室内用户语义位置预测研究

位置预测技术可以提前预知用户下一时刻的位置,在基于位置的服务（Location-based Service,LBS）领域中发挥着极其重要的作用。现有的位置预测技术大多仅使用用户的地理轨迹,仅使用地理轨迹挖掘出来的用户移动模式易受地理特性的限制缺乏深层次的语义信息。本文基于某商场群体用户的室内轨迹数据和语义信息预测用户下一个时刻语义位置。语义位置预测包括停留区域识别、停留区域语义匹配、语义位置建模。在停留区域识别阶段,为减少室内停留时间不固定对停留区域识别的影响,本研究提出了一种新型的时空凝聚层次聚类算法（Spatial-Temporal Agglomerative Nesting, ST-AGNES),该算法具有思想简单、超参数少、自动生成聚类个数等优点。在语义匹配阶段,引入了吸引度规则,充分利用停留区域所有轨迹点与室内高密度的商铺名称信息做匹配。最后,采用长短型记忆神经网络模型（Long Short-Term Memory,LSTM）挖掘群体用户的语义位置模式并预测用户未来的语义位置,实验预测正确率达到61.3%。

Research on Semantic Location Prediction of Indoor Users

The location prediction technology can predict the location of the user at the next moment in advance, and plays an extremely important role in the field of Location-based Service (LBS).Most of the existing location prediction techniques only use the geographical location information and time information of the user's historical trajectory. The geographic trajectory is composed of a series of geographically-pointed time-stamped latitude and longitude points, and the geographic trajectory only mines users. Mobile mode is limited by geographic features. In this paper, we propose a novel approach for predicting the next semantic location of a user's movement based on the geographic and semantic characteristics of the group user trajectory. The semantic location prediction based on group users generally consists of three steps: Firstly, the specific algorithm is used to identify the staying area in the user's trajectory; Next, the semantic matching algorithm is used to associate the user's staying area with the semantic information; Finally, Mining the semantic location pattern of group users, using this pattern to predict the semantic location of the user at the next moment. In the stage of staying area identification, in order to reduce the influence of indoor stay time unfixed on the recognition of stay area, this paper proposes a new type of spatial-temporal agglomerative nesting (ST-AGNES), which can automatically identify the number of staying areas in the user's trajectory using only the distance threshold. In the semantic matching stage, this paper proposes a semantic matching method based on attractance rules, which makes uses all trajectory points in the stay area to be associated with indoor high-density semantic information. In the final forecasting stage, this paper uses Long Short-Term Memory (LSTM) to mine the semantic location patterns of group users and predict the future semantic location of users. The experimental results have achieved a prediction accuracy rate of 61.3%.