratings can also be gathered implicitly from

the users purchase history, web logs, hyper-link visits, browsing habits or other

types of information access patterns.

2. Demographic data, which refer to information such as the age, the gender and

the education of the users. This kind of data is usually difﬁcult to obtain. It is

normally collected explicitly from the user.

3. Content data, which are based on content analysis of items rated by the user. The

features extracted via this analysis are used as input to the ﬁltering algorithm in

order to infer a user proﬁle.

1.2.2 The Output of a Recommender System

The output of a RS can be either a prediction or a recommendation.

• A prediction is expressed as a numerical value, R

a, j

= R(u

, i

), which represents

the anticipated opinion of active user u

for item i

. This predicted value should

necessarily be within the same numerical scale (example: 1-bad to 5-excellent) as

the input referring to the opinions provided initially by active user u

. This form

of RS output is also known as Individual Scoring.

• A recommendation is expressed as a list of N items, where N ≤ n, which the active

user is expected to like the most. The usual approach in that case requires this list to

include only items that the active user has not already purchased, viewed or rated.

This form of RS output is also known as Top-N Recommendation or Ranked

Scoring.

1.3 Methods of Collecting Knowledge About User Preferences 5

1.3 Methods of Collecting Knowledge About

User Preferences

To generate personalized recommendations that are tailored to the speciﬁc needs of

the active user, RS collect ratings of items by users and build user-proﬁles in ways

that depend on the methods that the RS utilize to collect personal information about

user preferences. In general, these methods are categorized into three approaches:

• an Implicit approach, which is based on recording user behavior,

• an Explicit approach, which is based on user interrogation,

• a Mixing approach, which is a combination of the previous two.

1.3.1 The Implicit Approach

This approach does not require active user involvement in the knowledge acquisition

task, but, instead, the user behavior is recorded and, speciﬁcally, the way that he/she

reacts to each incoming piece of data. The goal is to learn from the user reaction

about the relevance of the data item to the user. Typical examples for implicit ratings

are purchase data or reading time of Usenet news [15]. In the CF system in [9], they

monitored reading times as an indicator for relevance. This revealed a relationship

between time spent on reviewing data items and their relevance. In [6], the system

learns the user proﬁle by passively observing the hyperlinks clicked on and those

passed over and by measuring user mouse and scrolling activity in addition to user

browsing activity. Also, in [14] they utilize agents that operate as adaptive Web site

RS. Through analysis of Web logs and web page structure, the agents infer knowl-

edge of the popularity of various documents as well as a combination of document

similarity. By tracking user actions and his/her acceptance of the agent recommen-

dations, the agent can make further estimations about future recommendations to the

speciﬁc user. The main beneﬁts of implicit feedback over explicit ratings are that

they remove the cognitive cost of providing relevance judgements explicitly and can

be gathered in large quantities and aggregated to infer item relevance [8].

However, the implicit approach bears some serious implications. For instance,

some purchases are gifts and, thus, do not reﬂect the active user interests. More-

over, the inference that purchasing implies liking does not always hold. Owing to the

difﬁculty of acquiring explicit ratings, some providers of product recommendation

services adopt bilateral approaches. For instance, Amazon.com computes recom-

mendations based on explicit ratings whenever possible. In case of unavailability,

observed implicit ratings are used instead.

6 1 Introduction

1.3.2 The Explicit Approach

Users are required to explicitly specify their preference for any particular item, usu-

ally by indicating their extent of appreciation on 5-point or 7-point Thurstone scales.

These scales are mapped to numeric values, e.g. R

i, j

∈[1, 2, 3, 4, 5]. Lower val-

ues commonly indicate least favorable preferences, while higher values express the

user’s liking.

Explicit ratings impose additional efforts on users. Consequently, users

often tend to avoid the burden of explicitly stating their preferences and either leave

the system or rely upon “free-riding” [2]. Ratings made on these scales allow these

judgments to be processed statistically to provide averages, ranges, or distributions.

A central feature of explicit ratings is that the user who evaluates items has to exam-

ine them and, then, to assign to them values from the rating scale. This imposes a

cognitive cost on the evaluator to assess the performance of an object [12].

1.3.3 The Mixing Approach

Newsweeder [10], a Usenet ﬁltering system, is an example of a system that uses

a combination of the explicit and the implicit approach, as it requires minimum

user involvement. In this system, the users are required to rate documents for their

relevance. The ratings are used as training examples for a machine learning algorithm

that is executed nightly to generate user interest proﬁles for the next day. Newsweeder

is successful in reducing user involvement. However, the batch proﬁling used in

Newsweeder is a shortcoming as proﬁle adaptation is delayed signiﬁcantly.

1.4 Motivation of the Book

The motivation of this book is based on the following facts that constitute important

open research problems in RS. It is well known that users hardly provide explicit

feedbacks in RS. More speciﬁcally, users tend to provide ratings only for items

that they are interested in and belong to their preferences and avoid, to provide

feedback in the form of negative examples, i.e. items that they dislike or they are

not interested in. As stated in [5, 17], “It has been known for long time in human

computer interaction that users are extremely reluctant to perform actions that are

not directed towards their immediate goal if they do not receive immediate beneﬁts”.

However, common RS based on machine learning approaches use classiﬁers that, in

order to learn user interests, require both positive (desired items that users prefer) and

The Thurstone scale was used in psychology for measuring an attitude. It was developed by Louis

Leon Thurstone in 1928, as a means of measuring attitudes towards religion. It is made up of

statements about a particular issue. A numerical value is associated with each statement, indicating

how favorable or unfavorable the statement is judged to be.

www.allitebooks.com

剩余134页未读，继续阅读

vanridin

粉丝: 108
资源: 1187

机器学习范式在推荐系统中的应用

"深入了解计算机体系结构：理论与实践"。

分布式系统原理与范型

算法设计指南：方法、范式与复杂性分析

Machine Learning Paradigms

Machine Learning Paradigms: Theory and Application

Paradigms-of-Artificial-Intelligence-Programming.pdf

clojure-conj-2014-paradigms-of-core-async:Clojure Conj 2014 演讲 - 核心异步范式

Software Paradigms - Stephen H. Kaisler

Programming-Paradigms-DD1361.Inet:实验室网络

Understanding Machine Learning - From Theory to Algorithms.zip

最新资源