| Modelling temporal,spatial,and semantic(TSS)features of human activities in urban areas based on geotagged data plays an important role in many fields,such as socio-economic status,crime dynamics research,epidemic prevention and control,etc.However,the effective applications of existing models are limited by the TSS uncertainties,which widely exist in the relevant processing of geotagged data.The positioning error,sampling frequency and model deviation in data acquisition,preprocessing,modelling,and application may all lead to uncertainties.Facing the uncertainty problems,many researchers work on measurement,visualization,control strategy,and other aspects.However,existing human activity models still lack enough research about dynamic,density change,multi-feature fusion,and other aspects,which influence the effective control of the uncertainties.Therefore,this study proposes corresponding control strategies of TSS uncertainties based on dynamic,density change,and multi-feature fusion,and then integrates the three control principles to present a comprehensive control scheme of TSS uncertainties to achieve reliability modelling.The core contributions of this dissertation are as follows:(1)A spatio-temporal prediction algorithm of human activities based on the dynamics of mobility patterns is proposed.Given the dynamics in human activities,including the changes in activity regions of interests and movements,the region change detection scheme and integrated prediction model are proposed to control the temporal uncertainty.The region change detection can effectively assign new data,and combine a matching scheme and a similarity index to describe the change degrees of the new and old regions.Additionally,the change detection decides to update or reconstruct the prediction model.Then,two prediction models are built respectively for fast change and slow change based on adaptive windows and decay factors.An adaptive weight method is designed to measure the importance of different activity features and evaluate the effects of the two models in real-time for model integration.The experiments show that there are widespread pattern changes in human activities,which are significantly affected by vacations,major public health events,and other factors.The proposed method can effectively detect changes to ensure the true and reliable extraction of human activity patterns.In terms of spatio-temporal prediction performance,the proposed method has higher accuracies than the existing advanced algorithms.Detailly,accuracy improvements ranging from 10% to 40% can be achieved on different datasets.Especially when the change degree is high,the improvements are more obvious.Additionally,the proposed method has a more stable prediction ability for different individuals.The experiments prove the effectiveness of this study,which detects changes in activity areas and activity patterns to control the temporal uncertainty.(2)A human activity area extraction algorithm based on density changes of spatial points is proposed.The distribution of spatial points in large-scale geotagged data is complex,and the spatial density of the same area is also significantly different,making spatial uncertainty difficult to be controlled.The existing algorithms don’t take the spatial uncertainty into account effectively,disturbing their reliabilities.Therefore,based on the density changes of spatial points,this study proposes a method to control the spatial uncertainty and extract reliable activity areas.Firstly,a distance parameter of point densities is adaptively extracted and the high-density points are clustered to reduce the impact of human subjective factors.Secondly,the spatial characteristics of each cluster are considered separately to assign remaining points and optimize the clusters’ spatial structure.Then,a re-segmentation strategy is designed,which can select appropriate large clusters for multiple segmentation,and realize the adaptation to different densities of spatial points.Finally,the extra noises in the re-segmentation are recycled to improve the data utilization and extract the potential active areas as many as possible.The experiments show that fine-grained activity areas can be more effectively extracted from large-scale spatial data with varying densities,and both highdensity and low-density clusters with less noise are simultaneously generated.The proposed algorithm also obtains optimal performances on all datasets and reaches 1.6times the effect of the suboptimal algorithm.The improvements are more obvious when the density differences are large.Additionally,the analysis of temporal and clustering characteristics proves that the method can effectively capture the spatio-temporal features of human activity areas.(3)A method for human activity topic modelling based on multi-features fusion is proposed.The existing research lacks comprehensive consideration of TSS features in geotagged data,and ignores the deviation between the texts and other features,resulting in semantic uncertainty in topic modelling.Therefore,based on the fusion of TSS features,a novel method is proposed to control the semantic uncertainty and realize more reliable and effective topics of human activities.The method combines the text with TSS information to extract the distributions of the words.Then the method measures the differences between the distributions of the words and the overall distributions with novel indices to describe the representation abilities of words.The differences and indices are used to extract TSS special words and solve the deviation problem in TSS feature fusion.Based on TSS special words,a variety of raw text revision schemes are proposed from the perspective of filtering,characteristics,and distributions of the words.The basic data structure is retained so that the mainstream models can be directly applied to the revised texts for enhancement.The experiments show that there are widespread deviations,and the TSS distributions are revised to improve the topic results.The proposed method obtains a clearer understanding of the activity contents and maintains a higher ranking in the comprehensive performance compared with the unrevised distributions used in the existing research.The results prove that the proposed method can effectively enhance the topic model based on the revised text to better reflect the TSS characteristics of human activities.(4)A multi-topic and dynamic attractiveness rating algorithm of human activity areas based on uncertainty control is proposed.After presenting the uncertainty control schemes based on dynamics,density changes,and multi-feature fusion,there are still challenges in how to realize multi-uncertainty control in the same scene.Therefore,this study proposes a reliable model for attractiveness rating by integrating three uncertainty control principles.Firstly,the area extraction based on the density changes and the topic modelling based on the multi-feature fusion are modified for dynamics.Then the methods can assign new data to existing areas and topics,and update or reconstruct the models,so the dynamic activity areas and topics are generated.Finally,the numerical scaling strategy is improved to calculate the scores of users and activities iteratively with adaptive windows and decay factors,and the dynamic attractiveness rating results of multi-topic areas are obtained.The experiments show that the proposed algorithm can obtain fine-grained,large-space coverage and real-time updated activity areas,which are more consistent with the real activity patterns,instead of few static areas extracted in the existing research.The dynamic topic modelling effectively reflects the changes in TSS features with higher reliability,compared with the tags used in other studies.In terms of the attractiveness rating results,the iterative convergence efficiency and effect are better,with wider coverage and finer granularity.Additionally,the multilevel rating of activities can be achieved through a transformation and the proposed algorithm better describes the latest activity pattern under different topics,and achieves the best performance under different circumstances. |
