Reciprocal Recommender Systems: Analysis of state-of-art literature, challenges and opportunities towards social recommendation

Highlights

• Formal characterization of Reciprocal Recommender Systems (RRS).

• Algorithmic, fusion and evaluation processes for reciprocal recommendation.

• Snapshot of RRS literature and their applications, analyzing representative models.

• Discussion of challenges and opportunities for future research on RRS.

Abstract

There exist situations of decision-making under information overload in the Internet, where people have an overwhelming number of available options to choose from, e.g. products to buy in an e-commerce site, or restaurants to visit in a large city. Recommender systems arose as a data-driven personalized decision support tool to assist users in these situations: they are able to process user-related data, filtering and recommending items based on the user’s preferences, needs and/or behavior. Unlike most conventional recommender approaches where items are inanimate entities recommended to the users and success is solely determined upon the end user’s reaction to the recommendation(s) received, in a Reciprocal Recommender System (RRS) users become the item being recommended to other users. Hence, both the end user and the user being recommended should accept the “matching” recommendation to yield a successful RRS performance. The operation of an RRS entails not only predicting accurate preference estimates upon user interaction data as classical recommenders do, but also calculating mutual compatibility between (pairs of) users, typically by applying fusion processes on unilateral user-to-user preference information. This paper presents a snapshot-style analysis of the extant literature that summarizes the state-of-the-art RRS research to date, focusing on the algorithms, fusion processes and fundamental characteristics of RRS, both inherited from conventional user-to-item recommendation models and those inherent to this emerging family of approaches. Representative RRS models are likewise highlighted. Following this, we discuss the challenges and opportunities for future research on RRSs, with special focus on (i) fusion strategies to account for reciprocity and (ii) emerging application domains related to social recommendation.

Introduction

In the last two decades, Recommender Systems (RS) have gained a lot of popularity as an effective information processing and personalized decision support tool to filter relevant information or resources to users in Internet platforms. They are used to assist people in decision-making situations under contexts of “information overload”, which are common not only in e-commerce sites, but also in entertainment portals like Spotify and Netflix, social media sites, tourism portals and any other services where optimizing the user’s experience in terms of their interaction with the system, becomes imperative [1], [2], [3], [4]. In essence, an RS gathers and analyzes users’ interaction data with items in the system to build knowledge about their preferences, whereby the system predicts items (e.g. products, services, media, things to see or do, etc.) that the user is likely to be interested in. Most RS learn users’ preferences from user–item ratings to undertake this predictive task.

The repertoire of RS approaches has amply expanded since the late 90s, with content-based –“recommending similar content to what the user liked” [5]–, collaborative filtering–“recommending what similar people to the user liked” [6]–and context-aware–“recommending items that suit the user’s current context” [7]–being popular approaches [8]. Likewise, there are RS models that provide group recommendations by aggregating members’ preferences or individually recommended items [9], as well as RS that integrate user information across multiple domains to build more insight about their taste or needs [10]. Besides, similar to multi-criteria decision-making scenarios [11] where humans naturally tend to judge options according to multiple criteria, in a multi-criteria RS a user may often prefer to rate items (e.g. hotels) using several criteria (e.g. food, cleanliness, service), hence these systems were proposed to exploit such ratings [12].

An emerging family of RS approaches arose in the last decade, Reciprocal RS (RRS) [13], [14], in which: (i) users become the item being recommended to other users, and (ii) success is determined not only by the end user who requested recommendations, but also by the user(s) being recommended, hence mutual preference or compatibility (reciprocity) needs to be measured. In other words, a vital requirement in any RRS is that both users should reciprocate, i.e. both of them should indicate positive feedback on the suggestion to connect with each other in order to deem the matching recommendation as successful. This requisite adds an additional layer of complexity in RRS with respect to the majority of conventional RS, the latter of which typically only seek satisfying the end user preferences. RRS are popular in dating [15], [16], recruitment [17], online learning environments [18], and social media platforms [19], [20], [21], [22] where reciprocity can yield better matchings between people [13].

Research into classical item-to-user RS has expanded to a sheer range of techniques, algorithms and application areas [1], [2], [8], [12], [23], [24]. Notwithstanding, although several of the most widespread families of RS algorithms have been translated into a reciprocal setting, the problem of personalized people-to-people recommendation via RRS is still comparatively less represented in the literature, hence it still largely poses a number of challenges and unanswered research questions that deserve further exploration. The most obvious complexity found relates to the reciprocity requirement. In this sense, fusion processes are a crucial and distinctive task in most RRS for determining the level of reciprocity or mutual compatibility between two users, predicated on unilateral preferences or recommendation information. Various approaches have been adopted in the RRS literature for fusing unidirectional information between (pairs of) users into reciprocal information [16], [25], [26], with the most salient ones relying on aggregation functions–e.g. harmonic or weighted means between unidirectional user-to-user preferences–for this end [27]. Other well-known examples of ongoing challenges in the research area include: balancing users’ different levels of popularity to prevent biased recommendations [25], an accentuated presence of the data sparsity and user cold-start problems [28] and the relative shortage of publicly available datasets to incentivize user studies, owing to privacy concerns. On another note, even though most RRS along the last decade have been developed for a limited range of applications, namely online dating, recruitment, learning and social media, there exist other social matching problems where they could play a very important role and exert significant impact to promote sociability and, ultimately, to contribute to social good. This opportunity goes in parallel with the increasing abundance of new social media platforms and apps built with more specific goals, e.g. house-share, skill-sharing, professional collaboration, and so on [21], [29], [30].

The aforesaid challenges and identifiable opportunities for future research on RRS, along with the gentle developments witnessed in the RRS literature in the last years and the growing spectrum of consolidated (or potential) real-world applications of RRS, motivate the need for a comprehensive analysis of RRS literature and an elaborated discussion on the current state of affairs in the field. There are recent state-of-the-art surveys in the literature that comprehensively analyze emerging – and comparatively more complex – families of RS along the last years, including: cross-domain RS [31], deep learning based RS [32], [33], sequence-aware RS [34], context-aware RS [35], RS leveraging multimedia contents [36] and adversarial RS [37], to name a few. Nonetheless, although Pizzato et al. [15] sat the bases for subsequent RRS developments in the domain of online dating, to our knowledge there are no exhaustive literature analyses focused on state-of-the-art RRS developments and their growing range of applications ever since.

Accordingly, this paper provides a fourfold contribution — for the readership interested in specific parts of this paper, Fig. 1 provides a mind map-like outline of its contents:

• A formal characterization of RRS with respect to other RS families and the definition of a general RRS conceptual model for RRS guided by preference fusion processes (Section 2).

• An outline of the algorithmic, fusion and evaluation aspects underlying RRS (Section 3).

• An exhaustive analysis of existing state-of-the-art studies in the RRS literature, with a triple objective: (i) signaling the key characteristics of existing RRS methodologies and studies conducted in several application domains; (ii) highlighting the recommendation techniques and fusion processes utilized to account for reciprocity; and (iii) analyzing some representative models in further detail (Section 4).

• A discussion of the challenges, research gaps, opportunities and future research directions in the landscape of RRS (Section 5), emphasizing the underlying fusion processes to capture reciprocity and emerging application areas of RRS.

Following these contributions, Section 6 summarizes the lessons learnt and concludes the paper.