The ability to recommend candidate locations for service facility placement is crucial for the success of urban planning. Whether a location is suitable for establishing new facilities is largely determined by its potential popularity. However, it is a non-trivial task to predict popularity of candidate locations due to three significant challenges: 1) the spatio-temporal behavior correlations of urban dwellers, 2) the spatial correlations between candidate locations and existing facilities, 3) the temporal auto-correlations of locations themselves. To this end, we propose a novel semi-supervised learn- ing model, Spatio-Temporal Graph Convolutional and Recurrent Networks, aiming for popularity prediction and location recommendation. Specifically, we first partition the urban space into spatial neighborhood regions centered by locations, extract the corresponding features, and develop the location correlation graph. Next, a contextual graph convolution module based on attention mechanism is introduced to incorporate local and global spatial correlations of locations. A recurrent neural network is proposed to capture temporal dependencies between locations. Furthermore, we adopt a location popularity approximation block to estimate the missing popularity from both spatial and temporal domains. Finally, the overall implicit characteristics are concatenated and then fed into the recurrent neural network to obtain the ultimate popularity. The extensive experiments on two real-world datasets demonstrate the superiority of the proposed model compared with state-of-the-art baselines. 

中文摘要：

POI服务设施的选址规划是城市计算的重要内容，良好的选址能为居民的工作生活带来便利，提高POI设施的服务效率，还能缓解城市的交通和环境等压力。现阶段大部分城市都已经发展到了一定的阶段，相应的服务设施都存在于城市的不同区域中。因此，选址工作不仅要考虑时空域上城市居民对候选地址的影响力，还要综合分析已经存在的服务设施对候选地址的影响效应，以及新建设施对现存设施的影响等因素。
为此，本文提出了半监督环境下的选址模型STGCRN。该模型首先以地址为中心，将城市区域分块划分并提取对应的多维度特征，以此为基础构建地址关联图。然后，引入基于多头注意力机制的上下文图卷积模块对地址设施的空间关联关系建模，捕获地址的局部空间相关性和全局空间相关性，并采用循环神经网络对地址设施的动态时间依赖关系建模，捕获周期性时间片内流行度的前后依赖性。最后，分别从空间域和时间域预估地址缺失的流行度分布，使用熵机制融合后，与时空特征一并再次送入循环神经网络模块，获得最终的候选地址流行度用于选址推荐。
在两个实验数据集中的共计6类具体服务设施上，本文所提预测模型STGCRN的性能都优于其他的比较算法。因为STGCRN重点分析和提取了城市区域内用户群体的时空轨迹特征和社交关联等特征，在半监督的环境下，采用基于多头注意力机制的上下文图卷积模块建模地址的空间局部相关性和空间全局相关性，并使用LSTM模块捕捉地址的动态时间依赖关系，从而解决了候选地址流行度预测中所面临的用户自相关、空间自相关和时间自相关等难题。
本文充分挖掘了用户签到、行程轨迹和社交行为等数据，与时空域结合得到的预测结果准确科学，在此基础上根据预测结果进行的增量选址，效率高、效果好，本研究可以为各类城市服务设施的选址提供理论支撑和算法参考。

Key words: location recommendation; popularity prediction; spatio-temporal correlation; deep learning;