What and who with: A social approach to double-sided recommendation

https://doi.org/10.1016/j.ijhcs.2017.01.001Get rights and content

Highlights

  • DSR are suggestions made of an item and a group with whom that item can be consumed.

  • We propose a DSR algorithm using information from the target user's social network.

  • The DSR algorithm performs better than a user-based CF one in suggesting items.

  • In the event domain our DSR algorithm outperforms a traditional, content-based one.

  • Users seem to focus on the suggested group when assessing DSR in the event domain.

Abstract

Double-sided recommendations (DSR) have been recently introduced for an item and a group that the item is destined for. Herein we present an algorithm which takes inspiration from the Social Comparison Theory to recommend items that had an average positive evaluation from other users on the target user's social network. Other users' judgments are weighted according to the influence these users have on the target. Moreover, for each recommended item, we propose a group that encompasses all the target users' contacts who expressed a positive opinion on it.

Our data show that users consider double-sided recommendations more useful than traditional recommendations which provide equivalent information. It was observed that our “social” DSR algorithm performs better in the event recommendation domain than a content-based one which has already been recognised as providing a good performance, in terms of precision, recall, accuracy and F1. This result is strengthened by our demonstrating that the good performance DSRs provide also depends on their peculiar structure and not only on the fact that they include “social” information. The item-recommendation part also performed better than a user-based collaborative filtering algorithm. Lastly, we found that users' scores for recommended item-group packages can be better predicted by considering only the system scores for the recommended groups, at least in the domain of social and cultural events.

Introduction

Everybody makes daily decisions, be they small or large, as to what to do and consume. During this decision making process, the presence of other people is usually implicitly or explicitly considered as part of the experience they are planning. For example when a choice has to be made as to which restaurant to eat at, cinema to go to or what film to see, planning holidays or just a weekend, or even deciding what to cook for dinner or collective buying practices. Even if all these activities are part of our everyday life, organizing them properly may turn into a complex task, since reasoning on different levels is involved, like which items to choose for consumption, who to involve and what items to consume with which people. Indeed, although most of these activities may usually be carried out with a predefined group of people (e.g., a person may normally dine in the family), it is not always so: let's take the example of organizing a dinner party where candidate guests may have different food preferences or vary on their ideas of a social event (e.g., intimate vs. large receptions) and be on different terms with each other.

Recommender systems emerged as a way to help users make choices (Jameson et al., 2014) by generating personalized suggestions, especially when there is too much information for a human decision maker to deal with effectively (the so-called problem of “information overload”). Recommenders usually follow either a content-based or a collaborative filtering approach. Content-based recommenders exploit some representation of users' interests and preferences and suggest items that appear to match it well, based on the idea that users will stick to their preferences and will continue to like items similar to those they liked in the past. Differently, collaborative filtering recommenders take advantage of rating data expressed by large numbers of users and suggest items that have been rated positively by users whose rating behaviour is similar to that of the target. Therefore, the ratio will be that users who agreed on their judgement for some items in the past will most likely have the same opinions on other items too. Most recommender systems suggest single items to single target users (Adomavicius and Tuzhilin, 2005). More advanced recommenders can either provide recommendations to groups (Jameson and Smyth, 2007) or suggest complex items, for example sequences or packages of items (Chao et al., 2005), friends (Hsu et al., 2006) or pre-existing groups of people (Baatarjav et al., 2008, Carmagnola et al., 2009). A previous study of ours (Vernero, 2011) proposed the concept of double-sided recommendations (DSRs), envisaging the need for an advanced recommendation technique which generates suggestions made up of an item and the group that the item should be consumed with. This matched situations with a strong social connotation, where the suggested items are at least as likely to be used by groups as they are by single individuals (as is the case of group recommendations), but where a group cannot be determined a priori.

Herein we present an algorithm able to generate DSRs (hereafter referred to as DSR algorithm) according to the so-called Social Comparison-based Recommendation Method, that we first introduced, in terms of high-level concepts and building blocks, in Vernero (2011). This method was chosen as it mainly uses information drawn from the target user's social network. In fact, relevant literature (Sinha and Swearingen, 2001), as well as our own previous research (Carmagnola et al., 2009), have evidenced how people tend to place a high value on recommendations based on the opinions and preferences of friends and other trusted users.

With the aim of validating our algorithm, we embraced a piece-wise approach and carried out a set of empirical evaluations able to make an assessment from various points of view and in different domains. Our goals were:

  • to evaluate the usefulness of DSRs generated by our “social” DSR algorithm and to compare them with other, more traditional, kinds of recommendations.

  • to understand whether the item-recommendation part of our DSR algorithm has comparable performance, in terms of standard measures such as precision and recall, to a collaborative filtering algorithm, in the context of a large offline study.

  • to understand whether, in the context of an online user study, our DSR algorithm performs better than a traditional content-based one which makes use of detailed information on the target user's preferences and proved to have a good performance with a consolidated user model. The algorithm and its evaluation are described in detail by Carmagnola et al. (2008) and Gena et al. (2013).

  • to assess how much importance users give to the recommended item and group, respectively, when they are evaluating a DSR in a particular domain (we surmised that user preferences might depend on the type of suggestion they receive). This information is useful to understand how these two elements should contribute to the overall predicted score for a recommendation:

  • to verify that the efficacy of DSRs does not depend solely on the fact that they provide information obtained from the target user's social network (to this aim, we compared the full DSR algorithm with its item-recommendation part).

Our empirical evaluations targeted three different domains, i.e. collaborative buying practices (in the context of a fictional system, first goal), music (second goal), and participation in social events (third, fourth and fifth goals) respectively. iCITY, a social adaptive recommender system we developed in the past, was used for the event domain, as a use-case.

Briefly, this paper makes a two-fold contribution: on the one hand, it proposes a new “social” algorithm for the generation of DSRs, starting from the concepts we introduced in a previous study of ours; on the other, it presents the results of the evaluations we carried out so as to assess it.

The paper is organized as follows: a presentation of the background information on DSRs is provided in Section 2. Section 3 reports on our work within pertinent research, whilst Section 4 presents our approach, detailing the algorithms. Section 5 discusses and describes our experiments and their results. Section 6 concludes the paper.

Section snippets

Background

The concept of DSRs was first introduced in Vernero (2011) to extend the scope of recommender systems to situations where the target user may need to be suggested, not only a personally appealing item, but also a group of people that the item could be consumed with. DSRs were formally defined as follows:

Definition 1

Given a set of users U, a target user tU, a set of contacts of the target user NtU and a set of candidate items I, we call a double-sided recommendation (DSR) a pair i,Gt(i) where iI and Gt(

Related work

The paper by Stefanidis and Pitoura (2013) is the most representative of the idea of DSRs proposed herein, which focused on the suggestion of vacation packages and addressed the problem of forming a group of users which is an appropriate target for the recommended packages. The authors proposed a greedy group construction algorithm that took into account a number of constraints as to the users' liking for the item as well as group composition. While DSRs are directly aimed at a single target

Double sided recommendation - a social approach

In the Social Comparison-based method (Vernero, 2011), target users are recommended items that were positively evaluated, on average, by other users in their social network, e.g., contacts. The users' judgment is weighted according to the influence they have on the target user, which is a function of the similarity and relevance of a certain contact for the target. For each recommended item, a group made up of all - and exclusively - the target user's contacts who expressed a positive opinion

Evaluation

A set of four different evaluations were used to assess our approach to DSRs. Firstly, a preliminary test was done, focusing on the goal of supporting users in adopting collaborative buying practices, with the aim of evaluating the usefulness of DSRs. More specifically, the aim was that of answering the following research question:

RQ1: are DSRs generated with the “social” DSR algorithm herein proposed more useful than traditional recommendations that have either a group or an item, but provide

Conclusions and future research

In this paper we presented an algorithm based on the Social Comparison Theory which generates double-sided recommendations starting from the preferences extrapolated from the social network of the target user. Our empirical evaluations assessed this algorithm from various points of view (usefulness, performance, parameter tuning) and in different domains, trying to answer five research questions (RQ1-5, see Section 5).

Firstly, our data showed that our double-sided recommendations were

References (51)

  • C. Cornelis et al.

    One-and-only item recommendation with fuzzy logic techniques

    Inf. Sci.

    (2007)
  • J.K. Kim et al.

    A group recommendation system for online communities

    Int. J. Inf. Manag.

    (2010)
  • C. Martínez-Cruz et al.

    A model to represent users trust in recommender systems using ontologies and fuzzy linguistic modeling

    Inf. Sci.

    (2015)
  • G. Adomavicius et al.

    Toward the next generation of recommender systems: a survey of the state-of-the-art and possible extensions

    IEEE Trans. Knowl. Data Eng.

    (2005)
  • S. Amer-Yahia et al.

    Group recommendation: semantics and efficiency

    PVLDB

    (2009)
  • O. Arazy et al.

    A theory-driven design framework for social recommender systems

    J. Assoc. Inf. Syst. AIS

    (2010)
  • L. Ardissono et al.

    Intrigue: personalized recommendation of tourist attractions for desktop and hand held devices

    Appl. Artif. Intell.

    (2003)
  • Baatarjav, E.-A., Phithakkitnukoon, S., Dantu, R., 2008. Group recommendation system for facebook. In: OTM '08:...
  • R. Bali et al.

    R Machine Learning By Example Community experience distilled

    (2016)
  • Baltrunas, L., Makcinskas, T., Ricci, F., 2010. Group recommendations with rank aggregation and collaborative...
  • Basu, C., Hirsh, H., Cohen, W., 1998. Recommendation as classification: Using social and content-based information in...
  • J. Beel et al.

    Towards reproducibility in recommender-systems research

    User Model. User-Adapt. Interact.

    (2016)
  • Berkovsky, S., Freyne, J., 2010. Group-based recipe recommendations: analysis of data aggregation strategies. In:...
  • Berkovsky, S., Freyne, J., Smith, G., 2012. Personalized network updates: Increasing social interactions and...
  • L. Boratto et al.

    State-of-the-art in group recommendation and new approaches for automatic identification of groups

  • Boratto, L., Carta, S., Chessa, A., Agelli, M., Clemente, M.L., 2009. Group recommendation with automatic...
  • Bourke, S., McCarthy, K., Smyth, B., 2011. Power to the people: Exploring neighbourhood formations in social...
  • Buhl, M., Famulare, J., Glazier, C., Harris, J., McDowell, A., Waldrip, G., Barnes, L.E., Gerber, M.S., April 2016....
  • Cantador, I., Brusilovsky, P., Kuflik, T., 2011. In: Proceedings of the 2nd workshop on information heterogeneity and...
  • F. Carmagnola et al.

    Sonars: a social networks-based algorithm for social recommender systems

  • F. Carmagnola et al.

    Advanced social recommendations with sonars++

    Interact. Comput.

    (2014)
  • F. Carmagnola et al.

    Tag-based user modeling for social multi-device adaptive guides

    User Model. Use.-Adapt. Interact.

    (2008)
  • F. Cena et al.

    Should i stay or should i go? Improving event recommendation in the social web

    Interact. Comput.

    (2016)
  • Chao, D.L., Balthrop, J., Forrest, S., 2005. Adaptive radio: achieving consensus using negative preferences. In: GROUP...
  • Chen, J.H., Chao, K.M., Shah, N., 2013. Hybrid recommendation system for tourism. In: 2013 IEEE Proceedings of the 10th...
  • Cited by (5)

    • Logical foundations of knowledge-based recommender systems: A unifying spectrum of alternatives

      2021, Information Sciences
      Citation Excerpt :

      With package recommendations, each suggestion consists in a set of items which are expected to be consumed “together” [3]. Examples of packages are playlists of songs or movies, travel plans or sets of points of interest which can be visited as part of a single trip, bundles of products which can be bought together, as well as teams of players or co-workers or a single item, such as a restaurant, accompanied by a group of people with whom that item can be enjoyed [30]. Sequence recommendations (see, e.g., [42]) extend the package concept by including temporal constraints on item consumption.

    • Multi-Sided recommendation based on social tensor factorization

      2018, Information Sciences
      Citation Excerpt :

      Lastly, the proposed method learns the range of the social network based on the user feedback. Recently, the double-sided recommendation (DSR) has been introduced to provide both items and users for the improvement of the recommendation quality [19,36]. The method name Multi-sided Recommendation based on Social Tensor (MSRST) was derived since this approach recommends multiple factors as follows: a movie as the item, two users (i.e., an active and another user for the watching of the recommended item together), time as the temporal context, and place as the spatial context.

    • Individual and Group Decision Making and Recommender Systems

      2022, Recommender Systems Handbook: Third Edition
    View full text