Human Mobility Prediction Based On Semantics Spatial Temporal Data

Location based services have accumulated a large number of semantics spatial-temporal data,which record users' behavior in the real world.Predicting human mo-bility based on semantics spatial-temporal data plays a key role in many applications.Related works of human mobility prediction can be divided into two categories:indi-vidual mobility prediction and crowd flow prediction.Individual mobility prediction focuses on model individual mobility pattern and predicting individual's future activity and location,which benefits location recommendation,travel planning and intelligent advertisement.Crowd flow prediction tries to understand the crowd flow and distri-bution through specific geo-spatial dimension,which sheds light to traffic manage,ur-ban region understanding and public epidemic prevention.Comparing with traditional spatial-temporal data,semantics spatial-temporal data contain users' activity and loca-tions',semantics information,creating opportunities for individual mobility and crowd flow prediction.Meanwhile,semantics spatial-temporal data are usually sparse and spo-radic,bringing challenges for human mobility prediction.In this article,after surveying advantages and disadvantages of related works,we develop several research works in human mobility prediction based on semantics spatial-temporal data.Firstly,we work on utilizing activity information for individual location predic-tion.Viewing that user activity serves as mobility motivation,activity prediction is introduced as auxiliary to reduce vast location candidate space.Unfortunately,many unresolved difficulties remain tough in location prediction,including how to integrate the sequential and temporal regularity of human mobility and how to capture individ-ual location preference under specific temporal context.In this paper,we tackle above challenges with a two-stage location prediction method.In the first stage,we design a Context-Aware Stacking module to integrate sequential and temporal regularity dy-namically for predicting individual activity.In the second stage,we adopt a Bayesian approach combining with Kernel Density Estimation for temporal location preference calculation.The experimental results on two public datasets validate that our model significantly outperforms state-of-the-art baselines in terms of both activity prediction accuracy and locations prediction accuracy.Secondly,we focus on the interplay of individual's activity and location prefer-ence.Considering that users' activity and location preferences interplay with each other,many scholars tried to figure out the relation between users' activities and lo-cations for improving prediction performance.However,most previous works enforce a rigid human-defined modeling strategy to capture these two factors,either activity purpose controlling location preference or spatial region determining activity prefer-ence.Unlike existing methods,we introduce spatial-activity topics as the latent factor capturing both users' activity and location preferences.We propose Multi-task Gated Recurrent Neural Network to leverage the spatial activity topic for activity and loca-tion prediction.More specifically,a novel Context Aware Recurrent Unit is designed to integrate the sequential dependency and temporal regularity of spatial activity topics.Besides,we employ a graph embedding method to for location and activity represen-tation learning.Extensive experimental results demonstrate that the proposed model significantly outperforms state-of-the-art approaches,including a case study for vividly explaining the spatial activity topics.At last,we study the crowd flow prediction of Point Of Interest(POI).POI crow flow prediction is a valuable for both visitors and managers of POI,and also benefits POI recommendation service.Scholars handle this problem with non-parametric time series method and parametric machine learning models for capturing the temporal pattern of POI crowd flow.After analysis of 8,000 POIs' crow flow data,we find that POI crow flow not only shows temporal pattern but also unexpected fluctuation.For capturing both the temporal pattern and unexpected fluctuation,we propose a Regular-Unexpected vis-itor Mixed method(RUM)for POI crow flow prediction.In RUM method,we adopt non-parametric method for temporal pattern feature extraction,design a graph convolu-tion neural network for region flow prediction and region-activity transition matrix for region-activity distribution calculation.Then we integrate all these features and apply regression model for POI crow flow prediction.The effectiveness of RUM method is proven in extensive experiments on real POI crow flow dataset.