时空嵌入式生成对抗网络的地点预测方法

摘要

The wide use of positioning technology makes the mining of the people movements easy and plenty of trajectory data are recorded. How to efficiently handle these data for location prediction is a popular research topic as it is fundamental to location-based services( LBS) . The existing methods focus either on long time ( days or months) visit prediction( i. e. point of interest recommendation) or on real time location prediction(i. e. trajectory prediction). In this paper, the location prediction problem in weak real time conditions is discussed to predict users′ movement in next minutes or hours. A spatial-temporal long-short term memory model ( ST-LSTM ) combining spatial-temporal influence into LSTM model naturally is proposed to mitigate the data sparse problem. Furthermore, following the idea of generative adversarial network(GAN) for seq2seq learning, the ST-GAN model is proposed, and it takes the proposed ST-LSTM as the generator and the proposed spatial-temporal convolutional neural network ( ST-CNN) as the discriminator. The minimax game of ST-GAN can produce more real enough data to train a better prediction model. The proposed ST-GAN is evaluated on a real world trajectory dataset and the results demonstrate the effectiveness of the proposed model.%定位技术的广泛使用可以积累大量的用户轨迹信息,为挖掘用户的行为轨迹提供便利.地点预测任务是众多基于位置服务的基础,学者们更关注如何有效利用这些轨迹数据进行地点预测.已有的方法或关注对长期模式(数天或数月)的预测,或致力于实时轨迹预测.文中研究的问题基于上述两者之间,即对弱实时条件下(数分钟或数小时)用户下一步的访问行为进行预测.为此,提出时空嵌入式的生成对抗网络模型(ST-GAN),在序列生成对抗网络的基础上,提出时空嵌入式长短时记忆生成模型(ST-LSTM)和时空嵌入式卷积神经网络判别模型(ST-CNN).ST-LSTM利用时空信息引导LSTM训练门机制,缓解数据的稀疏性.ST-CNN利用时空信息增强判别真伪访问序列的能力.此外,ST-GAN的训练优化机制使模型可以生成更多逼近真实的数据以引导模型学习,从而得到更好的预测效果.最后在真实的轨迹数据集上的实验验证ST-GAN的有效性.