Research On Activity Trajectory Similarity Based On Deep Representation Learning

The popularity of global positioning systems and mobile devices supporting wireless communication technologies have laid a good foundation for the generation and develop-ment of trajectory data.This phenomenon brings a series of studies and applications on tra-jectories.Trajectory similarity computation is a popular research on trajectory data,which is widely applied into POI recommendation,route planning,and so on.With the development of location-based social networks,the generation of activity trajectory provides traditional trajectory data with additional semantic information.Existing trajectory similarity algo-rithms not only ignore the rich semantic information of activity trajectories,but also have difficulty working on trajectory data with low sampling rates(heterogeneity problem).To solve these two problems,this paper proposes a similarity computation model of activity trajectories based on deep representation learning.In order to optimize the model's short-comings in dealing with long trajectories and distinguishing multiple features,a similarity computation model based on hierarchical attention mechanism is designed.The detailed work is as follows:Aiming at the problem that existing trajectory similarity algorithms only consider the spatiotemporal characteristics,this paper proposes an activity trajectory vectorization model(At2vec)based on deep representation learning and the semantic information of the activity trajectory.Firstly,the low-frequency trajectory is used as the model input and the high-frequency trajectory is the target,which solves the problem of data heterogeneity in tradi-tional algorithms.Second,matrix stitching is utilized to fuse the semantic information of the activity trajectory with the spatiotemporal features to lay the data foundation for model generation.Finally,a semantic proximity perception(STA)loss function is proposed,which mainly solves the problem of slow convergence during model training.Since At2vec model cannot capture long-term dependencies in trajectories and has difficulty distinguishing key elements of multiple features,our paper proposes a trajectory similarity computation model(At2vec+Atten)based on hierarchical attention mechanism.The model proposes three-level attention mechanism at point-level,feature-level and encoder-level,which realize the func-tion of automatically weighting trajectory points and features.The LSTM network is used to improve the Seq2Seq model by learning deeper sequence dependencies.To make up the shortcomings of language model processing spatio-temporal data,a data pre-training al-gorithm based on the spatio-temporal approximate distribution is proposed to improve the model by enhancing the quality of the data.The experiments of this paper are performed on two trajectory datasets in New York and Los Angeles.Experimental results illustrate that At2vec model is superior to existing methods on the evaluation criteria self-similarity measure,cross-distance and KNN query.The results of the At2vec+Atten model prove that it can further improve the precision of similarity computation based on the At2vec model.In addition,the model can maintain validity for trajectories of different sampling rates,which proves that it can solve the problem of trajectory heterogeneity commendably.