基于教师-学生时空半监督网络的城市事件预测方法

A Teacher-Student Spatiotemporal Semi-Supervised Method for Urban Event Forecasting

离散时空事件预测是城市计算领域中的重要科学问题之一.现有工作主要聚焦于使用多样化的时空神经网络对城市动态特征与事件时空关联进行建模,且已经取得了一定成效,但仍然存在以下问题:首先,城市事件具有诱因多源和时空稀疏性,而这种时空稀疏性可能同时源于事件本身的稀少性和采集的不完整性,现有工作尚未能解决短期预测中的稀疏性挑战及零膨胀问题;其次,已发生事件倾向于继续向周边区域传播事件风险,但由于现有工作同质化了动态特征和事件之间的交互关联,因此其不能捕捉历史事件对未来事件风险带来的交互影响.鉴于此,为协同地利用事件标记信息和时空特征,本文提出基于教师-学生时空半监督学习框架以预测短期离散事件的时空分布.在教师网络中,为应对事件标记的稀疏性,本文在时空学习中引入半监督机制,提出基于自编码器的特征重建和时空方差异常描述引导的动态特征表示学习;在学生网络中,本文设计了特征-事件解耦的双管道学习机制,并提出时空衰减图卷积网络与长短期记忆网络来模拟事件在时空范围内发生的风险传播.此外,本文发展了时空多粒度预测机制,通过易学的粗粒度预测任务指导细粒度的高质量预测,最终实现粗-细粒度协同提名的离散时空事件预测.实验基于纽约和苏州工业园区数据集开展,本文模型能够在事件击中准确率上分别超越最好的基线模型5.46%和10.65%,充分验证了提出方法的有效性.

Discrete spatiotemporal event forecasting is one of critical scientific problems in the field of urban computing.Existing works mostly focus on leveraging various spatiotemporal neural networks to model spatiotemporal correlations among dynamic urban features and events,and have achieved promising results.However,it still remains the following problems.First,urban events are naturally induced by multiple causes and distributed spatiotemporal sparsely,while such spatiotemporal sparsity can be induced by the inherent infrequent occurrence and its collection incompleteness.Given that,existing works cannot well address such sparsity challenge and zero-inflated issue in short-term forecasting.Second,the occurred events have the potential to raise future risks on neighboring regions.Unfortunately,off-the-shelf literatures tend to homogenize the correlations of feature-event and event-event,and fail to capture the wane-and-wax influences of historical event sequences on future events.Therefore,to cooperatively exploit the event labels and spatiotemporal features,this paper proposes a teacher-student spatiotemporal semi-supervised learning framework,addressing the challenge of shortterm spatiotemporal event forecasting.In the teacher network,to tackle the sparsity challenge of event labels,this paper introduces the semi-supervised scheme into spatiotemporal learning where it designs an AutoEncoder-based feature reconstruction learning and spatiotemporal variance-based anomaly descriptor to facilitate feature representations.In the student network,this work designs a feature-event disentangled dual pipeline and proposes the spatiotemporal attenuation graph convolution network(GCN)and long-short term memory network(LSTM)to imitate the natural risk propagation along spatiotemporal domains.In addition,this paper also develops the spatiotemporal multi-granularity risk prediction task,which emphasizes the easy-to-learn coarse-grained prediction to guide the high-quality fine-grained forecasting,and finally realizes the high-risk discrete region nomination with coarse-to-fine learning mechanism.Experiments on NYC and SIP datasets illustrate that the proposed event forecasting framework outperforms the best baselines by respectively 5.46%and 10.65%,verifying the effectiveness of our work.