PaperlayerView exact revision on arXiv

Source-linked AI summary

GroupIM: A Mutual Information Maximization Framework for Neural Group Recommendation

Aravind Sankar, Yanhong Wu, Yuhang Wu, Wei Zhang, Hao Yang, Hari Sundaram

arXiv:2006.03736v2cs.IRcs.LGcs.SI

TL;DR

Ephemeral groups have little shared interaction history, unlike persistent groups that prior studies primarily target. GroupIM combines mutual-information regularization and contextual member-preference weighting in an architecture-agnostic framework; across four public datasets, it reports higher recommendation metrics than state-of-the-art methods.

  • Problem

    Ephemeral groups have limited or no shared item history, while existing studies primarily target persistent groups with substantial group activity.

    Source: §1 · p. 1

  • Method

    GroupIM is an architecture-agnostic framework that maximizes mutual information between group and member representations and contextually weights informative members’ preferences.

    Source: §6 · pp. 9–10

  • Results

    31–62% relative NDCG@20 and 3–28% Recall@20 gains over state-of-the-art group recommenders across four publicly available datasets.

    Source: §1 · p. 2

  • Takeaways & Limitations

    GroupIM can integrate arbitrary neural preference encoders and aggregators for ephemeral group recommendation.

    Source: §6 · pp. 9–10

    Estimating user–group mutual information to identify informative members may be challenging when users have sparse individual interaction histories.

    Source: §5.8 · p. 9

Abstract

from arXiv · show

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% relative NDCG@20) over state-of-the-art group recommendation techniques.

1 INTRODUCTION

Ephemeral groups have little or no shared interaction history, while members’ preferences can vary by group context. GroupIM addresses this sparsity with data-driven mutual-information regularization and contextual weighting, and reports gains over state-of-the-art methods.

1 INTRODUCTION

  • Ephemeral groups lack shared interaction histories, and members’ individual preferences can depend on which other users are present.The motivating example contrasts a group’s possible Italian dining choice with Alice’s Mexican-food preferences among other friends.Source: §1 · p. 1
  • Prior methods target persistent groups and may ignore individual activity or apply the same individual-preference influence across groups.External side information is used by some methods but may be unavailable.Source: §1 · p. 1
  • Group members may have rich individual histories even when their shared group history is sparse, motivating data-driven guidance for learning informative group representations.The paper organizes the formal problem and base recommender in Section 3, its framework in Section 4, and experiments in Section 5.Source: §1 · pp. 1–2
  • GroupIM maximizes mutual information between group and member representations, then weights members’ preferences according to their contextual informativeness.The weighting is proportional to user–group mutual information and regularizes the group representation with informative members’ preferences.Source: §1 · p. 2
  • 31–62% relative NDCG@20 and 3–28% Recall@20 gains over state-of-the-art group recommenders across four public datasets.Gains are stronger for larger groups and groups with diverse member preferences.Source: §1 · p. 2
  • GroupIM is architecture-agnostic, integrating arbitrary neural preference encoders and aggregators; meanpool achieves state-of-the-art results with reduced inference costs and no loss of expressivity.Its data-driven regularization uses preference covariance among group members and the contextual relevance of their personal preferences.Source: §1 · pp. 1–2

2 RELATED WORK

Related work covers persistent and ephemeral groups, two broad aggregation strategies, and methods that use user histories or side information. GroupIM instead aims to support varied neural recommenders through data-driven regularization without extra side information.

2 RELATED WORK

  • Ephemeral groups have few interactions together, unlike persistent groups with stable membership and rich shared activity; the paper focuses on the former.Treating persistent groups as virtual users permits personalized recommenders, but does not handle ephemeral groups’ sparse interactions.Source: §2 · p. 2
  • Group recommendation aggregates either members’ scores late or their preferences early; heuristic score aggregation has no clear winner across group sizes and coherence levels.Least misery, average, maximum satisfaction, and relevance and disagreement are cited as hand-crafted score-aggregation strategies.Source: §2 · p. 2
  • Model-based early aggregation includes probabilistic approaches that model member influence and neural approaches that learn data-driven preference aggregators.MoSAN targets persistent groups and ignores personal activities, while AGREE applies the same degree of personal-preference regularization across groups.Source: §2 · p. 2
  • Some sparsity approaches use social networks, traits, demographics, or relationships, whereas this paper uses only user and item IDs and implicit feedback.Its stated goal is to let neural group recommenders integrate suitable preference encoders and aggregators.Source: §2 · p. 2
  • GroupIM applies the InfoMax principle to shared preference covariance and uses MI-guided weighting to regularize group embeddings with highly relevant members’ preferences.Prior neural MI work uses contrastive objectives to preserve information between structurally related inputs.Source: §2 · p. 2

3 PRELIMINARIES

The section defines recommendation for strict ephemeral groups and presents a neural recommender that encodes users, aggregates their preferences, and trains on user and group interactions. It motivates data-driven regularization to address sparse group histories and the limits of uniform user-preference regularization.

3.1 Problem Definition

  • Strict ephemeral group recommendation ranks items for a group that has never interacted together during training, despite its sparse group history.The task maps an ephemeral group and an item to a relevance score.Source: §3.1 · p. 3

3.2 Base Neural Group Recommender

  • The base recommender encodes individual preferences, aggregates member embeddings into a group representation, and jointly trains on user and group interactions.Its architecture unifies existing neural methods, including AGREE and MoSAN as specific instances.Source: §3.2 · p. 3

3.3 Motivation

  • The base user-loss regularizes user embeddings to predict personal histories uniformly across groups, which can overfit or over-regularize when group preferences differ.The section identifies inflexible use of individual activity as counter-productive because groups interact differently from individual members.Source: §3.3 · p. 4
  • The proposed remedies contrastively regularize user and group representations to capture member associations, and use group-specific weights to prioritize relevant members’ preferences.The motivation draws on covarying preferences among users in the same group and comparatively richer individual histories.Source: §3.3 · pp. 3–4

4 GROUPIM FRAMEWORK

GroupIM maximizes mutual information between member and group representations to capture their associations and estimate members’ contextual relevance. It uses those relevance scores to weight personal preferences in group representations, within an architecture-agnostic framework trained by alternating optimization.

4.1 Mutual Information Maximization.

  • User-group mutual information (MI) both shapes representations to capture member associations and supplies group-specific relevance weights for prioritizing personal preferences.The paper frames MI as measuring a member’s contextual informativeness through the reduction in group decision uncertainty when that member is included.Source: §4.1 · p. 4

4.2 User-Group MI Maximization.

  • The discriminator distinguishes observed member-group pairs from negative pairs formed by pairing a group with sampled non-member users.The objective maximizes MI between member and group representations; preference-biased negatives favor non-members who purchased group items, and outperform random negative sampling in the reported experiments.Source: §4.2 · p. 4

4.3 Contextual User Preference Weighting

  • Discriminator scores estimate member relevance, and GroupIM uses them to regularize group representations with the personal histories of contextually informative members.The paper reports that directly optimizing the group representation is more effective, especially with sparse group activity.Source: §4.3 · p. 5

4.4 Model Details

  • GroupIM is designed to integrate arbitrary neural preference encoders and aggregators, including maxpool, meanpool, and attention variants.The paper describes a two-layer MLP user encoder and a bilinear discriminator; further discriminator architecture variants are left for future work.Source: §4.4 · pp. 5–6
  • GroupIM combines group recommendation loss, contextually weighted user loss, and MI loss, training them with alternating optimization.The recommender is optimized with the discriminator held constant, then the discriminator is trained on MI loss; at inference, recommendations use the regularized encoder and aggregator without the discriminator.Source: §4.4 · p. 6

5 EXPERIMENTS

The experiments evaluate ephemeral-group recommendation on POI and event-venue datasets against heuristic and data-driven group recommenders. GroupIM gains 31–62% NDCG@20 and 3–28% Recall@20 over competing group recommenders.

5.1 Datasets

  • POI datasets come from three location-based social networks; Douban supplies venue recommendations from social events.POI group interactions are constructed from check-ins at the same venue within 15 minutes by a user and a subset of friends.Source: §5.1 · p. 6

5.2 Baselines

  • Baselines include popularity, neural VAE-CF with score aggregation, probabilistic COM and CrowdRec, and neural preference aggregators MoSAN and AGREE.VAE-CF is paired with averaging, least-misery, maximum-satisfaction, and relevance-disagreement strategies.Source: §5.2 · p. 6
  • 31–62% NDCG@20 and 3–28% Recall@20 gains over competing group recommenders are reported across four datasets.Meanpool and attention variants achieve comparable performance across all datasets.Source: Table 2 · p. 7

5.3 Experimental Setup

The setup trains on ephemeral groups and evaluates on held-out user combinations, using ranking metrics at varying cutoffs. Model dimensions, regularization, and discriminator negatives are tuned or specified.

5.3 Experimental Setup

  • Groups are split 70%/10%/20% into training, validation, and test sets, with each group appearing in only one split.Test groups are strict ephemeral combinations absent from training; individual interactions from all users are used for training.Source: §5.3 · p. 7
  • Performance is evaluated with NDCG@K and Recall@K.The text identifies these as the evaluation metrics without specifying particular K values here.Source: §5.3 · p. 7
  • Latent dimensions are tuned over {32, 64, 128}, and GroupIM uses two 64-unit fully connected encoder layers and five negatives per true user-group pair.The regularization parameter λ is tuned over {2−4, 2−3, . . . , 26}.Source: §5.3 · p. 7

5.4 Experimental Results

GroupIM consistently produces more precise recommendations across datasets, with smaller gains on Douban. Its contextually weighted regularization and simple aggregators distinguish its results from competing approaches.

5.4 Experimental Results

  • VAE-CF with heuristic score aggregation performs comparably to, and often better than, probabilistic models COM and CrowdRec.The authors attribute this to neural methods' ability to identify latent groups of similar users from individual interactions.Source: §5.4 · p. 7
  • Least-misery aggregation leads on Gowalla and Douban, while averaging leads on Yelp and Weeplaces, with no single strategy best across datasets.The compared strategies are VAE-CF paired with predefined score aggregation.Source: §5.4 · p. 7
  • MoSAN outperforms probabilistic models and fixed score aggregators on most datasets, while static regularization in AGREE performs poorly under sparse group interactions.The text says static regularizers cannot distinguish user roles across groups, resulting in models that lack generalization to ephemeral groups.Source: §5.4 · p. 7
  • GroupIM variants significantly outperform baselines, and attention achieves the best overall results.The maxpool variant is noticeably inferior, which the authors attribute to greater sensitivity to outlier members.Source: §5.4 · p. 7
  • Meanpool performs comparably to attention in GroupIM, while the discriminator contextually differentiates users' roles across groups.Once trained, simple aggregators such as meanpool suffice for state-of-the-art performance; MoSAN's sub-attentional networks scale quadratically with group size.Source: §5.4 · pp. 7–8
  • Contrastive representation learning improves the base recommender, but is substantially more effective with group-adaptive preference weighting.Table 3 reports the ablation at NDCG and Recall K = 50.Source: Table 3 · p. 8
  • GroupIM consistently generates more precise recommendations across all datasets, with smaller gains on Douban.Figure 2 compares NDCG across rank-list sizes; the caption describes strong gains from contextual identification of informative members and representation regularization.Source: §5.4 · p. 8, Figure 2 · p. 8

5.5 Model Analysis

The ablations compare the base recommender with contrastive learning, contextual preference weighting, and encoder pre-training. The results favor combining contrastive representation learning with group-adaptive weighting.

5.5 Model Analysis

  • The base recommender is stronger than MoSAN on NDCG but weaker on Recall; the authors suggest the difference may reflect multinomial-likelihood versus ranking-loss training.Static regularization on individual interactions yields higher gains on Gowalla, which has richer user-item interactions, with relatively larger margins for Recall than NDCG.Source: §5.5 · p. 8
  • Contrastive representation learning improves the base recommender, but still lacks user-preference prioritization without the user-group loss.The contrastive-only variant optimizes LG + LMI without the LUG term.Source: §5.5 · p. 8
  • GroupIM outperforms fixed preference-weighting strategies by contextually identifying informative members across groups.Uniform weights exceed static regularization, and cosine-similarity weights achieve stronger gains through more accurate correlation-guided weighting.Source: §5.5 · p. 8
  • Pre-training the encoder on individual interactions yields noticeable gains over random initialization and identifies better initialization points for convergence.The comparison is between the model without pre-training and the model with encoder pre-training on XU.Source: §5.5 · p. 8

5.6 Impact of Group Characteristics

GroupIM’s relative gains vary with group characteristics: they are larger for bigger groups, groups with lower coherence, and groups with higher aggregate diversity.

5.6.1 Group Size

  • GroupIM has larger gains for larger groups, with stronger gains for groups of sizes 8–9 and ≥10 in Gowalla.The analysis bins test groups into five size ranges and reports NDCG@50 across Weeplaces and Gowalla.Source: §5.6 · pp. 8–9, Figure 3 · p. 9

5.6.2 Group Coherence

  • GroupIM has stronger gains for groups in the low-coherence quartiles Q1 and Q2.Coherence is the mean pair-wise correlation of members’ personal activities; Figure 4 reports NDCG@50 across coherence quartiles.Source: §5.6 · p. 9, Figure 4 · p. 9

5.6.3 Group Aggregate Diversity

  • GroupIM has higher relative gains for groups in the high aggregate-diversity quartiles Q3 and Q4.Aggregate diversity is the number of distinct items interacted with across group members; Figure 5 reports NDCG@50 by diversity quartile.Source: §5.6 · p. 9, Figure 5 · p. 9

5.7 Qualitative MI Discriminator Analysis

The analysis examines how discriminator-score variation relates to group size and coherence, and reports sensitivity to the preference-regularization weight λ.

MI Variation

  • MI variation increases with group size, as larger groups often comprise users with divergent roles and interests.MI variation is the standard deviation of discriminator scores across group members.Source: §5.7 · p. 9
  • Among groups of similar sizes, MI variation is negatively correlated with coherence and is higher in quartiles Q1 and Q2.For highly coherent groups in Q4, the discriminator assigns comparable scores across members.Source: §5.7 · p. 9

Parameter Sensitivity

  • Low λ values result in overfitting to group activities, while larger values result in degenerate solutions lacking group distinctions.The sensitivity plot is excluded for brevity.Source: §5.7 · p. 9

5.8 Limitations

The authors identify two limitations: preference regularization depends on λ, and MI-based member prioritization may be challenging with sparse individual histories.

Limitations

  • The overall strength of preference regularization is controlled by λ, whose optimal value varies across datasets and applications.The authors plan to explore meta-learning approaches to eliminate this hyper-parameter.Source: §5.8 · p. 9
  • User-group MI estimation may be challenging when users have sparse individual interaction histories.The authors note that social-network side information or contextual factors such as location and interaction time can prove effective in this scenario.Source: §5.8 · p. 9

6 CONCLUSION

GroupIM is an architecture-agnostic framework for ephemeral group recommendation that regularizes user-group representations using mutual information and contextual preference weighting. Experiments on four real-world datasets show significant gains over state-of-the-art methods.

6 CONCLUSION

  • GroupIM integrates arbitrary neural preference encoders and aggregators for ephemeral group recommendation.The framework regularizes user-group representations by maximizing user-group mutual information and dynamically prioritizing informative members’ preferences.Source: §6 · pp. 9–10
  • For groups of a given size, MI variation decreases as coherence increases, while MI variation increases with group size.Figure 6 describes MI variation as the standard deviation of discriminator scores over members.Source: Figure 6 · p. 10
  • Experiments on four real-world datasets show significant gains over state-of-the-art methods.The framework is designed to overcome group interaction sparsity through user-group MI and contextual preference weighting.Source: §6 · p. 10

Related papers

Loading 2006.03736v2…