Point-of-interest (POI) recommendation models based on graph neural networks achieve strong performance by propagating collaborative signals over user-item interactions, yet they struggle with the cold-start problem, where items with few or no interactions are not represented. In this paper, we propose LLM-augmented Multi-Graph Contrastive Learning (LLM-MGCL), a multi-graph neural network that uses semantic and spatial information about items to extend the LightGCN backbone with two auxiliary item-item graphs: a semantic graph constructed from sentence embeddings of LLM-generated photo summaries and keywords, and a geographic graph derived from Haversine distances between business locations. Item embeddings are propagated over all three graphs in parallel, fused additively, and aligned across views through a bidirectional InfoNCE contrastive objective that connects behavioral, semantic, and spatial representations of the same items. Experiments on the Yelp Multimodal Recommendation Dataset show that LLM-MGCL outperforms classical collaborative filtering, matrix factorization, and non-contrastive graph neural network baselines. It improves Recall@20 by 48.5% and NDCG@20 by 59.0% over LightGCN, while trailing the strongest contrastive baseline, Self-supervised Graph Learning (SGL), slightly in aggregate. A popularity-stratified evaluation shows that LLM-MGCL achieves the highest recall on cold and warm items, whereas SGL remains better on popular items. An ablation study reveals that the cross-view contrastive alignment (CA) is the primary driver of the aggregate gains. Our results suggest that externally grounded, LLM-derived item knowledge can partially compensate for missing collaborative signal and mitigate the item cold-start problem in POI recommendation.
Figures & tables
Figure 1: Overview of the LLM-MGCL methodology. Source: Authors’ own.
Type
Model
Recall@10
NDCG@10
Recall@20
NDCG@20
\svhline GNN
NGCF
0.0391 ± 0.0020
0.0287 ± 0.0012
0.0667 ± 0.0025
0.0380 ± 0.0013
MF
NeuMF
0.0457 ± 0.0008
0.0339 ± 0.0009
0.0773 ± 0.0007
0.0446 ± 0.0009
GNN
GC-MC
0.0461 ± 0.0019
0.0335 ± 0.0013
0.0803 ± 0.0030
0.0451 ± 0.0017
GNN
LightGCN
0.0453 ± 0.0015
0.0327 ± 0.0010
0.0783 ± 0.0023
0.0439 ± 0.0013
MF
MF-BPR
0.0554 ± 0.0017
0.0412 ± 0.0015
0.0923 ± 0.0029
0.0537 ± 0.0020
CF
ItemKNN
0.0534
0.0398
0.0881
0.0517
Table 1: Performance comparison on the Yelp dataset. Results are averaged over three runs (mean ± std). Deterministic methods (ItemKNN, UserKNN) are reported without standard deviation. Best results are shown in bold, second-best are underlined. Source: Authors’ own.
Model Variant
R@20
Δ R@20
N@20
Δ N@20
\svhline LightGCN
0.0783 ± 0.0023
–
0.0439 ± 0.0013
–
LLM-MGCL w/o CL
0.0960 ± 0.0004
+22.6%
0.0555 ± 0.0002
+26.4%
LLM-MGCL w/o GGEO
0.1126 ± 0.0006
+43.8%
0.0669 ± 0.0001
+52.4%
LLM-MGCL w/o GSEM
0.1193 ± 0.0014
+52.4%
0.0715 ± 0.0008
+62.9%
LLM-MGCL (full)
0.1163 ± 0.0005
+48.5%
0.0698 ± 0.0003
+59.0%
Table 2: Ablation study of LLM-MGCL components on the Yelp dataset. Results are averaged over three random seeds (42, 123, 2024) and reported as mean ± std. Δ denotes relative improvement over LightGCN baseline. Source: Authors’ own.
Model
Cold
Warm
Hot
\svhline LightGCN
0.0000±0.0000
0.0027±0.0004
0.1244±0.0039
SGL
0.0030±0.0006
0.0194±0.0012
0.1854±0.0013
LLM-MGCL
0.0106±0.0013
0.0397±0.0005
0.1674±0.0007
Table 3: Recall@20 stratified by item popularity on the Yelp dataset (mean ± std over three seeds). Best value per bucket in bold. Source: Authors’ own.
We introduce Agent4POI, the first POI recommendation framework that generates context-conditioned multimodal representations at recommendation time, rather than relying on static POI embeddings pre-computed independently of context. Existing multimodal systems encode each POI once as a static embedding, a design that precludes reasoning about why the same cafe affords solo work on Monday but group celebration on Friday evening. We formally prove that no pre-computed encoder can satisfy context-sensitive ranking under standard bilinear scoring, motivating inference-time item-side representation. Agent4POI inverts this computation: given a situational context, a four-phase LLM agent generates dynamic, context-specific affordance queries (Phase 1) and executes a five-step cross-modal chain-of-thought over image, review, and metadata evidence (Phase 2). The resulting uncertainty-aware affordance representation is grounded in Gibsonian affordance theory. These cross-modal verdicts form a structured, uncertainty-adjusted affordance representation (Phase 3), which is aligned with user preferences via a semantic caching system for low-latency ranking (Phase 4). On three POI benchmarks and three evaluation configurations (standard, cold-start, context-shift), Agent4POI achieves a 23.2% relative gain over the strongest baseline and degrades by only 7.5% under context-shift versus 16--17% for the strongest baselines. In cold-start scenarios, Agent4POI outperforms the best content-based baseline by up to 2.4x, whereas ID-based methods fail to generalize.
Jinze Wang, Yangchen Zeng, Tiehua Zhang +5
Tongji University · Swinburne University of Technology · Southeast University +4
Generative point-of-interest (POI) recommendation models based on large language models (LLMs) have shown promising results by formulating next POI prediction as a sequence generation task. However, the knowledge encoded in these models remains fixed after training, making them unable to perceive evolving real-world conditions that shape user mobility decisions, such as local events and cultural trends. To bridge this gap, we propose AWARE (Agent-based World knowledge Augmented REcommendation), which employs an LLM agent to generate location- and time-aware contextual narratives that capture regional cultural characteristics, seasonal trends, and ongoing events relevant to each user. Rather than introducing generic or noisy information, AWARE further anchors these narratives in each user's behavioral context, grounding external world knowledge in personalized spatial-temporal patterns. Extensive experiments on three real-world datasets demonstrate that AWARE consistently outperforms competitive baselines, achieving up to 12.4% relative improvement.
Graph Collaborative Filtering models have achieved remarkable success in recommendation systems. However, they inherently suffer from a catastrophic representation collapse in strict cold-start scenarios due to the topological isolation of unobserved items. While integrating multi-modal semantic features offers a potential remedy, directly mapping isolated semantic attributes into a globally topology-driven collaborative space frequently precipitates representation degradation and optimization conflicts. To bridge this fundamental structural divide, we propose DualGCN, an innovative Dual-Branch Graph Knowledge Distillation Framework. Operating under a decoupled two-stage training paradigm, DualGCN first constructs a robust collaborative teacher network where item representations are anchored by a topology-smoothed semantic prior, effectively reshaping the latent manifold. Subsequently, a semantic student branch employs a residual mapping mechanism and a magnitude-agnostic cosine distillation objective to achieve high-fidelity cross-modal alignment without disrupting established collaborative signals. Furthermore, we introduce a dynamic norm reconstruction mechanism to rigorously calibrate the physical scales of novel and warm items during global retrieval. Extensive experiments across three real-world benchmark datasets demonstrate that DualGCN significantly outperforms state-of-the-art baselines in strict cold-start scenarios while comprehensively preserving recommendation accuracy for highly active items.
Yuanfeng Qiu
The Hong Kong University of Science and Technology (Guangzhou)