Lots of activities, like watching a movie or going to the restaurant, are intrinsically group-based. To recommend such activities to groups, traditional single-user recommendation techniques are not appropriate and, as a consequence, over the years a number of group recommender systems have been developed. Recommending items to be enjoyed together by a group of people poses many ethical challenges: in fact, a system whose unique objective is to achieve the best recommendation accuracy might learn to disadvantage submissive users in favor of more aggressive ones. In this work we investigate the ethical challenges of context-aware group recommendations, in the general case of ephemeral groups (i.e., groups where the members might be together for the first time), using a method that can recommend also items that are new to the system. We show the goodness of our method on two real-world datasets. The first one is a very large dataset containing the personal and group choices regarding TV programs of 7,921 users w.r.t. sixteen contexts of viewing, while the second one gathers the musical preferences (both individual and in groups) of 280 real users w.r.t. two contexts of listening. Our extensive experiments show that our method always manages to obtain the highest recall while delivering ethical guarantees in line with the other fair group recommender systems tested.
Recommender Systems are software tools and techniques that provide suggestions for items to
be of use to a user [
The majority of the existing approaches to Recommender Systems do not take into
consideration any contextual information, however, in many applications, the context of use might be
fundamental in guiding the current preference of a user [
Group Recommender Systems are systems that produce a common recommendation for a group
of users [
In single-user Recommender Systems, fairness is usually assessed with regard to sensitive attributes which are generally prone to discrimination (e.g., gender, ethnicity or social belonging) by verifying the presence of a discriminated class within the user set [16, 17]. When fairness is evaluated considering Group Recommender Systems, it should be computed within groups. Since the groups we consider in this work are composed of few users, evaluating fairness in the way just described is not a suitable solution. Instead of detecting unfairness towards a protected group of users, we aim to detect and prevent unfairness towards single users within a group whose desires are not considered when forming a recommendation for the whole group.
Some aggregation strategies exist that, despite not having been developed to explicitly address
ethical issues, aggregate individual preferences in a way that resembles fairness. Least misery,
used in [
In this section we review a previous approach of ours, introduced in [
We considere a set of items and a set of users , from which any group ∈ ℘( ) can be extracted. is the set of possible contexts in the given scenario, where a context is the conjunction of a set of dimension/value pairs: e.g., for the TV dataset, a context might be = ⟨_ = ∧ = ⟩. We assume the availability of a log ℒ recording the history of the items previously chosen by groups formed in the past, where each element of ℒ is a 4-ple ( , , , ), being the time instant in which the item ∈ has been chosen by the group ∈ ℘( ) in the context ∈ . A contextual scoring function (, , ), with ∈ , ∈ , ∈ , assigning to each user the score given to the items in the various contexts, is computed ofline on the basis of the log of the past individual choices and of the item descriptions in terms of their attributes, using any context-aware recommender system for single users from the literature. (, , ) is the function that returns the list of the items preferred by user in context , according to the values of (, , ), for each ∈ available at instant . Given a target group ∈ ℘( ), a context ∈ and a time instant , the group recommendation is obtained by recommending to the users in a list (i.e., an ordered set) of items, considered interesting in context , from those items in that are available at time instant , according to the following procedure:
The group preference for an item is obtained by aggregating the individual preferences of the group members on the basis of their influence. In each context , the influence (, ) of a given user is derived ofline by comparing the behavior of when alone (i.e., ’s individual preferences) with ’s behaviors in groups (i.e., the interactions contained in the log ℒ). Basically, the influence of tells us how many times the groups containing have selected one of ’s favorite items. Let (, , ) be the list of the items preferred by user in context , according to the values of (, , ) for each ∈ available at instant . The contextual influence is defined as follows: (, ) = | ∈ ℒ : = ∧ ∈ ∧ ∈ (, , )|
| ∈ ℒ : = ∧ ∈ | The value of (, ) quantifies the ability of user to direct the group’s decision towards ’s own tastes while in context .
Top- recommendations are computed online, when a group of users requires that the system
suggests some interesting items to be enjoyed together. The system must compute the group
preferences for the items, and then determine the items with the highest scores. Given a
group ∈ ℘( ), its preference (, , ) for ∈ in the context ∈ is computed as
the average of the preferences of its members weighed on the basis of each member’s influence
(Eq. 1) in context :
(, , ) =
∑︀∈ (, ) · (, , )
∑︀∈ (, )
(1)
(2)
Then, the top- list of items preferred by a certain group in context at time instant is
determined by retrieving the items with the highest scores among those available at time .
3.2. FARGO
Being CtxInfl based on the concept of influence, it inevitably privileges the preferences of the
most influential users. As a consequence, the results of the recommendation process are biased
towards the preference of one user or few users of the group who can be considered as the
leaders, or, using a more contemporary word, “influencers". Our aim is to add an element of
fairness to CtxInfl while maintaining its general structure, which already proved to be very
eficient and scalable [
In this section we present the results obtained by applying the proposed approach to two diferent real-world datasets. To evaluate the recommendation performance we use recall, considering for (number of items to be recommended) the values 1, 2 and 3. To evaluate the ethical properties of our method we used the two metrics proposed in [28] for estimating user discrimination, called score disparity and recommendation disparity, adapted to our needs. Score disparity is computed as the Gini coeficient of user satisfaction, i.e., the relative gain achieved by the user due to the actual recommendation with respect to the optimal recommendation strategy from the user perspective. Recommendation disparity is computed as the Gini coeficient of user gains, i.e., how many of the recommended items match the user Top-K items.
We compare our approach to the following methods: average (AVG) [8, 23, 18, 19, 9, 10, 13,
11, 12, 21], Fair Lin [24], Fair Prop [25, 26], Envy Free [25, 26], minimum disagreement (Dis)
[22], least misery (LM) [
This dataset contains TV viewing information related to 7,921 users and 119 channels, broadcasted both over the air and by satellite. The dataset is composed of an Electronic Program Guide (EPG) containing the description of 21,194 distinct programs, and a log containing both individual and group viewings performed by the users. The log spans from December 2012 to
Recall 37.94% 33.914% 33.22% 32.99% 29.33% 33.57% 30.35% 30.47% February 2013 and contains 4,968,231 entries, among which we retained just the syntonizations longer than three minutes. 3,519,167 viewings were performed by individual users, and are used to compute the individual preferences of the group members. The remaining 1,449,064 viewings have been done by more than one person. The two context dimensions considered are day of the week (weekday vs. weekend) and the time slot. The available values for the time slot are: graveyard slot, early morning, morning, daytime, early fringe, prime access, primetime, and late fringe. Group viewings are split into a training set (1,210,316 entries), and a test set (238,748 entries) with a 80%-20% ratio. Results are reported in Table 1. Note that the superiority of our method, recall-wise, is very pronounced. For what regards the ethical guarantees, FARGO, delivers a very good score disparity, while, for what regards the recommendation disparity, it seems to perform generally worse than the other methods (except for = 1, for which its performance is on par with the other methods). Note that for GFAR it is not possible to compute the score disparity as it does not involve the computation of group scores for the items.
This dataset1 has been created within the scope of a user study by asking participants to fill in two diferent forms: an individual form collecting demographic data (i.e., age and gender) and contextual individual preferences about music artists, and a group form to be filled in groups asking for a collective choice of a music artist that was available at the time of the choice in a particular context. The following two listening contexts have been selected, considering that both are common situations users can relate to both when alone and when with other people, and that users’ preferences would likely be diferent in each of them: during a car trip and at dinner as background music. The dataset obtained contains data gathered from 280 users. For each user, preferences regarding both the car trip and dinner contexts are gathered. From the group forms, 498 context-aware collective preferences have been gathered. Of this, 272 groups were composed of 2 users, 158 of 3 users, 32 of 4 users and 36 of 5 users. As for the previous dataset, we used a 80%-20% split for training and test sets. Results are reported in Table 2. Also in this case FARGO delivers the best recall. Contrarily to the previous dataset, in this case our method achieves a very good recommendation disparity. For what regards the score disparity, all methods provide very low (i.e., good) values.
1The dataset can be downloaded at https://github.com/azzada/FARGO.
Recall 25.00% 12.50% 11.11% 13.19% 12.50% 22.92% 13.89% 6.06%
In this paper we have introduced FARGO, a new method for providing fair, context-aware recommendations to ephemeral groups able also to recommend items new in the system. Considering both recall and fairness, it is not possible to identify a best overall method across all datasets and values of . Even if we ignored recall, a clear winner fairness-wise is not evident (all methods tested, except for Dis, perform best fairness-wise for at least a value of in at least one dataset). We argue that the relationship between fairness and recommendation accuracy should be seen as a tradeof. On both datasets of our experiments, FARGO provides the best solution to such tradeof by achieving the best recall across all values of while delivering similar ethical guarantees to the other fair methods tested. Contrarily to what one might think, LM is not the best method fairness-wise, and this implies that the problem of maximizing both recall and fairness is not a simple one. This is a complex problem that deserves further investigations, as recall and fairness seem not to be inversely correlated in a trivial manner. [8] J. Masthof, Group modeling: Selecting a sequence of television items to suit a group of viewers, in: Personalized Digital Television, Springer, 2004, pp. 93–141. [9] E. Ntoutsi, K. Stefanidis, K. Nørvåg, H.-P. Kriegel, Fast group recommendations by applying user clustering, in: Proc. ER, 2012, pp. 126–140. [10] A. J. Chaney, M. Gartrell, J. M. Hofman, J. Guiver, N. Koenigstein, P. Kohli, U. Paquet, A large-scale exploration of group viewing patterns, in: Proc. TVX, 2014, pp. 31–38. [11] T. De Pessemier, S. Dooms, L. Martens, Comparison of group recommendation algorithms,
Multimedia Tools Appl. 72 (2014) 2497–2541. [12] N.-r. Kim, J.-H. Lee, Group recommendation system: Focusing on home group user in tv domain, in: Proc. SCIS, 2014, pp. 985–988. [13] I. Ali, S.-W. Kim, Group recommendations: approaches and evaluation, in: Proc. IMCOM, 2015, pp. 1–6. [14] M. Gartrell, X. Xing, Q. Lv, A. Beach, R. Han, S. Mishra, K. Seada, Enhancing group recommendation by incorporating social relationship interactions, in: Proc. GROUP, 2010, pp. 97–106. [15] S. Berkovsky, J. Freyne, Group-based recipe recommendations: analysis of data aggregation strategies, in: Proc. RecSys, 2010, pp. 111–118. [16] S. Yao, B. Huang, New fairness metrics for recommendation that embrace diferences,
CoRR abs/1706.09838 (2017). [17] Y. Li, Y. Ge, Y. Zhang, Tutorial on fairness of machine learning in recommender systems, in: Proc. SIGIR, 2021, pp. 2654–2657. [18] L. Baltrunas, T. Makcinskas, F. Ricci, Group recommendations with rank aggregation and collaborative filtering, in: Proc. RecSys, 2010, pp. 119–126. [19] C. Senot, D. Kostadinov, M. Bouzid, J. Picault, A. Aghasaryan, C. Bernier, Analysis of strategies for building group profiles, in: Proc. UMAP, 2010, pp. 40–51. [20] J. Gorla, N. Lathia, S. Robertson, J. Wang, Probabilistic group recommendation via information matching, in: Proc. WWW, 2013, pp. 495–504. [21] S. Ghazarian, M. A. Nematbakhsh, Enhancing memory-based collaborative filtering for group recommender systems, Expert Syst. Appl. 42 (2015) 3801–3812. [22] S. Amer-Yahia, S. B. Roy, A. Chawlat, G. Das, C. Yu, Group recommendation: Semantics and eficiency, in: Proc. VLDB, 2009, pp. 754–765. [23] Z. Yu, X. Zhou, Y. Hao, J. Gu, Tv program recommendation for multiple viewers based on user profile merging, User Model. User-Adapt. Int. 16 (2006) 63–82. [24] L. Xiao, Z. Min, Z. Yongfeng, G. Zhaoquan, L. Yiqun, M. Shaoping, Fairness-aware group recommendation with pareto-eficiency, in: Proc. RecSys, 2017, pp. 107–115. [25] S. Qi, N. Mamoulis, E. Pitoura, P. Tsaparas, Recommending packages to groups, in: Proc.
ICDM, 2016, pp. 449–458. [26] D. Serbos, S. Qi, N. Mamoulis, E. Pitoura, P. Tsaparas, Fairness in package-to-group recommendations, in: Proc. WWW, 2017, pp. 371–379. [27] M. Kaya, D. Bridge, N. Tintarev, Ensuring fairness in group recommendations by ranksensitive balancing of relevance, in: Proc. RecSys, 2020, pp. 101–110. [28] J. Leonhardt, A. Anand, M. Khosla, User fairness in recommender systems, in: Proc.
WWW, 2018, pp. 101–102.