[论文解读] Multifaceted Scenario-Aware Hypergraph Learning for Next POI Recommendation

Yuxi Lin, Yuanzhe Li+2 more|arXiv (Cornell University)|Jan 9, 2026

Recommender Systems and Techniques被引用 0

一句话总结

MSAHG 通过构建情景特定的子超图并使用自适应参数分裂来解决情景间冲突，显式建模多情景移动，在三个真实数据集上达到最先进的结果。

ABSTRACT

Among the diverse services provided by Location-Based Social Networks (LBSNs), Next Point-of-Interest (POI) recommendation plays a crucial role in inferring user preferences from historical check-in trajectories. However, existing sequential and graph-based methods frequently neglect significant mobility variations across distinct contextual scenarios (e.g., tourists versus locals). This oversight results in suboptimal performance due to two fundamental limitations: the inability to capture scenario-specific features and the failure to resolve inherent inter-scenario conflicts. To overcome these limitations, we propose the Multifaceted Scenario-Aware Hypergraph Learning method (MSAHG), a framework that adopts a scenario-splitting paradigm for next POI recommendation. Our main contributions are: (1) Construction of scenario-specific, multi-view disentangled sub-hypergraphs to capture distinct mobility patterns; (2) A parameter-splitting mechanism to adaptively resolve conflicting optimization directions across scenarios while preserving generalization capability. Extensive experiments on three real-world datasets demonstrate that MSAHG consistently outperforms five state-of-the-art methods across diverse scenarios, confirming its effectiveness in multi-scenario POI recommendation.

研究动机与目标

• 在多面性上下文情景（用户类型、时间、空间）下为下一个 POI 推荐提供动机。
• 开发一个框架，通过多视角子超图捕捉情景特定的移动模式。
• 通过自适应参数分裂在保持泛化的同时缓解情景间优化冲突。

提出的方法

• 从协同、时序与地理视角以及一个共享过渡视角构建八个情景特定子超图。
• 使用带残差连接的两步子超图卷积来更新节点/超边。
• 将情景特定嵌入与视角启发的门控融合，并使用对比学习（InfoNCE）对齐多视图表示。
• 引入自适应参数分裂，将冲突参数复制并按梯度相似性将情景分组，以实现情景感知的训练。
• 以将对比项和推荐损失相结合的最终损失进行优化，以平衡多视图协作与准确性。

实验结果

研究问题

• RQ1 如何在多情景（用户类型、时间、空间）下有效建模下一个 POI 的推荐？
• RQ2 情景特定的子超图是否能比单一图基线更好地捕捉不同的移动模式？
• RQ3 自适应参数分裂是否能降低情景间干扰并改善多情景学习？
• RQ4 多视图融合和对比学习对在多情景下的推荐质量有何影响？

主要发现

Model	Scenario	ACC@1	ACC@5	ACC@10	ACC@20	MRR	ACC@1	ACC@5	ACC@10	ACC@20	MRR	ACC@1	ACC@5	ACC@10	ACC@20	MRR
Gowalla	Local	0.1314	0.2758	0.3369	0.3934	0.2014	0.2020	0.3762	0.4237	0.4583	0.2809	0.2159	0.3878	0.4573	0.5203	0.2978
Gowalla	Tourist	0.1079	0.2349	0.2999	0.3623	0.1714	0.1298	0.2576	0.3056	0.3398	0.1886	0.1679	0.3137	0.3762	0.4367	0.2387
Gowalla	HSTLSTM	0.1405	0.2909	0.3558	0.4163	0.2122	0.2065	0.3731	0.4276	0.4650	0.2334	0.4236	0.5002	0.5626	0.3229
Gowalla	Suburban	0.1102	0.2391	0.3019	0.3614	0.1746	0.1454	0.2877	0.3327	0.3658	0.2107	0.1785	0.3227	0.3828	0.4452	0.2489
Gowalla	Workday	0.1366	0.2846	0.3512	0.4133	0.2085	0.1723	0.3295	0.3793	0.4152	0.2439	0.2118	0.3813	0.4497	0.5121	0.2927
Gowalla	Weekend	0.0820	0.1901	0.2471	0.3023	0.1373	0.0240	0.0524	0.0689	0.0778	0.0394	0.0728	0.1715	0.2275	0.2897	0.1239
NYC	Local	0.1343	0.2984	0.3644	0.4242	0.2122	0.1603	0.3311	0.3871	0.4268	0.2382	0.1756	0.3791	0.4575	0.5232	0.2698
NYC	Tourist	0.1074	0.2435	0.3066	0.3699	0.1751	0.1178	0.2322	0.2827	0.3272	0.1722	0.1455	0.3155	0.3861	0.4411	0.2254
NYC	DeepMove	0.1441	0.3114	0.3807	0.4402	0.2238	0.1544	0.3078	0.3606	0.3950	0.2235	0.1806	0.3897	0.4727	0.5392	0.2773
NYC	Suburban	0.1096	0.2505	0.3131	0.3758	0.1789	0.1305	0.2660	0.3192	0.3648	0.1944	0.1574	0.3406	0.4114	0.4714	0.2428
TKY	Downtown	0.1556	0.3385	0.4104	0.4768	0.2432	0.1782	0.2970	0.3762	0.5149	0.2398	0.1416	0.3625	0.4969	0.6247	0.2479
TKY	Suburban	0.0785	0.1983	0.2438	0.3058	0.1329	0.0714	0.2143	0.3571	0.4464	0.1519	0.0213	0.1277	0.2128	0.3404	0.0771
TKY	DCHL	0.1929	0.4180	0.5145	0.5627	0.2968	0.1000	0.2000	0.3000	0.5000	0.1600	0.1511	0.4106	0.5668	0.7162	0.2753
TKY	Suburban	0.1262	0.2816	0.3398	0.4107	0.2011	0.1460	0.2774	0.3796	0.4891	0.2155	0.1207	0.2783	0.3781	0.4741	0.1978
TKY	Workday	0.1494	0.3274	0.3980	0.4651	0.2337	0.1450	0.2748	0.3969	0.5115	0.2157	0.1438	0.3598	0.4912	0.6259	0.2480
TKY	Weekend	0.1031	0.2422	0.2915	0.3498	0.1713	0.1154	0.2308	0.2308	0.3846	0.1719	0.1223	0.3476	0.4871	0.5923	0.2304
MSAHG	Downtown	0.1871	0.3680	0.4349	0.4882	0.2723	0.2947	0.6974	0.8237	0.8605	0.4666	0.2132	0.4858	0.5689	0.6410	0.3358
MSAHG	Suburban	0.1540	0.3151	0.3801	0.4316	0.2310	0.1816	0.4447	0.5658	0.6000	0.3013	0.1993	0.4073	0.4789	0.5372	0.2927
MSAHG	Workday	0.1882	0.3585	0.4301	0.4945	0.2704	0.2316	0.5474	0.6526	0.7158	0.3736	0.2171	0.4804	0.5620	0.6216	0.3351
MSAHG	Weekend	0.0960	0.2250	0.2857	0.3346	0.1580	0.1868	0.4395	0.5500	0.5763	0.3011	0.1446	0.3218	0.3902	0.4534	0.2274

• MSAHG 在 Gowalla、NYC 与 TKY 数据集的多情景下均达到最先进的性能。
• 情景特定子超图与自适应参数分裂对性能提升均有显著贡献。
• 相较于基线，MSAHG 在本地/游客以及市区/郊区情景中更准确地保留真实 POI 类别分布和距离模式。
• 消融实验表明，在 NYC 的 ACC@1/5/10/20 与 MRR 指标上，完整模型优于无分裂或无子超图的变体。
• MSAHG 相对于若干基线在效率指标上更有利，梯度缓冲区带来的内存开销可控。