GroupIM: A Mutual Information Maximization Framework for Neural Group Recommendation

We study the problem of making item recommendations to ephemeral groups, which comprise users with limited or no historical activities together. Existing studies target persistent groups with substantial activity history, while ephemeral groups lack historical interactions. To overcome group interaction sparsity, we propose data-driven regularization strategies to exploit both the preference covariance amongst users who are in the same group, as well as the contextual relevance of users' individual preferences to each group. We make two contributions. First, we present a recommender architecture-agnostic framework GroupIM that can integrate arbitrary neural preference encoders and aggregators for ephemeral group recommendation. Second, we regularize the user-group latent space to overcome group interaction sparsity by: maximizing mutual information between representations of groups and group members; and dynamically prioritizing the preferences of highly informative members through contextual preference weighting. Our experimental results on several real-world datasets indicate significant performance improvements (31-62 recommendation techniques.



There are no comments yet.


page 9


Overcoming Data Sparsity in Group Recommendation

It has been an important task for recommender systems to suggest satisfy...

Hierarchical Hyperedge Embedding-based Representation Learning for Group Recommendation

In this work, we study group recommendation in a particular scenario, na...

Double-Scale Self-Supervised Hypergraph Learning for Group Recommendation

With the prevalence of social media, there has recently been a prolifera...

Graph Neural Netwrok with Interaction Pattern for Group Recommendation

With the development of social platforms, people are more and more incli...

Group Recommendation Techniques for Feature Modeling and Configuration

In large-scale feature models, feature modeling and configuration proces...

Group Recommendations: Axioms, Impossibilities, and Random Walks

We introduce an axiomatic approach to group recommendations, in line of ...

Social Influence-based Attentive Mavens Mining and Aggregative Representation Learning for Group Recommendation

Frequent group activities of human beings have become an indispensable p...
This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1. Introduction

We address the problem of recommending items to ephemeral groups, which comprise users who purchase very few (or no) items together (Quintarelli et al., 2016). The problem is ubiquitous, and appears in a variety of familiar contexts, e.g., dining with strangers, watching movies with new friends, and attending social events. We illustrate key challenges with an example: Alice (who loves Mexican food) is taking a visitor Bob (who loves Italian food) to lunch along with her colleagues, where will they go to lunch? There are three things to note here: first, the group is ephemeral, since there is no historical interaction observed for this group. Second, individual preferences may depend on other group members. In this case, the group may go to a fine-dining Italian restaurant. However, when Alice is with other friends, they may go to Mexican restaurants. Third, groups comprise users with diverse individual preferences, and thus the group recommender needs to be cognizant of individual preferences.

Prior work primarily target persistent groups which refer to fixed, stable groups where members have interacted with numerous items as a group (e.g.

, families watching movies). They mainly fall into two categories: heuristic pre-defined aggregation (

e.g., least misery (Baltrunas et al., 2010)) that disregards group interactions; data-driven strategies such as probabilistic models (Yuan et al., 2014; Rakesh et al., 2016) and neural preference aggregators (Cao et al., 2018; Vinh Tran et al., 2019). A key weakness is that these methods either ignore individual user activities (Vinh Tran et al., 2019; Wang et al., 2019) or assume that users have the same likelihood to follow individual and collective preferences, across different groups (Yuan et al., 2014; Rakesh et al., 2016; Cao et al., 2018). Lack of expressivity to distinguish the role of individual preferences across groups results in degenerate solutions for sparse ephemeral groups. A few methods exploit external side information in the form of a social network (Yin et al., 2019; Cao et al., 2019), user personality traits and demographics (Delic et al., 2018b), for group decision making. However, side information may often be unavailable.

We train robust ephemeral group recommenders without resorting to any extra side information. Two observations help: first, while groups are ephemeral, group members may have rich individual interaction histories; this can alleviate group interaction sparsity. Second, since groups are ephemeral with sparse training interactions, base group recommenders need reliable guidance to learn informative (non-degenerate) group representations, but the guidance needs to be data-driven, rather than a heuristic.

To overcome group interaction sparsity, our key technical insight is to regularize the latent space of user and group representations in a manner that exploits the preference covariance amongst individuals who are in the same group, as well as to incorporate the contextual relevance of users’ personal preferences to each group.

Thus, we propose two data-driven regularization strategies. First, we contrastively regularize the user-group latent space to capture social user associations and distinctions across groups. We achieve this by maximizing mutual information (MI) between representations of groups and group members, which encourages group representations to encode shared group member preferences while regularizing user representations to capture their social associations. Second, we contextually identify informative group members and regularize the corresponding group representation to reflect their personal preferences. We introduce a novel regularization objective that contextually weights users’ personal preferences in each group, in proportion to their user-group MI. Group-adaptive preference weighting precludes degenerate solutions that arise during static regularization over ephemeral groups with sparse activities. We summarize our key contributions below:

  • [leftmargin=*]

  • Architecture-agnostic Framework: To the best of our knowledge, Group Information Maximization (GroupIM) is the first recommender architecture-agnostic framework for group recommendation. Unlike prior work (Vinh Tran et al., 2019; Cao et al., 2018) that design customized preference aggregators, GroupIM can integrate arbitrary neural preference encoders and aggregators. We show state-of-the-art results with simple efficient aggregators (such as meanpool) that are contrastively regularized within our framework. The effectiveness of meanpool signifies substantially reduced inference costs without loss in model expressivity. Thus, GroupIM facilitates straightforward enhancements to base neural recommenders.

  • Group-adaptive Preference Prioritization

    : We learn robust estimates of group-specific member relevance. In contrast, prior work incorporate personal preferences through static regularization 

    (Yuan et al., 2014; Rakesh et al., 2016; Cao et al., 2018). We use Mutual Information to dynamically learn user and group representations that capture preference covariance across individuals in the same group; and prioritize the preferences of highly relevant members through group-adaptive preference weighting; thus effectively overcoming group interaction sparsity in ephemeral groups. An ablation study confirms the superiority of our MI based regularizers over static alternatives.

  • Robust Experimental Results: Our experimental results indicate significant performance gains for GroupIM over state-of-the-art group recommenders on four publicly available datasets (relative gains of 31-62% NDCG@20 and 3-28% Recall@20). Significantly, GroupIM achieves stronger gains for: groups of larger sizes; and groups with diverse member preferences.

We organize the rest of the paper as follows. In Section 3, we formally define the problem, introduce a base group recommender unifying existing neural methods, and discuss its limitations. We describe our proposed framework GroupIM in Section 4, present experimental results in Section 5, finally concluding in Section 6.

2. Related Work

Group Recommendation: This line of work can be divided into two categories based on group types: persistent and ephemeral. Persistent groups have stable members with rich activity history together, while ephemeral groups comprise users who interact with very few items together (Quintarelli et al., 2016). A common approach is to consider persistent groups as virtual users (Hu et al., 2014), thus, personalized recommenders can be directly applied. However, such methods cannot handle ephemeral groups with sparse interactions. We focus on the more challenging scenario—recommendations to ephemeral groups.

Prior work either aggregate recommendation results (or item scores) for each member, or aggregate individual member preferences, towards group predictions. They fall into two classes: score (or late) aggregation (Baltrunas et al., 2010) and preference (or early) aggregation (Yuan et al., 2014).

Popular score aggregation strategies include least misery (Baltrunas et al., 2010), average (Berkovsky and Freyne, 2010), maximum satisfaction (Boratto and Carta, 2010), and relevance and disagreement (Amer-Yahia et al., 2009). However, these are hand-crafted heuristics that overlook real-world group interactions.  Baltrunas et al. (2010) compare different strategies to conclude that there is no clear winner, and their relative effectiveness depends on group size and group coherence.

Early preference aggregation strategies (Yu et al., 2006) generate recommendations by constructing a group profile that combines the profiles (raw item histories) of group members. Recent methods adopt a model-based perspective to learn data-driven models. Probabilistic methods (Liu et al., 2012; Yuan et al., 2014; Rakesh et al., 2016) model the group generative process by considering both the personal preferences and relative influence of members, to differentiate their contributions towards group decisions. However, a key weakness is their assumption that users have the same likelihood to follow individual and collective preferences, across different groups. Neural methods explore attention mechanisms (Bahdanau et al., 2014) to learn data-driven preference aggregators (Cao et al., 2018; Vinh Tran et al., 2019; Wang et al., 2019). MoSAN (Vinh Tran et al., 2019) models group interactions via sub-attention networks; however, MoSAN operates on persistent groups while ignoring users’ personal activities. AGREE (Cao et al., 2018) employs attentional networks for joint training over individual and group interactions; yet, the extent of regularization applied on each user (based on personal activities) is the same across groups, which results in degenerate solutions when applied to ephemeral groups with sparse activities.

An alternative approach to tackle interaction sparsity is to exploit external side information, e.g., social network of users (Cao et al., 2019; Yin et al., 2019; Sankar et al., 2020b; Krishnan et al., 2019), personality traits (Zheng, 2018), demographics (Delic et al., 2018b), and interpersonal relationships (Delic et al., 2018a; Gartrell et al., 2010). In contrast, our setting is conservative and does not include extra side information: we know only user and item ids, and item implicit feedback. We address interaction sparsity through novel data-driven regularization and training strategies (Krishnan et al., 2018). Our goal is to enable a wide spectrum of neural group recommenders to seamlessly integrate suitable preference encoders and aggregators.

Mutual Information: Recent neural MI estimation methods (Belghazi et al., 2018) leverage the InfoMax (Linsker, 1988) principle for representation learning. They exploit the structure of the input data (e.g., spatial locality in images) via MI maximization objectives, to improve representational quality. Recent advances employ auto-regressive models (Oord et al., 2018) and aggregation functions (Hjelm et al., 2019; Veličković et al., 2019; Yeh and Chen, 2019)

with noise-contrastive loss functions to preserve MI between structurally related inputs.

We leverage the InfoMax principle to exploit the preference covariance structure shared amongst group members. A key novelty of our approach is MI-guided weighting to regularize group embeddings with the personal preferences of highly relevant members.

3. Preliminaries

In this section, we first formally define the ephemeral group recommendation problem. Then, we present a base neural group recommender that unifies existing neural methods into a general framework. Finally, we analyze the key shortcomings of to discuss motivations for maximizing user-group mutual information.

3.1. Problem Definition

We consider the implicit feedback setting (only visits, no explicit ratings) with a user set , an item set , a group set , a binary user-item interaction matrix , and a binary group-item interaction matrix . We denote , as the corresponding rows for user and group in and , with , indicating their respective number of interacted items.

An ephemeral group comprises a set of users with sparse historical interactions .

Ephemeral Group Recommendation: We evaluate group recommendation on strict ephemeral groups, which have never interacted together before during training. Given a strict ephemeral group during testing, our goal is to generate a ranked list over the item set relevant to users in , i.e., learn a function that maps an ephemeral group and an item to a relevance score, where is the power set of .

3.2. Base Neural Group Recommender

Several neural group recommenders have achieved impressive results (Cao et al., 2018; Vinh Tran et al., 2019). Despite their diversity in modeling group interactions, we remark that state-of-the-art neural methods share a clear model structure: we present a base group recommender that includes three modules: a preference encoder; a preference aggregator; and a joint user and group interaction loss. Unifying these neural group recommenders within a single framework facilitates deeper analysis into their shortcomings in addressing ephemeral groups.

The base group recommender first computes user representations from user-item interactions using a preference encoder , followed by applying a neural preference aggregator to compute the group representation for group . Finally, the group representation is jointly trained over the group and user interactions, to make group recommendations.

3.2.1. User Preference Representations.

User embeddings constitute a latent representation of their personal preferences, indicated in the interaction matrix

. Since latent-factor collaborative filtering methods adopt a variety of strategies (such as matrix factorization, autoencoders, etc.) to learn user embeddings

, we define the preference encoder with two inputs: user

and associated binary personal preference vector



We can augment with additional inputs, including contextual attributes, item relationships, etc. via customized encoders (Zhang et al., 2019).

3.2.2. Group Preference Aggregation.

A preference aggregator models the interactions among group members to compute an aggregate representation for ephemeral group . Since groups are sets of users with no inherent ordering, we consider the class of permutation-invariant functions (such as summation or pooling operations) on sets (Zaheer et al., 2017). Specifically, is permutation-invariant to the order of group member embeddings . We compute using an arbitrary preference aggregator as:


3.2.3. Joint User and Group Loss.

The group representation is trained over the group-item interactions with group-loss . The framework supports different recommendation objectives, including pairwise (Rendle et al., 2009) and pointwise (He et al., 2017) ranking losses. Here, we use a multinomial likelihood formulation owing to its impressive results in user-based neural collaborative filtering (Liang et al., 2018). The group representation

is transformed by a fully connected layer and normalized by a softmax function to produce a probability vector

over . The loss measures the KL-divergence between the normalized purchase history ( indicates items interacted by group ) and predicted item probabilities , given by:


Next, we define the user-loss that regularizes the user representations with user-item interactions , thus facilitating joint training with shared encoder and predictor () layers (Cao et al., 2018). We use a similar multinomial likelihood-based formation, given by:


where denotes the overall loss of the base recommender with balancing hyper-parameter . AGREE (Cao et al., 2018) trains an attentional aggregator with pairwise regression loss over both and , while MoSAN (Vinh Tran et al., 2019) trains a collection of sub-attentional aggregators with bayesian personalized ranking (Rendle et al., 2009) loss on just . Thus, state-of-the-art neural methods AGREE (Cao et al., 2018) and MoSAN (Vinh Tran et al., 2019) are specific instances of the framework described by base recommender .

3.3. Motivation

To address ephemeral groups, we focus on regularization strategies that are independent of the base recommender . With the rapid advances in neural methods, we envision future enhancements in neural architectures for user representations and group preference aggregation. Since ephemeral groups by definition purchase very few items together, base recommenders suffer from inadequate training data in group interactions. Here, the group embedding receives back-propagation signals from sparse interacted items in , thus lacking evidence to reliably estimate the role of each member. To address group interaction sparsity towards robust ephemeral group recommendation, we propose two data-driven regularization strategies that are independent of the base recommendation mechanisms to generate individual and group representations.

3.3.1. Contrastive Representation Learning

We note that users’ preferences are group-dependent; and users occurring together in groups typically exhibit covarying preferences (e.g., shared cuisine tastes). Thus, group activities reveal distinctions across groups (e.g., close friends versus colleagues) and latent user associations (e.g., co-occurrence of users in similar groups), that are not evident when the base recommender only predicts sparse group interactions.

We contrast the preference representations of group members against those of non-member users with similar item histories, to effectively regularize the latent space of user and group representations. This promotes the representations to encode latent discriminative characteristics shared by group members, that are not discernible from their limited interacted items in .

3.3.2. Group-adaptive Preference Prioritization.

To overcome group interaction sparsity, we critically remark that while groups are ephemeral with sparse interactions, the group members have comparatively richer individual interaction histories. Thus, we propose to selectively exploit the personal preferences of group members to enhance the quality of group representations.

The user-loss (equation 4) in base recommender attempts to regularize user embeddings based on their individual activities . A key weakness is that forces to uniformly predict preferences across all groups containing user . Since groups interact with items differently than individual members, inaccurately utilizing can become counter-productive. Fixed regularization results in degenerate models that either over-fit or are over-regularized, due to lack of flexibility in adapting preferences per group.

To overcome group interaction sparsity, we contextually identify members that are highly relevant to the group and regularize the group representation to reflect their personal preferences. To measure contextual relevance, we introduce group-specific relevance weights for each user where is a learned weighting function of both user and group representations. This enhances the expressive power of the recommender, thus effectively alleviating the challenges imposed by group interaction sparsity.

In this section, we defined ephemeral group recommendation, and presented a base group recommender architecture with three modules: user representations, group preference aggregation, and joint loss functions. Finally, we motivated the need to: contrastively regularize the user-group space to capture member associations and group distinctions; and learn group-specific weights to regularize group representations with individual user preferences.

4. GroupIM Framework

In this section, we first motivate mutual information towards achieving our two proposed regularization strategies, followed by a detailed description of our proposed framework GroupIM.

4.1. Mutual Information Maximization.

Figure 1. Neural architecture diagram of GroupIM depicting model components and loss terms appearing in Equation 9.

We introduce our user-group mutual information maximization approach through an illustration. We extend the introductory example to illustrate how to regularize Alice’s latent representation based on her interactions in two different groups. Consider Alice who first goes out for lunch to an Italian restaurant with a visitor Bob, and later dines at a Mexican restaurant with her friend Charlie.

First, Alice plays different roles across the two groups (i.e., stronger influence among friends than with Bob) due to the differences in group context (visitors versus friends). Thus, we require a measure to quantify the contextual informativeness of user in group .

Second, we require the embedding of Alice to capture association with both visitor Bob and friend Charlie, yet express variations in her group activities. Thus, it is necessary to not only differentiate the role of Alice across groups, but also compute appropriate representations that make her presence in each group more coherent.

To achieve these two goals at once, we maximize user-group mutual information (MI) to regularize the latent space of user and group representations, and set group-specific relevance weights in proportion to their estimated MI scores. User-group MI measures the contextual informativeness of a member towards the group decision through the reduction in group decision uncertainty when user is included in group

. Unlike correlation measures that quantify monotonic linear associations, mutual information captures complex non-linear statistical relationships between covarying random variables. Our proposed MI maximization strategy enables us to achieve our two-fold motivation (Section 


  • Altering Latent Representation Geometry: Maximizing user-group MI encourages the group embedding to encode preference covariance across group members, and regularizes the user embeddings to capture social associations in group interactions.

  • Group-specific User Relevance: By quantifying through user-group mutual information, we accurately capture the extent of informativeness for user in group , thus guiding group-adaptive personal preference prioritization.

4.2. User-Group MI Maximization.

Neural MI estimation (Belghazi et al., 2018)

has demonstrated feasibility to maximize MI by training a classifier

(a.k.a, discriminator network) to accurately separate positive

samples drawn from their joint distribution from

negative samples drawn from the product of marginals.

We maximize user-group MI between group member representations and group representation (computed in equations 1 and 2 respectively). We train a contrastive discriminator network , where represents the probability score assigned to this user-group pair (higher scores for users who are members of group ). The positive samples for are the preference representations of pairs such that , and negative samples are derived by pairing with the representations of non-member users sampled from a negative sampling distribution . The discriminator is trained on a noise-contrastive type objective with a binary cross-entropy (BCE) loss between samples from the joint (positive pairs), and the product of marginals (negative pairs), resulting in the following objective:


where , is the number of negative users sampled for group and is a shorthand for . This objective maximizes MI between and based on the Jensen-Shannon divergence between the joint and the product of marginals (Veličković et al., 2019).

We employ a preference-biased negative sampling distribution , which assigns higher likelihoods to non-member users who have purchased the group items . These hard negative examples encourage the discriminator to learn latent aspects shared by group members by contrasting against other users with similar individual item histories. We define as:


where is an indicator function and controls the sampling bias. We set across all our experiments. In comparison to random negative sampling, our experiments indicate that preference-biased negative user sampling exhibits better discriminative abilities.

When is trained jointly with the base recommender loss (equation 4), maximizing user-group MI enhances the quality of user and group representations computed by the encoder and aggregator . We now present our approach to overcome the limitations of the fixed regularizer (Section 3.3).

4.3. Contextual User Preference Weighting

In this section, we describe a contextual weighting strategy to identify and prioritize personal preferences of relevant group members, to overcome group interaction sparsity. We avoid degenerate solutions by varying the extent of regularization induced by each (for user ) across groups through group-specific relevance weights . Contextual weighting accounts for user participation in diverse ephemeral groups with different levels of shared interests.

By maximizing user-group MI, the discriminator outputs scores that quantify the contextual informativeness of each pair (higher scores for informative users). Thus, we set the relevance weight for group member to be proportional to . Instead of regularizing the user representations with in each group ( in eqn 4), we directly regularize the group representation with in proportion to for each group member . Direct optimization of (instead of ) results in more effective regularization, especially with sparse group activities. We define the contextually weighted user-loss as:


where effectively regularizes with the individual activities of group member with contextual weight .

The overall model objective of our framework GroupIM includes three terms: , , and , which is described in Section 4.4.4 in detail.  GroupIM regularizes the latent representations computed by and through user-group MI maximization () to contrastively capture group member associations; and contextual MI-guided weighting () to prioritize individual preferences.

4.4. Model Details

We now describe the architectural details of preference encoder , aggregator , discriminator , and an alternative optimization approach to train our framework GroupIM.

4.4.1. User Preference Encoder

To encode individual user preferences into preference embeddings

, we use a Multi-Layer Perceptron with two fully connected layers, defined by:

with learnable weight matrices and , biases , and activations for non-linearity .

Pre-training: We pre-train the weights and biases of the first encoder layer () on the user-item interaction matrix with user-loss (equation 4). We use these pre-trained parameters to initialize the first layer of before optimizing the overall objective of GroupIM. Our ablation studies in Section 5.5 indicate significant improvements owing to this initialization strategy.

4.4.2. Group Preference Aggregators

We consider three preference aggregators Maxpool, Meanpool and, Attention, which are widely used in graph neural networks (Xu et al., 2019; Hamilton et al., 2017; Sankar et al., 2020a) and have close ties to aggregators examined in prior work, i.e., Maxpool and Meanpool mirror the heuristics of maximum satisfaction (Boratto and Carta, 2010) and averaging (Berkovsky and Freyne, 2010), while attentions learn varying member contributions (Cao et al., 2018; Vinh Tran et al., 2019). We define the three preference aggregators below:

  • [leftmargin=*]

  • Maxpool: The preference embedding of each member is passed through MLP

    layers, followed by element-wise max-pooling to aggregate group member representations, given by:

    where max denotes the element-wise max operator and is a nonlinear activation. Intuitively, the MLP layers compute features for each member, and max-pooling over each of the computed features effectively captures different aspects of group members.

  • Meanpool: We similarly apply an element-wise mean-pooling operation after the MLP, to compute group representation as:

  • Attention: To explicitly differentiate group members’ roles, we employ neural attentions (Bahdanau et al., 2014) to compute a weighted sum of members’ preference representations, where the weights are learned by an attention network, parameterized by a single MLP layer.

    where indicates the contribution of a user towards the group decision. This can be trivially extended to item-conditioned weighting (Cao et al., 2018), self-attention (Wang et al., 2019) and sub-attention networks (Vinh Tran et al., 2019).

4.4.3. Discriminator Architecture

The discriminator architecture learns a scoring function to assign higher scores to observed pairs relative to negative examples, thus parameterizing group-specific relevance . Similar to existing work (Veličković et al., 2019), we use a simple bilinear function to score user-group representation pairs.


where is a learnable scoring matrix and is the logistic sigmoid non-linearity function to convert raw scores into probabilities of being a positive example. We leave investigation of further architectural variants for the discriminator to future work.

4.4.4. Model Optimization

The overall objective of GroupIM is composed of three terms, the group-loss (Equation 3), contextually weighted user-loss (Equation 7), and MI maximization loss (Equation 5). The combined objective is given by:


We train GroupIM using an alternating optimization schedule. In the first step, the discriminator is held constant, while optimizing the group recommender on . The second step trains on , resulting in gradient updates for both parameters of as well as those of the encoder and aggregator .

Thus, the discriminator only seeks to regularize the model (i.e., encoder and aggregator) during training through loss terms and . During inference, we directly use the regularized encoder and aggregator to make group recommendations.

5. Experiments

Dataset Yelp Weeplaces Gowalla Douban
# Users 7,037 8,643 25,405 23,032
# Items 7,671 25,081 38,306 10,039
# Groups 10,214 22,733 44,565 58,983
# U-I interactions 220,091 1,358,458 1,025,500 1,731,429
# G-I Interactions 11,160 180,229 154,526 93,635
Avg. # items/user 31.3 58.83 40.37 75.17
Avg. # items/group 1.09 2.95 3.47 1.59
Avg. group size 6.8 2.9 2.8 4.2
Table 1. Summary statistics of four real-world datasets with ephemeral groups. Group-Item interactions are sparse: average number of interacted items per group ¡ 3.5.

In this section, we present an extensive quantitative and qualitative analysis of our model. We first introduce datasets, baselines, and experimental setup (Section 5.15.2, and 5.3), followed by our main group recommendation results (Section 5.4). In Section 5.5, we conduct an ablation study to understand our gains over the base recommender. In Section 5.6, we study how key group characteristics (group size, coherence, and aggregate diversity) impact recommendation results. In Section 5.7, we visualize the variation in discriminator scores assigned to group members, for different kinds of groups. Finally, we discuss limitations in Section 5.8.

5.1. Datasets

First, we conduct experiments on large-scale POI (point-of-interest) recommendation datasets extracted from three location-based social networks. Since the POI datasets do not contain explicit group interactions, we construct group interactions by jointly using the check-ins and social network information: check-ins at the same POI within 15 minutes by an individual and her subset of friends in the social network together constitutes a single group interaction, while the remaining check-ins at the POI correspond to individual interactions. We define the group recommendation task as recommending POIs to ephemeral groups of users. The datasets were pre-processed to retain users and items with five or more check-ins each. We present dataset descriptions below:

  • [leftmargin=*]

  • Weeplaces 111 we extract check-ins on POIs over all major cities in the United States, across various categories including Food, Nightlife, Outdoors, Entertainment and Travel.

  • Yelp 222 we filter the entire dataset to only include check-ins on restaurants located in the city of Los Angeles.

  • Gowalla (Liu et al., 2014): we use restaurant check-ins across all cities in the United States, in the time period upto June 2011.

Second, we evaluate venue recommendation on Douban, which is the largest online event-based social network in China.

  • [leftmargin=*]

  • Douban (Yin et al., 2019): users organize and participate in social events, where users attend events together in groups and items correspond to event venues. During pre-processing, we filter out users and venues with less than 10 interactions each.

Groups across all datasets are ephemeral since group interactions are sparse (average number of items per group in Table 1)

Dataset Yelp (LA) Weeplaces Gowalla Douban
Metric N@20 N@50 R@20 R@50 N@20 N@50 R@20 R@50 N@20 N@50 R@20 R@50 N@20 N@50 R@20 R@50
Predefined Score Aggregators
Popularity (Cremonesi et al., 2010) 0.000 0.000 0.001 0.001 0.063 0.074 0.126 0.176 0.075 0.088 0.143 0.203 0.003 0.005 0.009 0.018
VAE-CF + AVG (Liang et al., 2018; Berkovsky and Freyne, 2010) 0.142 0.179 0.322 0.513 0.273 0.313 0.502 0.666 0.318 0.362 0.580 0.758 0.179 0.217 0.381 0.558
VAE-CF + LM (Liang et al., 2018; Baltrunas et al., 2010) 0.097 0.120 0.198 0.316 0.277 0.311 0.498 0.640 0.375 0.409 0.610 0.750 0.221 0.252 0.414 0.555
VAE-CF + MAX (Liang et al., 2018; Boratto and Carta, 2010) 0.099 0.133 0.231 0.401 0.229 0.270 0.431 0.604 0.267 0.316 0.498 0.702 0.156 0.194 0.339 0.517
VAE-CF + RD (Liang et al., 2018; Amer-Yahia et al., 2009) 0.143 0.181 0.321 0.513 0.239 0.279 0.466 0.634 0.294 0.339 0.543 0.723 0.178 0.216 0.379 0.557
Data-driven Preference Aggregators
COM (Yuan et al., 2014) 0.143 0.154 0.232 0.286 0.329 0.348 0.472 0.557 0.223 0.234 0.326 0.365 0.283 0.288 0.417 0.436
Crowdrec (Rakesh et al., 2016) 0.082 0.101 0.217 0.315 0.353 0.370 0.534 0.609 0.325 0.338 0.489 0.548 0.121 0.188 0.375 0.681
AGREE (Cao et al., 2018) 0.123 0.168 0.332 0.545 0.242 0.292 0.484 0.711 0.160 0.223 0.351 0.605 0.126 0.173 0.310 0.536
MoSAN (Vinh Tran et al., 2019) 0.470 0.494 0.757 0.875 0.287 0.334 0.548 0.738 0.323 0.372 0.584 0.779 0.193 0.239 0.424 0.639
Group Information Maximization Recommenders (GroupIM)
GroupIM-Maxpool 0.488 0.501 0.676 0.769 0.479 0.505 0.676 0.776 0.433 0.463 0.628 0.747 0.291 0.313 0.524 0.637
GroupIM-Meanpool 0.629 0.637 0.778 0.846 0.518 0.543 0.706 0.804 0.476 0.504 0.682 0.788 0.323 0.351 0.569 0.709
GroupIM-Attention 0.633 0.647 0.782 0.851 0.521 0.546 0.716 0.813 0.477 0.505 0.686 0.796 0.325 0.356 0.575 0.714
Table 2. Group recommendation results on four datasets, R@K and N@K denote the Recall@K and NDCG@K metrics at and . The GroupIM variants indicate maxpool, meanpool, and attention as preference aggregators in our MI maximization framework.  GroupIM achieves significant gains of 31 to 62% NDCG@20 and 3 to 28% Recall@20 over competing group recommenders. Notice that meanpool and attention variants achieve comparable performance across all datasets.

5.2. Baselines

We compare our framework against state-of-the-art baselines that broadly fall into two categories: score aggregation methods with predefined aggregators, and data-driven preference aggregators.

  • [leftmargin=*]

  • Popularity (Cremonesi et al., 2010): recommends items based on item popularity, which is measured by its interaction count in the training set.

  • User-based CF + Score Aggregation: We utilize a state-of-the-art neural recommendation model VAE-CF (Liang et al., 2018), followed by score aggregation via: averaging (AVG), least-misery (LM), maximum satisfaction (MAX), and relevance-disagreement (RD).

  • COM (Yuan et al., 2014): a probabilistic generative model that considers group members’ individual preferences and topic-dependent influence.

  • CrowdRec (Rakesh et al., 2016): a generative model that extends COM through item-specific latent variables capturing their global popularity.

  • MoSAN (Vinh Tran et al., 2019): a neural group recommender that employs a collection of sub-attentional networks to model group member interactions. Since MoSAN originally ignores individual activities , we include into as pseudo-groups with single users.

  • AGREE (Cao et al., 2018): a neural group recommender that utilizes attentional preference aggregation to compute item-specific group member weights, for joint training over individual and group activities.

Figure 2. NDCG@K across size of rank list

. Variance bands indicate 95% confidence intervals over 10 random runs. Existing methods underperform since they either disregard member roles (VAE-CF variants) or overfit to the sparse group activities.  GroupIM contextually identifies informative members and regularizes their representations, to show strong gains.

We tested GroupIM by substituting three preference aggregators, Maxpool, Meanpool, and Attention (Section 4.4.2). All experiments were conducted on a single Nvidia Tesla V100 GPU with PyTorch (Paszke et al., 2017) implementations on the Linux platform. Our implementation of GroupIM and datasets are publicly available333

5.3. Experimental Setup

We randomly split the set of all groups into training (70%), validation (10%), and test (20%) sets, while utilizing the individual interactions of all users for training. Note that each group appears only in one of the three sets. The test set contains strict ephemeral groups (i.e., a specific combination of users) that do not occur in the training set. Thus, we train on ephemeral groups and test on strict ephemeral groups. We use NDCG@K and Recall@K

as evaluation metrics.

We tune the latent dimension in the range and other baseline hyper-parameters in ranges centered at author-provided values. In GroupIM, we use two fully connected layers of size 64 each in and tune in the range . We use 5 negatives for each true user-group pair to train the discriminator.

5.4. Experimental Results

We note the following key observations from our experimental results comparing GroupIM with its three aggregator variants, against competing baselines on group recommendation (Table 2).

First, heuristic score aggregation with neural recommenders (i.e.VAE-CF) performs comparable to (and often beats) probabilistic models (COM, Crowdrec

). Neural methods with multiple non-linear transformations, are expressive enough to identify latent groups of similar users just from their individual interactions.

Second, there is no clear winner among the different pre-defined score aggregation strategies, e.g., VAE-CF + LM (least misery) outperforms the rest on Gowalla and Douban, while VAE-CF + LM (averaging) is superior on Yelp and Weeplaces. This empirically validates the non-existence of a single optimal strategy for all datasets.

Third, MoSAN (Vinh Tran et al., 2019) outperforms both probabilistic models and fixed score aggregators on most datasets.  MoSAN achieves better results owing to the expressive power of neuural preference aggregators (such as sub-attention networks) to capture group member interactions, albeit not explicitly differentiating personal and group activities. Notice that naive joint training over personal and group activities via static regularization (as in AGREE (Cao et al., 2018)) results in poor performance due to sparsity in group interactions. Static regularizers on cannot distinguish the role of users across groups, resulting in models that lack generalisation to ephemeral groups.

GroupIM variants outperform baselines significantly, with attention achieving overall best results. In contrast to neural methods (i.e., MoSAN and AGREE),  GroupIM regularizes the latent representations by contextually weighting the personal preferences of informative members, thus effectively tackling group interaction sparsity. The maxpool variant is noticeably inferior, due to the higher sensitivity of max

operation to outlier group members.

Note that Meanpool performs comparably to attention. This is because in GroupIM, the discriminator does the heavy-lifting of contextually differentiating the role of users across groups to effectively regularize the encoder and aggregator modules. If and are expressive enough, efficient meanpool aggregation can achieve near state-of-the-art results (Table 2)

An important implication is the reduced inference complexity of our model, i.e., once trained using our MI maximizing framework, simple aggregators (such as meanpool) suffice to achieve state-of-the-art performance. This is especially significant, considering that our closest baseline MoSAN (Vinh Tran et al., 2019) utilizes sub-attentional preference aggregation networks that scale quadratically with group size.

We compare the variation in NDCG scores with size of rank list in figure 2. We only depict the best aggregator for VAE-CF.  GroupIM consistently generates more precise recommendations across all datasets. We observe smaller gains in Douban, where the user-item interactions exhibit substantial correlation with corresponding group activities. GroupIM achieves significant gains in characterizing diverse groups, evidenced by our results in section 5.6.

Dataset Weeplaces Gowalla
Metric N@50 R@50 N@50 R@50
Base Group Recommender Variants
(1) Base () 0.420 0.621 0.369 0.572
(2) Base () 0.427 0.653 0.401 0.647
GroupIM Variants
(3) GroupIM () 0.431 0.646 0.391 0.625
(4) GroupIM (Uniform weights) 0.441 0.723 0.418 0.721
(5) GroupIM

(Cosine similarity)

0.488 0.757 0.445 0.739
(6) GroupIM (No pre-training) 0.524 0.773 0.472 0.753
(7) GroupIM () 0.543 0.804 0.505 0.796
Table 3.  GroupIM ablation study (NDCG and Recall at ). Contrastive representation learning (row 3) improves the base recommender (row 1), but is substantially more effective with group-adaptive preference weighting (row 7).

5.5. Model Analysis

In this section, we present an ablation study to analyze several variants of GroupIM, guided by our motivations (Section 3.3). In our experiments, we choose attention as the aggregator due to its consistently high performance. We conduct studies on Weeplaces and Gowalla to report NDCG@50 and Recall@50 in Table 3.

First, we examine the base group recommender (Section 3.2) which does not utilize MI maximization for model training (Table 3).

Base Group Recommender.

We examine two variants below:
(1) We train on just group activities with loss (equation 3).
(2) We train jointly on individual and group activities with static regularization on using joint loss (Equation 4).

In comparison to similar neural aggregator MoSAN, our base recommender is stronger on NDCG but inferior on Recall. The difference is likely due to the multinomial likelihood used to train , in contrast to the ranking loss in MoSAN. Static regularization via (row 1) results in higher gains for Gowalla (richer user-item interactions) with relatively larger margins for Recall than NDCG.

Next, we examine model variants of GroupIM in two parts:

GroupIM: Contrastive Representation Learning.

We analyze the benefits derived by just training the contrastive discriminator to capture group member associations, i.e., we define a model variant (row 3) to optimize just , without the term. Direct MI maximization (row 3) improves over the base recommender (row 1), validating the benefits of contrastive regularization, however still suffers from lack of user preference prioritization.

GroupIM: Group-adaptive Preference Prioritization

We analyze the utility of data-driven contextual weighting (via user-group MI), by examining two alternate fixed strategies to define :
(4) Uniform weights: We assign the same weight for each group member in group , when optimizing .
(5) Cosine similarity: To model user-group correlation, we set the weight as the cosine similarity between and .

From table 3 (rows 4 and 5), the uniform weights variant of loss (row 4) surpasses the statically regularized model (row 2), due to more direct feedback from to group embedding during model training. Cosine similarity (row 5) achieves stronger gains owing to more accurate correlation-guided user weighting across groups. Our model GroupIM (row 7) has strong gains over the fixed weighting strategies as a result of its regularization strategy to contextually identify informative members across groups.

GroupIM: Pre-training on

We depict model performance without pre-training (random initializations) in row 6. Our model (row 7) achieves noticeable gains; pre-training identifies good model initialization points for better convergence.

5.6. Impact of Group Characteristics

In this section, we examine our results to understand the reason for GroupIM’s gains. We study ephemeral groups along three facets: group size; group coherence; and group aggregate diversity.

Figure 3. Performance (NDCG@50), across group size ranges. GroupIM has larger gains for larger groups due to accurate user associations learnt via MI maximization.

5.6.1. Group Size

We classify test groups into bins based on five levels of group size (2-3, 4-5, 6-7, 8-9, and 10). Figure 3 depicts the variation in NDCG@50 scores on Weeplaces and Gowalla datasets.

We make three key observations: methods that explicitly distinguish individual and group activities (such as COM, CrowdRec, GroupIM), exhibit distinctive trends wrt group size. In contrast, MoSAN (Vinh Tran et al., 2019) and AGREE (Cao et al., 2018), which either uniformly mix both behaviors or apply static regularizers, show no noticeable variation; Performance generally increases with group size. Although test groups are previously unseen, for larger groups, subsets of inter-user interactions are more likely to be seen during training, thus resulting in better performance;  GroupIM achieves higher (or steady) gains for groups of larger sizes owing to its more accurate prioritization of personal preferences for each member, e.g.,  GroupIM clearly has stronger gains for groups of sizes 8-9 and in Gowalla.

5.6.2. Group Coherence

We define group coherence as the mean pair-wise correlation of personal activities () of group members, i.e.

, if a group has users who frequently co-purchase items, it receives greater coherence. We separate test groups into four quartiles by their coherence scores. Figure 

4 depicts NDCG@50 for groups under each quartile (Q1 - Lower values), on Weeplaces and Gowalla.

GroupIM has stronger gains for groups with low coherence (quartiles Q1 and Q2), which empirically validates the efficacy of contextual user preference weighting in regularizing the encoder and aggregator, for groups with dissimilar member preferences.

Figure 4. Performance (NDCG@50), across group coherence quartiles (Q1: lowest, Q4: highest). GroupIM has larger gains in Q1 & Q2 (low group coherence).
Figure 5. Performance (NDCG@50), across group aggregate diversity quartiles (Q1: lowest, Q4: highest). GroupIM has larger gains in Q3 & Q4 (high diversity).

5.6.3. Group Aggregate Diversity

We adapt the classical aggregate diversity metric (Vargas and Castells, 2011) to define group aggregate diversity as the total number of distinct items interacted across all group members, i.e., if the set of all purchases of group members covers a wider range of items, then the group has higher aggregate diversity. We report NDCG@50 across aggregate diversity quartiles in figure 5.

Model performance typically decays (and stabilizes), with increase in aggregate diversity. Diverse groups with large candidate item sets, pose an information overload for group recommenders, leading to worse results. Contextual prioritization with contrastive learning, benefits diverse groups, as evidenced by the higher relative gains of GroupIM  for diverse groups (quartiles Q3 and Q4).

Figure 6. MI variation (std. deviation in discriminator scores over members) per group coherence quartile across group sizes. For groups of a given size, as coherence increases, MI variation decreases. As groups increase in size, MI variation increases.

5.7. Qualitative MI Discriminator Analysis

We examine the contextual weights estimated by GroupIM over test ephemeral groups, across group size and coherence.

We divide groups into four bins based on group sizes (2-3, 4-6, 7-9, and ), and partition them into quartiles based on group coherence within each bin. To analyze the variation in contextual informativeness across group members, we compute MI variation

as the standard deviation of scores given by

over group members. Figure 6 depicts letter-value plots of MI variation for groups in corresponding coherence quartiles across group sizes on Weeplaces.

MI variation increases with group size, since larger groups often comprise users with divergent roles and interests. Thus, the discriminator generalizes to unseen groups, to discern and estimate markedly different relevance scores for each group member. To further examine the intuition conveyed by the scores, we compare MI variation across group coherence quartiles within each size-range.

MI variation is negatively correlated with group coherence for groups of similar sizes, e.g., MI variation is consistently higher for groups with low coherence (quartiles Q1 and Q2). For highly coherent groups (quartile Q4), assigns comparable scores across all members, which is consistent with our intuitions and earlier results on the efficacy of simple averaging strategies for such groups.

We also analyze parameter sensitivity to user-preference weight . Low values result in overfitting to the group activities , while larger values result in degenerate solutions that lack group distinctions (plot excluded for the sake of brevity).

5.8. Limitations

We identify two limitations of our work. Despite learning to contextually prioritize users’ preferences across groups, controls the overall strength of preference regularization. Since optimal varies across datasets and applications, we plan to explore meta-learning approaches to eliminate such hyper-parameters (Ren et al., 2018).

GroupIM relies on user-group MI estimation to contextually identify informative members, which might become challenging when users have sparse individual interaction histories. In such a scenario, side information (e.g., social network of users), or contextual factors (e.g., location, interaction time) (Krishnan et al., 2020) can prove effective.

6. Conclusion

This paper introduces a recommender architecture-agnostic framework GroupIM that integrates arbitrary neural preference encoders and aggregators for ephemeral group recommendation. To overcome group interaction sparsity, GroupIM regularizes the user-group representation space by maximizing user-group MI to contrastively capture preference covariance among group members. Unlike prior work that incorporate individual preferences through static regularizers, we dynamically prioritize the preferences of informative members through MI-guided contextual preference weighting. Our extensive experiments on four real-world datasets show significant gains for GroupIM over state-of-the-art methods.


  • (1)
  • Amer-Yahia et al. (2009) Sihem Amer-Yahia, Senjuti Basu Roy, Ashish Chawlat, Gautam Das, and Cong Yu. 2009. Group recommendation: Semantics and efficiency. VLDB 2, 1 (2009), 754–765.
  • Bahdanau et al. (2014) Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473 (2014).
  • Baltrunas et al. (2010) Linas Baltrunas, Tadas Makcinskas, and Francesco Ricci. 2010. Group recommendations with rank aggregation and collaborative filtering. In RecSys. 119–126.
  • Belghazi et al. (2018) Mohamed Ishmael Belghazi, Aristide Baratin, Sai Rajeshwar, Sherjil Ozair, Yoshua Bengio, Devon Hjelm, and Aaron Courville. 2018. Mutual Information Neural Estimation. In ICML. 530–539.
  • Berkovsky and Freyne (2010) Shlomo Berkovsky and Jill Freyne. 2010. Group-based recipe recommendations: analysis of data aggregation strategies. In RecSys. ACM, 111–118.
  • Boratto and Carta (2010) Ludovico Boratto and Salvatore Carta. 2010. State-of-the-art in group recommendation and new approaches for automatic identification of groups. In Information retrieval and mining in distributed environments. Springer, 1–20.
  • Cao et al. (2018) Da Cao, Xiangnan He, Lianhai Miao, Yahui An, Chao Yang, and Richang Hong. 2018. Attentive group recommendation. In SIGIR. ACM, 645–654.
  • Cao et al. (2019) Da Cao, Xiangnan He, Lianhai Miao, Guangyi Xiao, Hao Chen, and Jiao Xu. 2019. Social-Enhanced Attentive Group Recommendation. IEEE TKDE (2019).
  • Cremonesi et al. (2010) Paolo Cremonesi, Yehuda Koren, and Roberto Turrin. 2010. Performance of recommender algorithms on top-n recommendation tasks. In RecSys. 39–46.
  • Delic et al. (2018a) Amra Delic, Judith Masthoff, Julia Neidhardt, and Hannes Werthner. 2018a. How to Use Social Relationships in Group Recommenders: Empirical Evidence. In UMAP. ACM, 121–129.
  • Delic et al. (2018b) Amra Delic, Julia Neidhardt, Thuy Ngoc Nguyen, and Francesco Ricci. 2018b. An observational user study for group recommender systems in the tourism domain. Information Technology & Tourism 19, 1-4 (2018), 87–116.
  • Gartrell et al. (2010) Mike Gartrell, Xinyu Xing, Qin Lv, Aaron Beach, Richard Han, Shivakant Mishra, and Karim Seada. 2010. Enhancing group recommendation by incorporating social relationship interactions. In GROUP. 97–106.
  • Hamilton et al. (2017) Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. In NIPS. 1024–1034.
  • He et al. (2017) Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural collaborative filtering. In WWW. 173–182.
  • Hjelm et al. (2019) R. Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Philip Bachman, Adam Trischler, and Yoshua Bengio. 2019. Learning deep representations by mutual information estimation and maximization. In ICLR.
  • Hu et al. (2014) Liang Hu, Jian Cao, Guandong Xu, Longbing Cao, Zhiping Gu, and Wei Cao. 2014. Deep modeling of group preferences for group-based recommendation. In AAAI.
  • Krishnan et al. (2019) Adit Krishnan, Hari Cheruvu, Cheng Tao, and Hari Sundaram. 2019. A modular adversarial approach to social recommendation. In CIKM. 1753–1762.
  • Krishnan et al. (2020) Adit Krishnan, Mahashweta Das, Mangesh Bendre, Hao Yang, and Hari Sundaram. 2020. Transfer Learning via Contextual Invariants for One-to-Many Cross-Domain Recommendation. arXiv preprint arXiv:2005.10473 (2020).
  • Krishnan et al. (2018) Adit Krishnan, Ashish Sharma, Aravind Sankar, and Hari Sundaram. 2018. An adversarial approach to improve long-tail performance in neural collaborative filtering. In CIKM. 1491–1494.
  • Liang et al. (2018) Dawen Liang, Rahul G Krishnan, Matthew D Hoffman, and Tony Jebara. 2018. Variational autoencoders for collaborative filtering. In WWW. 689–698.
  • Linsker (1988) Ralph Linsker. 1988. Self-organization in a perceptual network. Computer 21, 3 (1988), 105–117.
  • Liu et al. (2012) Xingjie Liu, Yuan Tian, Mao Ye, and Wang-Chien Lee. 2012. Exploring personal impact for group recommendation. In CIKM. ACM, 674–683.
  • Liu et al. (2014) Yong Liu, Wei Wei, Aixin Sun, and Chunyan Miao. 2014. Exploiting geographical neighborhood characteristics for location recommendation. In CIKM. ACM, 739–748.
  • Oord et al. (2018) Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018).
  • Paszke et al. (2017) Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. 2017. Automatic differentiation in PyTorch. In NIPS-W.
  • Quintarelli et al. (2016) Elisa Quintarelli, Emanuele Rabosio, and Letizia Tanca. 2016. Recommending new items to ephemeral groups using contextual user influence. In RecSys. ACM, 285–292.
  • Rakesh et al. (2016) Vineeth Rakesh, Wang-Chien Lee, and Chandan K Reddy. 2016. Probabilistic group recommendation model for crowdfunding domains. In WSDM. ACM, 257–266.
  • Ren et al. (2018) Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. 2018.

    Learning to Reweight Examples for Robust Deep Learning. In

    ICML. 4331–4340.
  • Rendle et al. (2009) Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme. 2009. BPR: Bayesian personalized ranking from implicit feedback. In UAI. AUAI Press, 452–461.
  • Sankar et al. (2020a) Aravind Sankar, Yanhong Wu, Liang Gou, Wei Zhang, and Hao Yang. 2020a. DySAT: Deep Neural Representation Learning on Dynamic Graphs via Self-Attention Networks. In WSDM. 519–527.
  • Sankar et al. (2020b) Aravind Sankar, Xinyang Zhang, Adit Krishnan, and Jiawei Han. 2020b. Inf-VAE: A Variational Autoencoder Framework to Integrate Homophily and Influence in Diffusion Prediction. In WSDM. 510–518.
  • Vargas and Castells (2011) Saúl Vargas and Pablo Castells. 2011. Rank and relevance in novelty and diversity metrics for recommender systems. In RecSys. ACM, 109–116.
  • Veličković et al. (2019) Petar Veličković, William Fedus, William L. Hamilton, Pietro Liò, Yoshua Bengio, and R Devon Hjelm. 2019. Deep Graph Infomax. In ICLR.
  • Vinh Tran et al. (2019) Lucas Vinh Tran, Tuan-Anh Nguyen Pham, Yi Tay, Yiding Liu, Gao Cong, and Xiaoli Li. 2019. Interact and Decide: Medley of Sub-Attention Networks for Effective Group Recommendation. In SIGIR. ACM, 255–264.
  • Wang et al. (2019) Haiyan Wang, Yuliang Li, and Felix Frimpong. 2019. Group Recommendation via Self-Attention and Collaborative Metric Learning Model. IEEE Access 7 (2019), 164844–164855.
  • Xu et al. (2019) Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. 2019. How Powerful are Graph Neural Networks?. In ICLR.
  • Yeh and Chen (2019) Yi-Ting Yeh and Yun-Nung Chen. 2019. QAInfomax: Learning Robust Question Answering System by Mutual Information Maximization. In EMNLP. 3368–3373.
  • Yin et al. (2019) Hongzhi Yin, Qinyong Wang, Kai Zheng, Zhixu Li, Jiali Yang, and Xiaofang Zhou. 2019. Social influence-based group representation learning for group recommendation. In ICDE. IEEE, 566–577.
  • Yu et al. (2006) Zhiwen Yu, Xingshe Zhou, Yanbin Hao, and Jianhua Gu. 2006. TV program recommendation for multiple viewers based on user profile merging. User modeling and user-adapted interaction 16, 1 (2006), 63–82.
  • Yuan et al. (2014) Quan Yuan, Gao Cong, and Chin-Yew Lin. 2014. COM: a generative model for group recommendation. In KDD. ACM, 163–172.
  • Zaheer et al. (2017) Manzil Zaheer, Satwik Kottur, Siamak Ravanbakhsh, Barnabas Poczos, Ruslan R Salakhutdinov, and Alexander J Smola. 2017. Deep sets. In NIPS. 3391–3401.
  • Zhang et al. (2019) Shuai Zhang, Lina Yao, Aixin Sun, and Yi Tay. 2019. Deep Learning Based Recommender System: A Survey and New Perspectives. ACM Comput. Surv. 52, 1 (2019), 5:1–5:38.
  • Zheng (2018) Yong Zheng. 2018. Identifying Dominators and Followers in Group Decision Making based on The Personality Traits. In IUI Workshops.