Characterizing context-aware recommender systems: A systematic literature review
Introduction
With the proliferation of big data & data analytics technologies, recommender systems (RS) are now crucial in seeking customer satisfaction through personalization [1]. RS aim at selecting and proposing the most relevant items, services and offers for their users, by considering their profiles, purchase history, preferences, opinions, interactions with offered products and services, as well as their relationships with other clients. At the same time, the generalization of smart-phones and ubiquitous computing has given RS access to context information [2]. Context-aware recommender systems (CARS) go one step further from traditional RS by exploiting context information such as time, location, and user activity to understand user situations and their influence on user preferences. The incorporation of context information into RS [2], [3] leverages the value of these systems by improving the relevance of possible recommendations with respect to changing user needs [4], [5].
The value of context information to improve the quality of recommendations has been demonstrated and supported by different researchers [6], [7], [8], [9], [10], [11]. Nevertheless, RS as well as context-awareness researchers and practitioners interested in combining the two areas still lack a guide that helps them understand how to exploit context information to smarten up RS. Evidence of this is the absence of comprehensive and domain-independent surveys, particularly systematic literature reviews, that not only consolidate the state of the art of the field, but also explain the most common techniques used to integrate context into the recommendation process. After a rigorous revision of the state of the art, we found that none of the available surveys comprehensively characterize recommendation processes from the perspective of the exploitation of context information. In the best cases, existing surveys focus only on the identification of used context types, and most of them address the problem from the perspective of a particular domain.
This paper presents the findings of a systematic literature review (SLR) [12] on CARS that we conducted with the goal of helping practitioners and researchers understand how context information can be effectively combined with recommendation mechanisms. To this end, we studied a final set of 87 CARS papers that were classified as content-based, collaborative filtering and hybrid approaches. For each paper, we identified recommendation techniques, means to exploit context information, context types, application domains, validation mechanisms including the used datasets, the improvements obtained through the exploitation of context (when measured quantitatively), and research opportunities. The main results of our study are reported in this paper in the form of a framework that characterizes recommendation processes in terms of: i) the recommendation techniques used at every stage of the process, ii) the techniques used to incorporate context, and iii) the stages of the process where context is integrated into the system. This manuscript aims at providing a clear understanding about where context information is usually integrated into the system, what techniques are available to exploit context information depending on the underlying recommendation approach and the phase of the process where context is included, what context types are more frequently exploited in the different application domains, and what validation mechanisms—explained in terms of the used datasets, properties, metrics and evaluation protocols—are generally used to evaluate the proposed approaches. Last, but not least, the paper discusses research opportunities relevant to CARS.
This paper is structured as follows. Section 2 explains foundational concepts on recommender systems and context information. Section 3 visits related work by analysing the contributions of our SLR with respect to other surveys published on CARS. Section 4 explains the methodology we followed to conduct the SLR. Sections 5–8 constitute the contributions of this manuscript: Section 5 presents the findings of our SLR and the characterization framework for CARS; Section 6 reports on the validation methods and datasets identified in the studied approaches; Section 7 presents quantitative data, reported in the studied papers, on the improvements obtained from the exploitation of context information; and Section 8 summarizes and classifies research opportunities. Finally, Section 9 concludes the paper.
Section snippets
Background
This section briefly presents the fundamentals of RS, and context information as an enabler to improve the quality of recommendations.
Related work
We found 15 RS surveys published in relevant venues and journals between 2004 and 2016. However, only 7 out of these 15 surveys, published between 2012 and 2014, relate to the improvement of RS through the incorporation of context information. Aiming at providing a comprehensive understanding of the state of the art of this field, our SLR not only follows a well defined research methodology, but also characterizes CARS along all application domains, context types, and techniques reported in the
Methodological aspects
We conducted this study by following the guidelines proposed by Kitchenham and Charters in [12]. With our long-term research goal in mind—to look for innovative and more effective ways of exploiting context information to improve the effectiveness of recommender systems, we defined the set of research questions that would allow us to understand the state of the art of CARS. These questions are stated as follows:
- •
RQ1: How is context information exploited along the recommendation process?
- •
RQ2:
Characterization of Context-Aware RS (CARS)
This section summarizes, for each type of recommender system, the findings of our SLR. We consider that the differences between content-based, collaborative filtering, and hybrid recommenders are too profound to analyze them all together, thus we set to do it independently.
To characterize content-based and collaborative filtering CARS, we first represented their recommendation processes using flow diagrams (cf. Figs. 1 and 2) that allow us to distinguish the different phases they comprise, and
Characterization of validation methods
The improvement of user experience is the ultimate goal of a recommender system. In order to measure it, a series of properties, each with a set of metrics, have been proposed and used since the first developments in the field. These properties allow us to determine the pertinence of the recommendations being suggested. Instances of these properties are predictive power, confidence, diversity, learning rate, coverage, scalability and user evaluation [111].
In this section we summarize the
The effect of incorporating context into RS
When conducting an SRL on CARS, a natural question is the level of improvement of RS performance (e.g., in terms of accuracy) obtained from the inclusion of a particular context type into the recommendation process. Nevertheless, answering this question results impractical, given the wide spectrum of recommendation techniques that can be combined with the different context types, through any of the three existing paradigms to include context information into RS. Furthermore, the performance of
Research opportunities
This section provides CARS researchers with a list of research opportunities, most of them borrowed from the studied articles. From each paper, we identified, categorized, and analyzed the challenges that authors defined as worthy of future work. Each subsection corresponds to one of the nine challenge categories that we identified: dynamic context management, context gathering, context reasoning, contextual modeling, problems inherent in RS, CARS evaluation, users in the loop, self-adaptation
Conclusions
This paper presented a comprehensive characterization of context-aware recommendation processes and systems, based on the findings of a systematic literature review (SLR) we conducted to survey CARS that were published between 2004 and 2016. This study was conducted with the goal of helping practitioners and researchers understand how context information can be effectively combined with recommendation mechanisms. The main results provide a clear understanding about where context information is
Acknowledgment
This work was funded by Universidad Icesi through its institutional research support program.
References (114)
- et al.
Recommender system application developments: a survey
Decis. Support Syst.
(2015) Newsweeder: learning to filter netnews
Proc. 12th Int. Conf. on Machine Learning
(1995)- et al.
Contextual music information retrieval and recommendation: State of the art and challenges
Comput. Sci. Rev.
(2012) - et al.
A systematic review of scholar context-aware recommender systems
Expert Syst. Appl.
(2015) - et al.
Utilizing contextual ontological user profiles for personalized recommendations
Expert Syst. Appl.
(2014) - et al.
Recommetz: a context-aware knowledge-based mobile recommender system for movie showtimes
Expert Syst. Appl.
(2015) - et al.
Making smartphone service recommendations by predicting users’ intentions: a context-aware approach
Inf. Sci.
(2014) - et al.
Context-aware system for proactive personalized service based on context history
Expert Syst. Appl.
(2009) - et al.
A novel mobile recommender system for indoor shopping
Expert Syst. Appl.
(2012) - et al.
Building and evaluating a location-based service recommendation system with a preference adjustment mechanism
Expert Syst. Appl.
(2009)