推荐系统文献综述：深度探索与分析

需积分: 9 85 浏览量更新于2024-07-17 收藏 1.65MB PDF 举报

"这篇PDF文献是对研究论文推荐系统的一个详尽的文献综述，适合对推荐系统感兴趣的初学者。文章作者包括Joeran Beel、Bela Gipp、Stefan Langer和Corinna Breitinger，来自Docear、Konstanz大学和林奈大学。文献中对过去十六年间发表的200多篇关于研究论文推荐系统的研究文章进行了回顾，并提供了相关的描述性统计，讨论了主要的发展成就、不足之处以及最常见的推荐概念和方法。" 在推荐系统领域，这篇文献调查报告是一个重要的参考资料。它深入探讨了在众多推荐算法中，内容过滤（Content-Based Filtering）占据主导地位，约55%的推荐方法采用了这一技术。内容过滤是基于用户过去的偏好或项目内容来推荐相似或匹配的物品，其核心是理解用户的历史行为和项目的特性，以此来预测用户可能的兴趣。另一方面，协同过滤（Collaborative Filtering）的应用则相对较少，仅占18%。协同过滤是通过分析用户之间的行为模式来预测用户可能喜欢的项目，分为用户-用户协同过滤和项目-项目协同过滤，尽管应用较少，但因其在大规模数据集上的有效性和广泛适用性而备受关注。此外，图基础推荐（Graph-Based Recommendations）占了16%，这类方法通常利用用户和项目之间的关系网络，通过计算节点间的相似度进行推荐。其他推荐概念如刻板印象（Stereotyping）、项目中心推荐（Item-Centric Recommendations）和混合推荐（Hybrid Recommendations）也被提及，这些方法结合了多种推荐策略，以提高推荐的准确性和多样性。文献综述还讨论了推荐系统的主要进步和缺陷。进步可能包括更有效的推荐算法、对用户动态兴趣的适应以及对冷启动问题的解决策略。不足之处可能涉及数据稀疏性、推荐结果的解释性以及用户隐私问题。通过对这些方法的分析，读者能获得推荐系统领域的全面理解，为进一步研究和开发提供基础。总结起来，这篇文献调查为想要深入了解推荐系统理论和实践的读者提供了宝贵的资源，涵盖了各种推荐策略、技术和当前的研究挑战，对于学术界和工业界的研究者都具有很高的参考价值。

User Studies

User studies typically measure user satisfaction through explicit ratings. Users receive

recommendations generated by different recommendation approaches, users rate the

recommendations, and the approach with the highest average rating is considered most

effective [264]. Study participants are typically asked to quantify their overall satisfaction

with the recommendations. However, they might also be asked to rate individual aspects

of a recommender system, for instance, how novel or authoritative recommendations are

[117], or how suitable they are for non-experts [72]. A user study can also collect

qualitative feedback, but qualitative feedback is rarely used in the field of (research-

paper) recommender systems [156], [159].

We distinguish between “lab” and “real-world” user studies. In lab studies, participants

are aware that they are part of a user study, which together with other factors might affect

user behavior and thereby the evaluation's results [267], [268]. In real-world studies,

participants are not aware of the study and rate recommendations for their own benefit,

for instance because the recommender system improves recommendations based on user

ratings (i.e. relevance feedback [263]), or user ratings are required to generate

recommendations (i.e. collaborative filtering [222]). All reviewed user studies were lab-

based.

Often, user studies are considered the optimal evaluation method [269]. However, the

outcome of user studies often depends on the questions asked. Cremonesi et al. found that

it makes a difference if users are asked for the “perceived relevance” or the "global

satisfaction" of recommendations [270]. Similarly, it made a difference whether users

were asked to rate the novelty or the relevance of recommendations [271]. A large number

of participants are also crucial to user study validity, which makes user studies relatively

expensive to conduct. The number of required participants, to receive statistically

significant results, depends on the number of approaches being evaluated, the number of

recommendations being displayed, and the variations in the results [272], [273].

However, as a rough estimate, at least a few dozen participants are required, often more.

Most participants rated only a few recommendations and four studies (15%) were

conducted with fewer than five participants [62], [123], [171]; five studies (19%) had five

to ten participants [66], [84], [101]; three studies (12%) had 11-15 participants [15], [146],

[185]; and five studies (19%) had 16-50 participants [44], [118], [121]. Six studies (23%)

were conducted with more than 50 participants [93], [98], [117]. Three studies (12%)

failed to mention the number of participants [55], [61], [149] (Table 4). Given these

findings, we conclude that most user studies were not large enough to arrive at meaningful

conclusions.

Table 4: Number of participants in user studies of reviewed recommendation approaches

Online Evaluations

Online evaluations were first used by the online advertising and e-commerce fields. They

measure the acceptance rates of recommendations in real-world recommender systems.

Acceptance rates are typically measured by click-through rates (CTR), i.e. the ratio of

clicked recommendations to displayed recommendations. For instance, if a recommender

system displays 10,000 recommendations and 120 are clicked, the CTR is 1.2%. Other

metrics include the ratio of downloaded or bought items to the number of items displayed.

Acceptance rate is typically interpreted as an implicit measure for user satisfaction. The

assumption is that when a user clicks, downloads, or buys a recommended item, the user

liked the recommendation. Of course, this assumption is not always reliable because users

might buy a book but rate it negatively after reading it. However, metrics such as CTR

can be an explicit measures of effectiveness, namely when the operator receives money,

e.g. for clicks on recommendations.

Online evaluations are not without drawbacks. Zheng et al. showed that CTR and

relevance do not always correlate and concluded that "CTR may not be the optimal metric

for online evaluation of recommender systems" and "CTR should be used with

precaution" [274]. In addition, conducting online evaluations requires significantly more

time than offline evaluations, they are more expensive, and they can only be conducted

by researchers who have access to a real-world recommender system.

Among the 75 approaches that included some form of evaluation, only six (8%) were

evaluated using an online evaluation [6], [92], [94]. Despite the active experimentation in

the field with a large number or evaluations being performed on research-paper

recommender systems, we observed that many researchers have no access to real-world

systems to evaluate their approaches. Interestingly, the researchers who do have access

to real-world recommender systems often do not make use of this resource to conduct

online evaluations, but rather perform offline evaluations or lab user studies. For instance,

Giles and his co-authors, who are some of the largest contributors in the field, could have

conducted online evaluations with their academic search engine CiteSeer. However, they

chose primarily to use offline evaluations. The reason for this might be that offline

evaluations are more convenient than conducting online evaluations or user studies.

Results are available within minutes or hours and not within days or weeks as is the case

for user studies and online evaluations. However, offline evaluations have a set of serious

drawbacks, as shown in the next section.

n/a <5 5-10 11-15 16-50 >50

Absolute 3 4 5 3 5 6

Relative 12% 15% 19% 12% 19% 23%

Number of Participants

Offline Evaluations

Offline evaluations typically measure the accuracy of a recommender system based on a

ground-truth. To measure accuracy, precision at position n (P@n) is often used to express

how many items of the ground-truth are recommended within the top n recommendations.

Other common evaluation metrics include recall, F-measure, mean reciprocal rank

(MRR), normalized discounted cumulative gain (nDCG), mean absolute error, and root

mean square error. Offline evaluations are also sometimes used to evaluate aspects such

as novelty or serendipity of recommendations [227]. For a comprehensive overview of

offline evaluations refer to [275], [276].

Offline evaluations were originally meant to identify a number of promising

recommendation approaches [225], [264], [277], [278]. These approaches should then be

evaluated in detail with a user study or online evaluation to identify the most effective

approaches. However, criticism has been raised on the assumption that offline evaluation

could predict an algorithm’s effectiveness in online evaluations or user studies. More

precisely, several researchers have shown that results from offline evaluations do not

necessarily correlate with results from user studies or online evaluations [93], [270],

[271], [279–282]. This means that approaches that are effective in offline evaluations are

not necessarily effective in real-world recommender systems. McNee et al. observed that

"the research community’s dependence on offline experiments [has]

created a disconnect between algorithms that score well on accuracy

metrics and algorithms that users will find useful." [87]

Other researchers also voiced criticism of offline evaluations. Jannach et al. stated that

"the results of oﬄine [evaluations] may remain inconclusive or even misleading" and

"real-world evaluations and, to some extent, lab studies represent probably the best

methods to evaluate systems" [283]. Knijnenburg et al. reported that "the presumed link

between algorithm accuracy […] and user experience […] is all but evident" [284]. Said

et al. consider "on-line evaluation [as] the only technique able to measure the true user

satisfaction" [269]. Rashid et al. observed that biases in the offline datasets may cause

bias in the evaluation [278]. The main reason for the criticism in the literature is that

offline evaluations focus on accuracy yet ignore human factors; however, human factors

strongly affect overall user satisfaction for recommendations. Despite the criticism,

offline evaluations are the predominant evaluation method in the recommender

community [285] and "surprisingly few studies [evaluate] algorithms in live experiments

with real users" [284].

Our review indicates that the voiced criticism of offline evaluations also applies to the

field of research-paper recommender systems. Some of the approaches were evaluated

using both an offline evaluation and a user study. In two evaluations, results from the

offline evaluations were indeed similar to results of the user studies [26], [84]. However,

剩余69页未读，继续阅读

GanD.GanD

粉丝: 3
资源: 90

推荐系统文献综述：深度探索与分析

Recommender Systems.pdf

recommender systems.pdf

知识图谱英语参考文献

给我推荐20个比较流行的推荐算法模型

给我推荐20个比较流行的用户画像算法模型源码地址

引用文献介绍基于内容的推荐算法

a survey on knowledge graph-based recommender systems

graph neural networks in recommender systems: a survey

找git上一个使用了django写的简单的协同过滤推荐系统的接口，提供下载地址

DDPG结合GCN做预测的GitHub代码

最新资源