Attentive Group Recommendation (AGR): An Attention‑Based Deep Learning Model for Group Recommendation

Abstract Group recommendation is challenging because a good model must capture the decision process of a group, where users tend to follow a few opinion leaders or experts. To address this, we introduce an attention‑based model (AGR) that learns dynamic influence weights for each group member and aggregates their preferences accordingly.

Introduction Existing group recommendation approaches fall into memory‑based (preference or score aggregation) and model‑based methods, but both fail to model interactions among members. Users often rely on experts in specific domains, and their influence varies across groups. Our motivation is to model these dynamics using attention.

Methodology AGR combines collaborative filtering with an attention network. Each user is represented by a latent vector and a user‑context vector. The attention sub‑network computes a weight for every member based on these vectors, allowing the model to emphasize influential users. The group representation is obtained by summing the weighted sub‑group representations, and final item scores are predicted using a BPR pairwise loss.

Attention Mechanism The attention network consists of two fully‑connected layers followed by a softmax to produce normalized weights. Formally, the attention score for user i in group g is computed as a_i = softmax(w^T ReLU(V u_i + b)) , where u_i is the user latent vector and V is the context matrix.

Experiments We evaluate AGR on four datasets (Plancast, Meetup, MovieLens‑Simi, MovieLens‑Rand) and compare it with six baselines: CF‑AVG, CF‑LM, CF‑RD, PIT, COM, and MF‑AVG. Metrics include precision, recall, and NDCG for top‑K (K=5,10,20). AGR consistently outperforms all baselines, achieving improvements of 6–28% depending on the dataset, with statistical significance (p<0.0001).

Results Analysis AGR’s advantage stems from its ability to capture dynamic user influence and inter‑member interactions, which traditional aggregation or probabilistic models lack. The model also remains robust on randomly generated groups (MovieLens‑Rand), demonstrating flexibility.

Conclusion AGR is the first attention‑based group recommendation model, dynamically learning user influence weights and modeling group interactions, leading to superior recommendation quality. Future work includes incorporating auxiliary information such as social links, textual descriptions, or temporal signals.

References The paper cites nine related works covering group recommendation, attention mechanisms, and deep learning frameworks.