The prevalence of Location-based Social Networks (LBSNs) services makes next personalized Point-of-Interest (POI) predictions become a trending research topic. However, due to device failure or intention camouflage, geolocation information missing prevents existing POI-oriented researches for advanced user preference analysis. To this end, we propose a novel model named Bi-STAN, which fuses bi-direction spatiotemporal transition patterns and personalized dynamic preferences, to identify where the user has visited at a past specific time, namely missing check-in POI identification. Furthermore, to relieve data sparsity issues, Bi-STAN explicitly exploits spatiotemporal characteristics by doing bilateral traceback to search related items with high predictive power from user mobility traces. Specifically, Bi-STAN introduces (1) a temporal-aware attention semantic category encoder to unveil the latent semantic category transition patterns by modeling temporal periodicity and attenuation; (2) a spatial-aware attention POI encoder to capture the latent POI transition pattern by modeling spatial regularity and proximity; (3) a multitask-oriented decoder to incorporate personalized and temporal variance preference into learned transition patterns for missing check-in POI and category identification. Based on the complementarity and compatibility of multi-task learning, we further develop Bi-STAN with a self-adaptive learning rate for model optimization. Experimental results on two real-world datasets show the effectiveness of our proposed method. Significantly, Bi-STAN can also be adaptively applied to next POI prediction task with outstanding performances.

基于位置的社交网络 (LBSN) 服务的流行使得下一个个性化的兴趣点 (POI) 预测成为一个趋势研究课题。然而，由于设备故障或意图伪装，地理定位信息的缺失阻碍了现有的面向 POI 的高级用户偏好分析研究。为此，我们提出了一种名为 Bi-STAN 的新模型，该模型融合了双向时空转换模式和个性化动态偏好，以识别用户在过去特定时间访问过的地方，即缺少签到 POI 识别。此外，为了缓解数据稀疏问题，Bi-STAN 通过双边回溯显式利用时空特征，从用户移动轨迹中搜索具有高预测能力的相关项目。具体来说，Bi-STAN 引入（1）一种时间感知注意力语义类别编码器，通过对时间周期性和衰减进行建模来揭示潜在语义类别转换模式；(2) 一种空间感知注意力 POI 编码器，通过对空间规律性和邻近性建模来捕获潜在的 POI 转换模式；(3) 一个面向多任务的解码器，将个性化和时间方差偏好纳入学习的转换模式中，用于丢失签到 POI 和类别识别。基于多任务学习的互补性和兼容性，我们进一步开发了具有自适应学习率的Bi-STAN，用于模型优化。两个真实世界数据集的实验结果表明了我们提出的方法的有效性。显著地，