协作过滤推荐系统在动态环境中的长期性能研究

73 浏览量更新于2024-08-28 收藏 1.48MB PDF 举报

本文主要探讨了在时间演进系统中，基于协同过滤的推荐系统的长期性能。随着推荐系统在日常生活中无处不在的影响，人们对单步推荐和静态用户-项目网络的关注度不断提高，然而，对于推荐系统在动态环境下的长期效果却鲜有深入研究。作者们针对这一问题，首先通过构建一个二分图模型来描绘一个在线商业系统，该模型捕捉了推荐过程中用户和项目之间的实时交互。在文中，提出了一种基于推荐的演化方法，这种方法模拟了随着时间推移，用户兴趣、项目内容以及用户与项目之间的关联如何随时间变化。这种演化策略考虑到了推荐过程中的动态性，旨在提供更为准确和持久的个性化推荐。它不仅关注于短期的推荐精度，而且着重于长期的用户体验和满意度。通过对长期性能的评估，研究人员对比了不同协同过滤算法在时间演进系统中的表现，包括但不限于基于内容的推荐、矩阵分解方法（如SVD和PMF）以及深度学习驱动的推荐技术。他们探究了这些算法如何应对用户行为的不稳定性和新项目的加入，以及如何在保持推荐准确性的同时，适应系统随时间的变化。此外，文章还可能分析了不同数据更新频率对推荐性能的影响，以及如何通过调整推荐策略来优化系统的长期稳定性和用户参与度。实验结果和案例分析展示了协同过滤推荐系统在面对动态环境时的优势和挑战，为业界提供了有价值的指导，特别是在维护推荐系统的长期有效性方面。这篇论文的重要性在于填补了现有研究的空白，为理解和改进推荐系统在现实世界中的长期性能提供了理论依据和实践策略。这对于不断发展的在线服务和个性化推荐领域具有深远的意义，将有助于提升用户的满意度和系统的可持续性。

ACCEPTED MANUSCRIPT

recommendation accuracy on the static dataset, which assumes

that rating matrix is fixed without incremental data. For

real-world online applications and systems, this assumption is

lack of practicability since the mutual feedback between system

and the user decisions are on-going [29]. Thus, the new ratings

happen continuously in recommender systems, and the

real-world sites with recommender system are temporally

evolving systems.

Most importantly, it is one of the most important goals but

not the ultimate goal that making good rating predictions for

unrated items. The original intention of employing

recommender system is that not only mining some useful

information to customer, but also bringing profits to the sites

[30-31]. Hence, the ultimate goal of employing the

recommender system is to broaden the scope of customer

interest, in addition to recommend unrated items with high

predicted rating to user.

In summary, the quality metrics of a healthy recommender

system should include accuracy, novelty, diversity and other

utilities. Furthermore, it is significant to evaluate the quality of

recommender system in a more comprehensive way that the

combination of short- and long-term effects.

Unfortunately, the long-term effects of a CF-based

recommender in the temporally evolving system are rarely

discussed up to date. For example, it is still unclear whether a

well-performed CF-based recommender in a one-step

recommendation can also provide high performance in the

long-term recommendation. Meanwhile, it is not sure that an

accurate-oriented recommender (e.g., an LFM-based one) can

also share the diversity of recommendation and the health of

temporally evolving system. Last but not least, there are several

unknown questions about: whether it will hurt the health of

temporally evolving system by using highly accurate

recommender; and which recommender will lead to this effect;

what extent this effect will be during the temporal evolving

process.

In such context, some works have been proposed to address

the aforementioned issues. In order to deal with customers’

preference for products which are keep drifting over time,

Koren et al. [29] proposed a time-aware factor model that

introducing the temporal dynamics of user rating criteria rating

and item’s popularity. Based on this model, Koenigstein et al.

[17] proposed a rich bias model with consideration of items’

type to improve the recommendation accuracy in Yahoo!

Music dataset.

Since the data of historical user behavior are ever growing in

real-world applications, Luo et al. proposed an incremental

collaborative filtering model to deal with the problem of data

explosion [32]. For the heterogeneity environment, Rosaci et al.

proposed a new agent-based system with consideration of the

effect of the device exploited by the user [33]. Shi et al. [34]

designed a user preference-oriented recommender to deal with

the temporal dynamics in the online system. However, these

works only focus on improving the accuracy of recommender

without considering other important performance indexes, such

as diversity of recommendation and health of system. Different

with above works, we give a comprehensive performance study

of recommender in temporally evolving system.

Moreover, there are several studies along this line. Ekstrand

et al. [35] proposed a novel MovieLens movie recommender

that allows users to choose recommender and studies its

performance in the long-term evolving process. and

Experimental results showed that how customers make use of

this power. In order to evaluate the long-term effects of using

recommender system, Zeng et al. [36] proposed an evolution

model to simulate the interaction between recommender system

and its users, but it not considered the temporal dynamics. Hu et

al. [37] investigated the recommendation accuracy of different

recommenders in evolving networks without considering the

temporal dynamics. Zhao et al. [38] reported the performance

of different recommender on the evolution of user-item

bipartite network. However, the paper focuses mainly on

studying the performance of entity relationships-based

recommender instead of involving numerical

optimization-based ones (e.g., LFM). In addition, the

recommendation with time is iterated for a small number of

rounds, making the long-term effects of co-evolution between

recommender and users decisions hard to detect.

This paper focuses on investigating the long-term

performance of CF-based recommenders in temporally

evolving systems. To do so, we firstly model a recommender

system with an evolving user-item bipartite network, and then

propose a recommendation-based evolution method to simulate

the temporal dynamics between recommender and the

decisions of its users. Meanwhile, the network is evolving

driven by recommender. In such settings, we study the

performance of three well-known CF-based recommenders,

i.e., U-NBM, I-NBM and LFM. For thoroughly evaluating the

performance of involved models, we have adopted four

performance metrics, i.e., Gini coefficient for ecosystem health,

intra-similarity and popularity for recommendation diversity,

and the root mean squared error (MASE) for recommender

accuracy.

To summarize, the main motivation of this work is to explore

the long-term performance of CF-based recommenders in the

temporally evolving system, in order to make them more

practical in real-world application. The main contributions

include:

--We will propose a recommendation-based evolution method

to capture the temporal dynamics between recommender

system and users’ behaviors in the temporally evolving system.

--We will investigate the ecosystem health and diversity of

recommendation as well as recommender accuracy. Through

empirical study, we find that numerical optimization-based

LFM has the lowest accuracy loss during the temporal evolving

process, while entity relationships-based ICF shows a relatively

better performance than LFM in terms of recommendation

diversity and ecosystem health.

--We will conduct experiment on two real datasets for studying

the performance of CF-based recommenders with the proposed

method.

The rest of this paper is organized as follows: Section II

presents the preliminaries. Section III develops the proposed

method. Section IV illustrates the experiments and discusses

剩余11页未读，继续阅读

weixin_38728360

粉丝: 4
资源: 926

协作过滤推荐系统在动态环境中的长期性能研究

Empirical Analysis of Collaborative Filtering-based Recommenders in Temporally Evolving Systems

Deep Collaborative Filtering via Marginalized Denoising

A User-based Cross Domain Collaborative Filtering Algorithm Based on a Linear Decomposition Model

2018-Variational Autoencoders for Collaborative Filtering.pdf

人工智能-协同过滤推荐算法-基于用户+基于项目-python实现-CollaborativeFiltering-master

Model-Based Engineering of Collaborative Embedded Systems

Movie-Recommendation-System-Using-Collaborative-Filtering

Improve the algorithmic performance of collaborative filtering by using the interevent time distribution of human behaviors

User-based-Collaborative-Filtering:Python中基于用户的协作过滤

Memory-based-collaborative-filtering:高效的协同过滤算法实现-推荐系统

最新资源